The Taste Harness

Here's the part that bothered me. I'm not bad at marketing and branding. I'd supplied the original brief myself, generated with AI beforehand from frameworks I've spent years building. That same day, four films later, an AI agent handed me a page that said what the work was for more clearly than my brief had.

I'll show you that page, and exactly what the agent had seen by the time it wrote it, in Chapter 15. For now I want the question, because the whole book hangs off it: why would the explanation of a piece of work get better after the work had been done?

The easy answer is that the model was clever. It wasn't a cleverer model than the one I'd started the day with. It was the same kind of model, running under the same kind of instructions. Something else had changed.

The objection I've been making for years

I've long argued against one particular way of using AI. You'll recognise it, because it's the way most people start.

I still hold that. Ask a model cold, "What's the best way to win this customer?", and you get a competent, general answer about awareness, trust, differentiation and a clear call to action. Nothing in it is wrong. None of it is yours. It's the answer to the question as anyone would ask it, not as your situation asks it.

The standard fix is the one in the quote: give it context. Load the brand guide, the persona, the offer, the history. That helps, and it's the right instinct as far as it goes. But it isn't what happened with the page.

The agent that wrote the page didn't have a longer brief than the one I'd started with. It had the same brief. What it had in addition was a day of consequences. It had watched films get built and fail review. It had seen two judges score the same cut eight point eight and six. It had heard me reject a voice as "crappy" and watched the fix. So when I asked it "why are we making these films?", it wasn't answering a question about films in general. It had a pile of specific contrasts to account for, and any answer that couldn't account for them was visibly wrong.

Do not expect the best explanation of a task from a system that has only been told about it. Give it access to consequential attempts, then ask it to account for what those attempts revealed.

That's the shape my oracle objection should have had all along. The trouble with the oracle was never only that it lacked context. It lacked contact. A system that has only been told about a task can describe the task. A system that has tried it, watched the attempts fail in specific ways and heard someone say why, can explain what the task is actually for. Chapter 12 unpacks the mechanism properly. Here I only want to put the claim on the table and say plainly that I believe it.

Two products, and only one of them compounds

Once you see the page that way, every round of work looks different. The obvious product of an iteration is the artefact: a better film, a sharper page, a cleaner proposal. There's a second product, and it's the one most pipelines throw away.

The artefact gets spent on the job. It ships, or it doesn't, and next week's job starts from zero. The account is the thing the next job could start from: the distinction you didn't have on Monday, the reason a lovely draft was wrong, the test that would have caught it earlier. Most teams keep the first product in a folder and let the second evaporate into chat history. They pay for their taste on every job, and it never compounds.

What I was actually doing that week, though I didn't have the words for it until afterwards, was using production to discover what "good" meant for this work, and then making that discovery operational. I said it at the time in rougher language:

That "sort of" was a real hedge, and I'll keep it. A test harness doesn't contain taste the way a person does. What it can contain is a growing set of distinctions that someone's taste produced, each with a way to check it. That's worth more than it sounds, because of the speed at which it was happening:

What I'm calling a taste harness

This book gives the practice a name, because a practice without a name doesn't get repeated. A taste harness is a way of running production so it forms the judgment nobody could brief, and a way of carrying that judgment so it survives into the next, unfamiliar job.

Each of those three moving parts has its own chapters. Rejection triage is Chapter 8. The three carriers are Chapter 9. The successor test is Chapter 26. Everything else in the book is either the reasoning that makes them necessary or the evidence that they work.

Where this sits among things I've already written

If you know my earlier work, you'll notice this book arrives next to three ideas that sound related. They are, and the difference matters.

The Taste Kernel, from The Cognition Dimension Ladder, is taste documented: your examples, anti-examples, rejection rules and "too-generic" detectors, applied as the evaluation function a search engine scores against, and kept honest by closing the loop with the real world. Worldview Recursive Compression takes each output, evaluates it, extracts the lesson and encodes it back into one maintained kernel. The Exemplar Ratchet freezes a win, with its cost, and extracts the direction of improvement from blind comparisons against it.

All three need something to hold. A documented taste to score against. An output to evaluate against an existing kernel. A win to freeze. This book is about the stretch before any of those exist: when the standard itself is still being formed, often for a client who couldn't have described it if you'd asked. It's the upstream of all three.

Why this is a problem now

For most of the history of professional work, making things was the expensive part. A film took a crew and weeks. A website took a sprint. A proposal took a partner's weekend. Taste got exercised slowly, a few times per job, by whoever was senior enough to say no.

That's gone. A one-minute film, a site, a thirty-page proposal can now be constructed and revised in hours. When making gets that cheap, the hard part moves. It stops being "can we make this?" and becomes "have we made the right thing?" And at the same time, the harnesses we use to check AI work have become capable of testing fit, not just fonts: whether a voice belongs with its audience, whether a scene connects to what the viewer actually notices, whether an opening assumes knowledge a stranger won't have. When making outruns judging, taste becomes the bottleneck. Chapter 11 takes that seriously, including the ways a harness that judges on taste goes wrong.

How the argument runs

The order of the book is the order in which each idea earns the next. Parts I and II say what taste is, why it can't be written into a brief, and where the missing context actually comes from. Part III builds the harness that forms taste and carries it: the three questions a rejection might be answering, the triage that tells them apart, the three carriers, and the guard against taste hardening into superstition. Part IV explains why the explanation improved, and how to stop a good explanation from promoting itself. Part V is the evidence: one day of films, from that page to a single rejection walked through all three of its lessons, with the before and after of the actual skill text. Part VI takes the hardest case, a client who can't brief you at all. Part VII says what being heard actually requires, and designs the test the whole harvest has to pass. Part VIII is the field kit.

That's the only time I'll say it. From here on I'll write with the conviction the evidence supports, inside that frame.

So: if the agent that explained the films didn't have more context than I'd given it at the start, what did it have? The answer needs a distinction that most people collapse, between what a system is given and what it discovers. That's the next chapter.

Start with a brand brief, read literally. A health clinic wants its films to feel warm, credible, reassuring, recognisably local.

Every word is right. Nobody would argue with any of them. Now watch what they allow.

None of those failures contradicts the brief. That's the problem with briefs, and it isn't a problem with this brief. A brief can't tell you where its own words stop. "Warm" doesn't contain the boundary between warm and patronising. Nobody writing it knew exactly where that boundary was, because nobody had seen the patronising version yet.

Where do the edges come from?

You find where the words' edges are by making something and reacting to it. You hear a narrator and something's wrong. "I don't like that voice." At that moment it's a preference, nothing more. But if you stay with it and ask why, it can turn into something more useful: the voice isn't low quality, it puts the speaker in the wrong relationship with the audience. Too much authority for a question the viewer hasn't asked yet. Too far away for a local service.

That's a candidate principle, and it reaches much further than "don't use voice number seven". It could shape casting, wording, pacing, the next brand's film. Provided you check that it actually explains the rejection, and isn't just a plausible story told after the fact. (The real voice rejection from my week was messier than this tidy version, and more instructive for it. Chapter 17 tells it straight.)

The brief supplies concepts. The encounters give those concepts boundaries.

That's why the agent's page in Chapter 1 could be better than my brief without knowing anything my brief didn't mention. It knew where the brief's words stopped, because it had watched them be overstepped.

A pedant will object that everything the supervising agent learned ended up in its context window anyway, so it's all context. Technically, yes. But calling everything "context" hides the mechanism, and the mechanism is the useful part. Some information was supplied at the start. Other information could only be obtained by attempting something and meeting resistance. Those two arrive by different routes, they need different handling, and only one of them is available on day one.

I noticed this before I could say it properly. My first attempt, talking it through at the end of the week, went like this:

"It's been exercised" is still the best short description I have. An exercised model isn't one with more reading behind it. It's one that has tried to do the thing and been told, specifically, where it went wrong.

I'd already written half of this, about reading

Perturbation isn't a new word in my work. In Cycle Compression I argued that a long conversation with a model that shares hundreds of hours of your vocabulary is "not mainly a sycophancy problem. It is a path dependence problem." Shared context creates conceptual gravity; new observations fall toward the concepts you already have. My answer then was that web research is "not a citations chore. It is an external perturbation source": someone who has never heard of your framework kicks the thought sideways, and "perturbation keeps the breath from becoming a pure sine wave of self-agreement."

That was perturbation by reading. The films showed me perturbation by making. A build that fails review, two judges who disagree about the same cut, a rejection with a reason attached: that's perturbation the work generates for itself, and it's aimed at exactly the part of the problem you're working on. Web research kicks the thought from outside. Consequential attempts kick it from inside the job.

The same extension applies to an argument I made in The Inference Field. Two contexts "can both be 200,000 tokens and be cognitively nothing alike", because of how the tokens got admitted. A context built by retrieval is a pile sorted by resemblance. A context built by walking, where each thing read changes what's worth reading next, is "a causal history of inquiry". Or, as I put it then: "Retrieval spends context on candidates. A traversal spends context on consequences." The supervising agent's context was a travelled argument too. Its steps just weren't all reads. They were constructions, listenings, rejections and repairs.

And I'd seen the design version of this before the films. Co-designing a system with AI a few months earlier, I wrote that "I could not completely explain what I wanted at the start. I had the shape of it", and that "the advances were perturbation, not pure decomposition." Same finding, different medium. There it was a system's shape nobody could specify upfront. Here it's taste.

It isn't just me

The idea that you can't fully specify what "good" means before you've judged some attempts turns up wherever people have looked for it carefully.

Researchers at Berkeley watched people build evaluators for language-model output and named the effect criteria drift: "users need criteria to grade outputs, but grading outputs helps users define criteria". Their conclusion is blunt: "it is impossible to completely determine evaluation criteria prior to human judging of LLM outputs."¹ Participants "wanted to add new criteria when they observed new 'types' of bad LLM outputs." They couldn't have listed those types in advance. The outputs taught them.

Donald Schön found the same thing in architecture studios forty years ago. Watching a student designer, he noted that "her intention was not fully established at the beginning of her design process, but evolved through her appreciation of an intermediate design product… An evolving intention is one of the outputs of her designing."² Software engineering gave it a name that everyone who's shipped to a client recognises: IKIWISI, "I'll Know It When I See It". "Requirements often emerge only after users have had an opportunity to view and provide feedback on models or prototypes."³ And AI-safety researchers at DeepMind put the asymmetry in one line: "It is often easier to evaluate whether an outcome has been achieved than to specify it explicitly."⁴

Four fields, one pattern. The standard isn't sitting in someone's head waiting to be written down. It forms when attempts meet judgment.

So does any activity count?

No, and this is the qualifier that keeps "perturbation" from becoming a fancy word for "more activity".

Ten near-identical versions, each greeted with "yeah, looks good", teach almost nothing. The system has been busy. Its understanding hasn't moved. One version that looks excellent and still fails its purpose can expose an entire missing distinction in an afternoon. The difference isn't the amount of disturbance. It's what the disturbance does to the working understanding.

Education research has a precise version of this. Variation theory holds that "discernment of the critical aspects requires students to experience variation in those aspects", and adds the part people usually forget: "variation is easier to discern when it occurs against a background of non-variance, or sameness… what is not varied is as important as what is varied."⁵ Ten films that differ in everything teach less than two films that differ in one thing. If you want an attempt to teach you where "warm" ends, change the voice and hold everything else still.

Which leaves a practical problem. If attempts supply what briefs can't, then every job generates context that didn't exist when it started: distinctions, reasons, tests. What happens to all of that when the job ends? Mostly, it evaporates. The artefact survives. The reasons usually don't.

Suppose you keep everything a film was built from: the source, the assets, the final render. Have you kept what you learned making it?

Most people would say yes. Code is the record; the record is complete. I'd have said the same a year ago. It's the wrong archive, and the reason is worth spelling out, because it decides where the rest of this book puts its effort.

Source code preserves what happened, but it may not preserve why it mattered.

The final cut of a film records the choices that survived. It can't record the alternatives they beat, or why those alternatives lost. The pause before the promise is there in the timeline; the version without the pause, and the reason it felt rushed, isn't. The local narrator is there; the imported-sounding one, and what was wrong with it, isn't. A repository of good outputs is a record of survivors with the causes of death removed.

The why lives upstream of the artefact

I made a version of this argument about software in The Prompt Is Source. "Source", I wrote, is always relative to a compiler stage. Generated code is source to the conventional toolchain, but relative to the agent that wrote it, it's already a downstream product. "Code becomes an intermediate representation only to the extent it can be regenerated from the upstream source package. Until then, parts of it are still the source of truth — whether you admit it or not." And I said where the missing part lives: "Transcripts and prompts are not 'process noise.' They are the source map — code is the what; the transcript is often the why. … When only the generated files remain, you are back to the legacy predicament — with prettier syntax."

Taste decisions are the purest case of that problem. A coding decision often leaves a trace in the code: a guard, a comment, a test. A taste decision usually leaves nothing. The reason a cut shipped lives in a rejection of a different cut, in a sentence somebody said while watching, in a judge's note about something the final version no longer does. Delete the chat, keep the file, and you've deleted the taste. What remains is what I've called trapped judgment: a decision that exists only in the output and the conversation around it, which nobody can regenerate because nobody wrote it down upstream.

Where does the missing context actually come from?

Here is the central claim of the first part of this book, in one sentence.

Not a better brief. Not a smarter model asked a better question. The process. It happens in four steps, and each step can fail in its own way, which is why the book spends the next twenty-five chapters on them.

If you do all four, the process leaves behind something more valuable than any single artefact. My thinking partner put it better than I had: the durable asset "is not just a repository of good videos. It is creative judgment connected to examples, counterexamples and outcomes." For any judged work, swap "videos" for your artefact. The asset is judgment with its cases still attached.

I'd already said the practical version, less elegantly, while it was happening:

Three destinations: tests, prompting, skill. I didn't know at the time that those three were the whole answer to where a lesson should live. Chapter 9 makes them the three carriers and argues that a lesson needs all three.

The harness is what compounds, and it has to hold more for taste

There's a close cousin of this argument in Replay-Driven Design Evolution. I asked what should still exist if the first build burned down, and answered: "If the answer is 'only some generated code,' the method failed. The durable outputs … are the harness and the history — not the first codebase." Or, more rudely: "Morphing code without a harness is amnesia with a Git log."

That holds for taste, with an extension. A design harness replays candidates against recorded reality and keeps its measures. A taste harness has to hold three things a design race didn't need. It has to hold distinctions with their conditions, because a taste lesson without its condition hardens into a rule it was never meant to be. It has to hold contrasting cases with verdicts attached, because the words alone drift. And it has to hold tests of its own judges, because in taste work the judge is often the thing that's wrong.

Anthropic's own advice on building agent skills lands in the same place from the practitioner side. Iterate with the model, it says; capture its successful approaches and common mistakes into reusable context. "This process will help you discover what context Claude actually needs, instead of trying to anticipate it upfront."⁶ Discover, not anticipate. That's the working process writing its own context, described by a model vendor.

The first habit: keep what you were told apart from what you learned

If the process is going to write context, you need somewhere to put it that doesn't blur it into the brief. The simplest habit I know is two columns, started on the first day of a job and kept to the last.

Shape The rows are the week's lessons, abbreviated; each is walked in full in Part V. Use your own job's rows. "Nowhere yet" is a legitimate entry in the last column, and the most useful one to notice.

Two columns sounds trivial. It does three jobs. It stops everything being quietly credited to the brief, or to the AI. It gives rejection triage (Chapter 8) its raw material. And it's the evidence you'll need later when someone, possibly you, tells a neat story about where an idea came from (Chapters 14 and 18).

How would you know the process worked?

One more thing belongs here, because it sets the standard for everything that follows. If a working process generates context, how would you know it actually did? The tempting answer is to look at the account it produces: the page, the principles, the essay. If the account is sharp and persuasive, surely something was learned.

That's the trap Chapter 14 names. An eloquent account can make the previous case sound inevitable without having learned anything that would help with the next one. The only test that can't be fooled that way is a successor: a fresh builder who never saw the process, handed what you kept, set to work on a company it has never seen. Does it refuse the attractive wrong answer? Does it recognise a good one that doesn't look like yours? Chapter 26 designs that test in enough detail to run it. Keep it in mind from here on, because it's what every carrier and every harvest in this book is ultimately answerable to.

Most of the context the process generates is about one thing. Not facts about the client, which the client could have told you. Not technique, which the tools already know. It's about fit: what counts as right for this audience, in this situation, for this purpose. Before we build the harness that forms and carries it, we need to say exactly what that means.

Two sentences. Read them side by side before anything else.

Both can be beautiful. Both can use the right colours, the right fonts, the right music. Both can be, in the everyday sense, on brand. Only the second closes a connection with the person watching. The first is about the company. The second is about the viewer, with the company in it.

I'm not inventing that contrast for effect. Two AI judges scored the same film on exactly those two sentences in the week this book comes from; one loved it and one marked it down, and both were right about what they were looking at. Chapter 16 walks it. Here I want the idea underneath it.

That definition does two things. It makes taste considerably richer than a catalogue of personal aesthetic preferences, which is what most people mean when they say "taste" and then give up on it. And it makes taste testable, because each of the four terms can fail on its own while the other three are fine. That's the property you need if you're going to build anything that checks it.

Four ways to fail, one at a time

Shapes These are the shapes of failure. The real instances, with their records, are in Part V.

Read the table down the third column and you'll notice none of those failures would be caught by a brand-guidelines check. They're all on-palette. They're all "professional". That's the gap this book is about.

Which "on brand" is your harness testing?

"On brand" means two different things, and most brand checks only test one of them.

When I first saw a harness reject a film on the second meaning, I was delighted, and slightly disbelieving:

"Doesn't properly match what we're trying to achieve" is the consequential meaning. A test that can only say "wrong shade of green" is a linter. A test that can say "this film is about the clinic's craft, and the viewer came here about their own tooth" is doing something closer to what a good creative director does in a review. It isn't a creative director. But it's pointed the same way.

A vendor has noticed too, with limits

Adobe's recent description of its Brand Intelligence product is a strikingly close parallel on part of this. It says the system "learns from what we call 'decision traces', or signals that reflect how teams interpret brands in practice — including comments, edits, rejected versions, and final approvals", and that its validation skill takes in "context from campaign briefs, asset types, and intended audiences" and can "pinpoint exactly where issues occur within an asset and link them back to specific guidelines or prior decisions."⁷ Its launch post names the problem precisely: brand decisions "are shaped by context, taste, and experience. Guidelines try to capture them, sometimes across hundreds of pages, but they never go far enough."⁸

That's the vendor's described architecture, not independent evidence that it works. But it matters that a company that size now describes brand validation as learned from rejections rather than written as rules. It says the second meaning of "on brand" is testable enough to build a product around.

Those three limits are where the rest of this book lives. Part III is about routing a rejection to the right place, including the judge. Part VI is about the client with no history to learn from. And the audience question stays honestly open throughout, because nothing about a harness, mine or Adobe's, substitutes for people actually responding.

Why a disagreement between judges is information

In The Exemplar Ratchet I split the criteria for judging a page into three families with different owners. Craft: the visual and verbal making. Doctrine: whether the page carries the company's actual commitments. Care: whether the experience is safe, true and non-manipulative. And I wrote that "a page has to win across families, not average them into mush."

The two sentences at the top of this chapter are a craft-versus-doctrine disagreement. The first scores beautifully on craft. The second is what doctrine asks for. A harness that averages its judges would pass the beautiful film, because one high score and one middling score average out to "good". A harness that knows the families can see what actually happened: craft won and doctrine lost. That's not a tie to be split. It's a finding.

How far can a harness go?

Not all the way, and I want to be exact about where the line is, because I drew it myself in an earlier book and I'm now moving it.

In Hidden Gates I wrote that hidden gates "guard completeness and process … They do not guard taste." Taste, I said, isn't a checklist a worker could game; it's the thing you were trying to protect. "Removing yourself from the inner loop is the design goal; removing yourself from the judgement of quality-as-such is drift."

The films moved that line, but only partly. A harness can now operationalise some failures of fit as inspectable tests. Does the craft connect to something the viewer notices? Is the voice cast from the audience's place? Does the opening assume knowledge a stranger won't have? Each of those can be asked of a film, by a judge or a cold viewer, and answered with a time and a reason. What a harness still can't do is decide that something is good. Promotion to "this is the standard now" stays with a person.

Anthropic's write-up of its own design harness arrives at the same boundary from the other side. "Aesthetics can't be fully reduced to a score — and individual tastes will always vary," its author writes, "but they can be improved with grading criteria that encode design principles and preferences. 'Is this design beautiful?' is hard to answer consistently, but 'does this follow our principles for good design?' gives Claude something concrete to grade against."⁹ Swap "beautiful" for "on brand" and you have the move this book makes.

Why fit is testable at all

There's a reason this works, and it's worth knowing because it tells you what a good test looks like. Experts don't usually decide things by reciting rules. They recognise. Kahneman and Klein, who disagree about a lot, agree on Herbert Simon's definition: "The situation has provided a cue: This cue has given the expert access to information stored in memory, and the information provides the answer. Intuition is nothing more and nothing less than recognition." And crucially: "valid cues must be specifiable, at least in principle — even if the individual does not know what they are."¹⁰

That's the whole trick. When I rejected a voice as "crappy compared to the rest, seems like older tech or something", I was responding to a cue I couldn't name; "or something" is the tell. A judge that could actually listen named it later: an incongruous British accent (Chapter 17 has the whole story). Once a cue has a name, it can become a test. Making taste testable is mostly the work of finding the cue a person is responding to and writing it down in a form someone else can check.

Which raises the question the next chapter answers. If the cue can't be written down in advance, if even the person responding to it can't name it, where does the named version come from? It isn't copied out of anyone's head. It gets made.

"See how we did that." It's what a master says to an apprentice with a hand on their shoulder, standing in front of something that went right.

I found myself saying a version of it to an AI that week, without noticing I was doing it. When I described the method afterwards, it came out like this:

The question is what exactly you take away when you "see how we did that". Because there are two very different things on offer, and most people grab the wrong one.

Two things you can extract from a session that went well

The first is a recipe. The second is what an apprentice actually takes away from a good master, and it's why apprenticeship works for skills that recipes can't carry.

The first reproduces a procedure. The second can support competent variation.

The sequence is easy to extract and nearly useless on the next client, because the next client's film won't want the same scenes. The body of judgment is hard to extract, and it's the only part that travels.

Why the rejected versions matter more than the final one

Here's the part people miss. The final artefact contains the choices that survived. It cannot, by itself, explain the alternatives they defeated. If you only keep the winner, you've kept the answer and thrown away the question it was answering. The production history supplies the missing contrast: this voice rather than that one, this opening rather than the one that assumed too much.

Learning science has known this for a long time. Schwartz and Bransford ran studies where students analysed contrasting cases before being taught the underlying idea, and found that "noticing the distinctions between contrasting cases creates a 'time for telling'; learners are prepared to be told the significance of the distinctions they have discovered."¹¹ Both halves mattered. Contrasts without an explanation didn't get students to the principle. An explanation without the contrasts didn't stick. The contrasts first, then the telling, beat either alone. Hold on to that order: it's also why the supervising agent's explanation in Chapter 1 landed after the work.

The cognitive-apprenticeship literature says it about skills directly. Some concepts can be formally described, "but many of the crucial subtleties of their meaning are best acquired through applying them in a variety of problem situations. Indeed, it is only through encountering them in real problem solving that most students will learn the boundary conditions and entailments of much of their domain knowledge."¹² And the same paper describes the move that turns tacit into usable: "By becoming critics as well as producers, students are forced to articulate their knowledge … once articulated, this knowledge can no longer simply reside in tacit form." Producer and critic. Hold that pair in mind for Part III, where it becomes a skill and a harness.

After the task, or while doing it?

I've written about extracting lessons from work before. In Worldview Recursive Compression the loop is generate, evaluate, extract, encode: after each output, you judge it, pull out the lesson and put it back into one maintained kernel of your judgment, so "the lesson goes back into frameworks.md, not just into the output." Fix the kernel, not the output, and the lesson helps forever.

That's still right. But it describes extraction after each output, from a worldview that already exists. What happened with the films was earlier and messier than that.

You are not only extracting knowledge after the task. You are developing the knowledge needed to complete the task, while completing it.

And the knowledge doesn't go back into one kernel. As Part III argues, it goes to three places that do different jobs.

There's an older cousin of this in A Blueprint for Future Software Teams, where I described mining AI coding conversations for repeated corrections and rationale, "giving your system dream time", and turning them into rules. The same chapter carries a warning I still stand by: "AI may misunderstand context, over-generalize from specific cases, or miss nuance. … Treat AI extraction as proposals, not decisions." That mining happens after the fact, over conversations that weren't set up to teach anything. "See how we did that" is different in kind: it's during the work, and the contrast is made on purpose. Keep the warning, though. It comes back in Chapter 11 as the superstition trap.

Where does the principle come from, if nobody had it?

This is the part I didn't expect. I'd assumed the process was taste transfer: my judgment, moved into the system. Some of it was. I had opinions before the session started. But some of the distinctions that ended up in the skill weren't in my head at the start in any form I could have written down. They became clear through the collaboration.

Neither party necessarily possessed the fully articulated principle at the beginning. That's the important sentence. It means "capture my taste" is only half the story.

I'd already named the middle step in Demonstrate Over Tell, where I called it assisted articulation. When someone rejects something, "the person may not have a fully formed explanation waiting to be recalled. They may only have a reaction. The machine supplies candidate interpretations of that reaction … and the person recognises, combines or corrects them." I added a guard: "The proposed reasons should help people find their meaning, not quietly replace it with the model's framing." In taste formation, that step does more than help me say what I meant. Sometimes it's where the meaning first gets made. The guard matters more, not less.

None of this would surprise a design researcher. Schön wrote that design knowledge "is knowing in action, revealed in and by actual designing. It is mainly tacit … designers know more than they can say, they tend to give inaccurate descriptions of what they know, and they can best (or only) gain access to their knowledge in action by putting themselves into the mode of doing."² It's Polanyi's old line, "we can know more than we can tell", with a practical consequence attached.¹³ If you can best reach your judgment by doing, then the doing has to be set up to catch it.

It also explains a gap in one of my own frameworks. The Exemplar Ratchet freezes a win and extracts its direction of improvement. It needs a win to freeze. Formation is what happens before there's any win at all, while you still don't know what "good" is for this audience.

That list is the raw material for everything in Part III. But formation leaves one particular residue before any of it: short phrases. "Wow." "You see me." Words that carry far more than they say, and that are worth almost nothing until they've been made precise.

I had three phrases I kept using that week. "Wow." "Oh, you see me." "Let the promise land."

They'd gone into the prompting early, because that's how I talk about what a film should do. Builders were steering by them. And partway through, I realised I didn't know exactly what they meant, which meant nobody else did either. So I asked the build agent to fix that:

Where the answer came from, and what else came with it, is in Chapter 18. This chapter is about what those phrases are worth, and what they need before they're worth anything.

What's a short phrase actually for?

Short phrases are powerful. "Let the promise land" lets me steer a two-minute sequence of decisions with four words. A director can say "more air" and a whole crew knows what to do. That's the upside, and it's real.

The downside is that a short phrase only works when it points into something: a set of worked examples, contrasts, "this, not that". Without that grounding it's just an adjective with ambitions. Hand "wow" to two builders and you get two unrelated films, one full of spectacle and one full of a clever reveal, and both of them can claim they delivered wow. Adjectives in a brief are where fit goes to die, because every failure complies with them.

The short phrase is the handle. The accumulated discrimination is the asset.

My thinking partner's first pass at mapping the handles to what they should mean was already a big improvement on the words alone:

And one sharpening of "wow" that I've kept ever since: the valuable wow is not merely "that looked expensive". It's "that understands my situation — and now I see what this would do for me." The spectacle has to carry the promise, not compete with it. A film can be gorgeous and have no wow at all in that sense. A plain film can have plenty.

The four terms, made precise

The version that actually went into the skill and the harness went further. Each handle became a term with three parts: what it is, what it isn't, and a test someone other than me can run.

Definitions and tests as they were written into the brand-film skill and its judge that week, lightly trimmed; the "isn't" for being heard is my wording of the distinction Chapter 24 develops.

Reading the table, not just looking at it

Three things about that table are worth more than any row of it.

The middle column does most of the work. Every term has a near miss, and the near miss is exactly the attractive wrong answer a builder would otherwise produce. Ask for recognition without saying what it isn't, and you get a smiling stock person and a narrator saying "we understand". Ask for demonstration without the still-life clause, and you get a beautiful tool beside a beautiful result, with nothing happening. The "isn't" is where the taste lives, because it's where the rejections came from.

One row wasn't in my vocabulary at all. I'd never said "being heard". It turned up during the work as a stronger form of recognition: being recognised is seeing your situation on screen; being heard is seeing your own words change what the service does. Nobody briefed it. It's formation, from Chapter 5, leaving a trace in the dictionary. It became the most important idea in the whole series, and Part VII is built on it.

Every test is a question a stranger can answer. None of them is "does this feel right to Scott?" They're phrased so a cold viewer, human or model, can be asked them and their answer checked. That's what lets the same distinction live in two places, telling a builder what to make and telling a reviewer what to look for. Chapter 10 is about why that double life matters.

Where the Taste Kernel came from

In The Cognition Dimension Ladder I argued that the real moat in AI-assisted decisions isn't the engine; it's the evaluation function the engine scores against. "Taste is not just human magic. You compress it into examples, anti-examples, rules, rejected patterns, preferred language — and then you apply it systematically." I called that documented judgment the Taste Kernel, and said it stays honest only by closing the loop with the real world.

That describes taste once it's documented. The four-term table shows how a kernel entry gets made. It starts as a loose handle. Rejections sharpen it into a definition. The near misses become its anti-examples. A test gets attached that can fail. The Taste Harness is the Kernel's missing upstream: the place where entries are formed before they're filed.

Why "wow" alone is a slogan

There's a test I wrote for design patterns in Generative Design Patterns that applies directly here. Give only the phrase to two independent builders. Do their results share a recognisable invariant and differ mainly in local fit? Then it's a pattern. If they share only a buzzword, it's a slogan. "If independent expansions share structure, it is a pattern. Heat without invariant is fashion." And the fix I proposed then is the fix now: "Do not rescue a slogan by adding adjectives … Add invariant structure … or admit you are still naming a desire."

"Wow" handed to two builders is a slogan. "Aha, with this test" is a pattern: two builders will make different films, and a third party could still say what both were trying to do. The structure that turns one into the other is always the same three parts: definition, near miss, test.

And the reason it works goes back to Chapter 4. Expert judgment runs on recognition of cues the expert often can't name. A test is a cue that someone finally named.

Once a distinction can be tested, it can fail a test. And the moment something fails, the harness has to answer a question most loops never ask: what, exactly, failed? The thing we made, the way we checked it, or the distinction itself?

Every AI workflow I've seen has a loop in it somewhere. Make something, check it, fix it, check it again. It's the most useful pattern in the field, and it has a blind spot built into its grammar.

The blind spot is the word "it". Fix it. Check it. The loop assumes there's one thing being improved: the output. But when a piece of judged work fails, there are three different things that might need to get better, and they answer three different questions.

I'll call these the three questions of improvement. (Readers of A Blueprint for Future Software Teams will remember a different set of three questions, about sticking points, surprises and never-agains, for extracting a lesson at the end of a ticket. Those are for pulling a lesson out. These are for locating which thing a lesson improves. Both are useful; they do different jobs.)

Each question has its own product, owner and evidence

The reason to separate them is practical. Each one produces something different, can only be answered by particular people, and has its own evidence that it worked.

Look at the last column. Only the first row's evidence is visible on the day. The other two only show up later: when the old failure is re-run against the new judge, or when a different job goes differently because of a principle. That's exactly why most loops skip them.

Why most loops only ask the first question

The artefact is visible, urgent and billable. The client is waiting for the film. The instrument and the principle are invisible until they fail again, usually on a different job, usually when someone else is doing it.

But look at what each one is worth afterwards. The artefact is spent on this job. The instrument and the principle carry to the next one. A loop that only asks "make" pays for taste on every job and never compounds. Chapter 1 said every iteration makes two products, the artefact and the account of what to keep. This is what the account is made of: a better way to tell, and a more precise thing to say.

That's also why the speed of these loops matters so much now. When a cycle runs in hours inside one job, you can ask all three questions several times before the job is done. The escalation I described in Chapter 1, getting better at the thing while doing the thing, comes from the second and third questions, not the first.

One rejection can move through all three

Here's where it gets interesting. A single failure doesn't have to belong to one question. It can travel.

Take a voice rejection, in its tidy form. At first it looks like a production defect: replace the voice. Then it exposes a perception defect: the judge was assessing the narration from a transcript, so it couldn't hear the quality being criticised. Then, if you keep going, it exposes a gap in the principle: the skill never distinguished the narrator's role from the customer's voice, or said what relationship each one should establish with the audience.

One rejection can improve the product, the instrument and the principle.

That's the tidy version. The real voice rejection from my week was messier, spread across four separate events in one long day, and I'll tell it straight in Chapter 17. The cleanest real case, with the actual skill text before and after, is Chapter 19. For now the claim is the important thing: if you stop at the first question, you get a better film and keep the blind judge and the missing principle. The next film fails the same way, and so does the next client's.

"Isn't this just iteration?"

It's the obvious objection, and it deserves a straight answer, because it's half right. Yes, it's a loop. No, it isn't the same loop.

A loop driven by a score, pass or fail, seven out of ten, can only move the artefact. A number carries no information about whether the judge could see the problem, or whether anyone knows what "better" means here. It tells the builder to try again, and nothing else. A loop driven by diagnosis in words can move all three, because a reason can be about the artefact, about the instrument, or about the principle.

The research has caught up with this. GEPA, a prompt optimiser from a group of university and industry researchers, argues that "the interpretable nature of language often provides a much richer learning medium for LLMs, compared to policy gradients derived from sparse, scalar rewards." It reflects on what happened in each attempt "to diagnose problems, propose and test prompt updates", and it distinguishes two kinds of trace. The text the system produced is the execution trace. The text the environment produced while evaluating it is the evaluation trace: "Beyond reflection on execution traces, we identify a second valuable source of diagnostic information in the evaluation traces."¹⁴ Execution trace and evaluation trace: that's the artefact and the instrument, in a paper about prompt optimisation.

Work on reward models points the same way. Training a judge on binary verdicts makes it "susceptible to guessing correct outcomes without sound critiques", which puts noise into everything learned from it.¹⁵ An earlier study concluded that "judgments hold greater potential than rewards" for aligning models.¹⁶ In plain English: a "no" without a reason lets the judge be right for the wrong reason, and gives you nothing to route. A "no, because…" can be sent to whichever of the three questions it belongs to.

What this adds to an older loop of mine

In Worldview Recursive Compression I wrote that "when you generate a good output, it validates your framework. When you generate a bad output, you learn what's missing." Still true. But "what's missing" now has three possible addresses. It might be missing from the artefact. It might be missing from the way we checked it. It might be missing from what we think "good" means.

The second address is the one that loop assumed away. Its evaluate step treated evaluation as reliable: you look at the output and you can tell whether it's good. In taste work that's often exactly what fails. The judge was reading a transcript. The "cold" viewer had been told it had already seen the ads. The brief told the reviewer where the film would be shown, and the reviewer forgave it accordingly. A bad output that exposes a blind judge teaches you something no amount of fixing the output would.

Knowing there are three questions doesn't tell you which one a particular rejection belongs to. A reviewer's "no" arrives as a single word, or a single sentence, and it doesn't come labelled. Sorting it takes a deliberate step before anyone touches the artefact. That step has a name.

The step is triage, and I didn't invent the idea of routing a rejection. I wrote the first version years before these films, and it's still right as far as it goes.

In Look Mum No Hands, writing about AI proposals that people approve or reject, I said: "Rejection isn't failure—it's feedback. And how you capture rejection feedback determines whether the system learns." Different rejections point to different fixes. "Wrong person" is a data problem: the underlying record is wrong. "Wrong action" is a playbook problem: the decision rules need updating. "Wrong timing" is a context problem. "Wrong tone" is a template problem. And the principle that holds them together: "feed back at the right layer. If the playbook is wrong, fix the playbook—don't just reject this one card and leave the playbook broken."

I reused it in Demonstrate Over Tell, where a menu of candidate reasons makes the routing almost automatic, because each reason can carry its layer with it.

What the old routing quietly assumes

Read those four layers again. Data, playbook, context, template. Every one of them is about how the artefact got made: what it was made from, what rules made it, what it knew, how it was worded. The routing assumes two things it never says out loud. It assumes the person who said "no" saw correctly. And it assumes everyone already knows what "right" would have looked like.

For operational proposals, those are usually safe assumptions. For taste work they're the two that break first. The judge often can't see the problem, or has been pointed at the wrong viewer. And "right" is frequently the thing nobody could say yet. So the routing needs two more layers.

The first four are all answers to "make". The last two are "tell" and "say", from the previous chapter. Triage is simply asking which row a rejection belongs to, and allowing the answer to be more than one row.

What does this failure teach us: a defect in the film, a defect in our perception, or an incomplete account of what we want?

That question, which came out of my conversation with the thinking partner after the week, is the whole of triage. Swap "film" for whatever you make. Ask it before anything else happens.

Why triage has to come before the fix

This is the rule people break most, because the fix is right there and the client is waiting.

How do you tell which leg it is?

Each leg has its own diagnostic questions. Run all three sets; more than one can come back yes.

That last question is the best tool I have for the principle leg, and I borrowed it from my own work on relationship files. In The Manila Folder, writing about offering people candidate labels for themselves, I noticed that "'That isn't me' is easier to say than 'this is me,' and it points the same way", and gave one instruction: "When they reject one, ask what the rejection protects. That sentence often belongs in the retelling."

Apply it to taste. "I don't like that voice." What does the refusal protect? Not "good audio", which is what it sounds like. It protects the relationship between the speaker and the audience: whose voice this is meant to be, where it's coming from, how close it stands. That sentence is the candidate principle. Sometimes the rejector can say it directly if you ask. Usually the model can propose it, and the rejector recognises it or refuses it (Chapter 5's formation loop, doing its job).

The instrument leg is a learning event

I want to be careful about the instrument leg, because there's a whole book's worth of argument about why AI judges miss what they miss: what they can perceive, what they're fed, how to measure their range. That's the sibling book, Code as Proxy, and I won't re-argue it here. For triage, the instrument leg asks a narrower question: what did this rejection teach us about how we tell?

The answer is often surprisingly mundane. It isn't always the model. It's the brief the judge was given, which mentioned where the work would be shown. It's the persona the "cold" viewer was handed, which gave away the answer. It's the builder's own habit of auditioning voices by whether the transcript comes out clean, which can't hear quality at all. Those are instruments too, and they're the ones people forget to suspect.

The discipline for this leg is one I wrote down in Your AI Partner Is the Challenger: "If a failure reveals that your scoreboard cannot see the real defect, that is not only a design loss — it is a measure discovery. Record the date and the motivating case." Version the evaluator. Keep the case that taught you.

The triage card

Here's the artefact. It's the most-used form in the book, and every lesson in Part V passes through one.

And one card filled in, from outside the films, so it's usable before you've read Part V.

Triage locates. It doesn't promote.

One guard, before the next chapter gives lessons their homes. A principle that comes out of a single rejection is a candidate. In Demonstrate Over Tell I put it this way: "A picked reason is evidence about this moment, not a permanent preference. … Don't promote a single situational rejection into a rule about the person." Triage tells you where a lesson belongs. It doesn't tell you the lesson is true everywhere. Chapter 11 is about what happens when that distinction gets lost.

The messiest real three-leg case from my week, a voice, is in Chapter 17. The cleanest, with the skill text and the judge's questions before and after, is Chapter 19.

Triage tells you where a lesson belongs. It doesn't give the lesson anywhere to live. The default home, a new line in the prompt, is the wrong one, and the next chapter explains why it takes three.

Picture what happens to a prompt after twenty rejections have been "harvested" into it.

It's two pages long. It opens with a sensible paragraph about purpose and audience, then turns into a list: always do this, never do that, make sure you, avoid. Half the lines were written in a hurry after a single bad afternoon. A few contradict each other. The builder that reads it has become timid; its output is safer and blander than it was at the start. And not one of those twenty lessons can be checked, because a line in a prompt is an instruction, not a test. Longer prompt, weaker lessons.

I'd already learned this the hard way, on websites, before the films started. When the build agent drafted its list of lessons for the general skill, it reminded me of my own earlier finding in one line: "prescriptive rules made results worse." Its proposal was the opposite of a rulebook: each lesson as "the principle, why, and a test, not rules."

The three carriers

When I asked, near the end of the week, how to retrofit a lesson into the prompting, the build agent described exactly this, in practice rather than theory. The skill carries a lesson "as craft: a principle, the real example and your words, and a test." The judge's rubric "carries the same test, so a film that misses it gets named with a time." Code "measures it wherever it can be measured." And one sentence I've since adopted as policy: "The briefs stay short north stars and never grow rules. Your verdicts are recorded as test labels, so each new judge can be checked against what you actually said." Chapter 19 shows that answer in full, with the skill text before and after.

What goes wrong with only one carrier

Why the contrasting cases are the carrier everyone forgets

Most teams have some version of the first two: a style guide and a review. Almost nobody keeps the third deliberately. It's the one that makes the other two mean something.

A positive example shows one good answer. It doesn't tell you which of its hundred features made it good. A pair does. The same opening with and without the person's own photo. The same line in a local voice and an imported one. The contrast points at the one thing that changed, and says: that. Chapter 2's variation theory is the principle; this is the carrier built on it.

The effect is measurable in language models too. Researchers at Intuit, trying to get models to write "in a certain style or tone that is hard to describe", found that giving the model "positive examples that illustrate the true intent, along with negative examples that show what characteristics we want LLMs to avoid" significantly beat ordinary few-shot prompting.¹⁷ A style that's hard to describe is exactly what taste is. You convey it better with this, not that than with a description, or with good examples alone.

There's a deeper reason the cases can't be dropped. Schön warned that "symbolic, procedural representations of tacit design knowledge are bound to be incomplete or inadequate in relation to the actual phenomena of designing."² The skill's words will always be less than the judgment that formed them. The cases carry the remainder.

Rule one: reinforce without converging

The three carriers have to agree. They mustn't become the same thing.

Here's how convergence actually happens. The first good film has a device that works: a note filling in on screen with the customer's own words. The skill describes the principle behind it. The cases keep the film as the example. And then someone writes the harness test as "does the film have a consultation note?" instead of "where did the person's words change what happened?" Now the harness is testing the example, not the principle. Every future film needs a note filling in on screen, whether or not it fits, and a genuinely better device will fail review for being different.

I've drawn this line twice before. In The Exemplar Ratchet it was the difference between extracting the shape and extracting the wallpaper: test on a different world against the same bar, and "if they only win when the wallpaper matches, you extracted wallpaper." In Demonstrate Over Tell it was the template trap: one good expression copied everywhere. "Aim at the reader's need, not at an aesthetic." "The family resemblance should be visible value, not identical chat bubbles." "Specify facts; nudge taste." The carriers version of it is simple to say and easy to break: test the principle; use the example only to show what the principle means.

Rule two: reinforce without duplicating

The opposite failure is quieter. Write the principle three times, once in the skill, once in the rubric, once in the notes beside the examples, and within a week you have three versions. Somebody tightens the rubric's wording after a judge misreads it. Somebody else adds a clause to the skill after a builder over-applies it. Now the builder is taught one thing and tested on another.

I solved the same problem for knowledge bases in The Wiki Is the Kernel. Named documents that restate each other drift apart; the fix was to demote them to "boot profiles": an entry point and a selection bias over one shared set of pages. "Nothing is copied, so nothing can drift." In The Wiki Playbook, the same idea: "There is exactly one register page"; every profile points at it.

For taste: one canonical statement of each distinction, with its conditions, and three carriers that point at it and use it differently. The skill uses it to make. The harness uses it to check. The cases use it to show where it holds and where it doesn't.

The routing record

Put together, each lesson that survives triage gets one record. It's the storage form of Chapter 3's "judgment connected to examples, counterexamples and outcomes", and of the "rejections with reopening conditions" that Replay-Driven Design Evolution treats as compounding assets.

Test the principle. Use the example only to show what the principle means.

Look at what the three carriers do to a single judgment. The skill puts it in front of the builder: this is what to make. The harness puts it in front of a reviewer: this is what to look for. Those are two very different working positions for the same idea, and the second one turns out to matter more than it looks.

What surprised me most that week wasn't that my rejections changed the prompts. It was where else they turned up.

"Obviously" is doing a lot of work in that sentence. Everyone expects feedback to end up in the instructions. What I hadn't expected was to find my judgment on the other side of production too, sitting in the checks, rejecting films I'd never seen. (How that particular voice lesson came to be is Chapter 17's story. Here it's just the example of a judgment living in two places.)

Two positions for one judgment

Those are complementary uses of the same judgment, and the second is the one that does the protecting. An instruction can be satisfied by intention alone. The builder can sincerely believe it made warmth, recognition, clarity; it read the instruction, it tried, it will tell you so. The reviewer's job is to inspect what was delivered, not what was meant. Without the test, every principle in the skill is on the honour system, and builders, human or AI, are generous graders of their own intentions.

From one rejection to a way of examining everything

The move that puts a judgment into the test has a shape, and it's worth drawing because it's the opposite of how most feedback gets stored.

Walk it with the voice, one step at a time. "I don't like that voice." Then the mismatch: the delivery puts the speaker in the wrong relationship with the audience. Then the principle: cast the voice from the audience's place. Then the examination: a listening check that asks whether this voice sounds like it belongs where the audience is.

Compare that with how most teams store the same event: a note somewhere that says "Scott doesn't like voice seven". That note can't examine anything. It can only forbid one voice, forever, for one person's reasons, which nobody remembers.

The last step is where most of the value is, and where the principle has to be stated carefully. The portable lesson isn't "always use an Australian voice". It's to consider where the audience is, who the speaker is meant to be, and what relationship the delivery should establish. A local accent is one application of that. For a service sold nationally, or a founder who speaks for their own company, the same principle picks a different voice. The test has to check the principle, not its first application; that's Chapter 9's first rule, doing its job.

Declared, then perceived

The cleanest way I've found to put a principle into the test is to make the builder commit to it before anyone checks. The builder declares, with times, where it meant the recognition to land and where it meant the viewer to understand how the service works. Then the harness asks a viewer who has seen nothing but the film to find those moments, and compares.

That turns "recognition in the first few seconds" from something the builder believes it delivered into something the harness can hold it to. The method itself, including why asking a viewer "did you have a moment?" tells you almost nothing while asking "did you find the moment we declared, at the time we declared it?" discriminates, belongs to the sibling book, and I'll leave it there. What matters here is what it's for: it's the place where the author's judgment lands inside the harness.

The instrument learning to state the standard

There's a subtler thing that happens once judgment lives in the test. The test starts to improve its own vocabulary.

The judges' rubric that week had a criterion called "connection". It was marked down often, and when the build agent looked at why, it found the criterion was mixing two different things: whether the viewer recognised their own situation, and whether they understood how the service worked. Those can come apart. A film can make you feel seen and leave you no wiser; another can explain beautifully to someone who never felt it was about them. So "connection" was split into recognition, being heard and the aha, each with its own test (the definitions are in Chapter 6's table). The standard didn't change. It got stated more precisely, and so it got observed more precisely.

Criteria shape the work before anyone judges it

Anthropic's write-up of its design harness has two lines I think are under-read. Its author "wrote four grading criteria that I gave to both the generator and evaluator agents in their prompts." Same criteria, both sides of production, exactly as above. And then the surprising finding: "Even on the first iteration, outputs were noticeably better than a baseline with no prompting at all, suggesting the criteria and associated language themselves steered the model away from generic defaults before any evaluator feedback led to further refinement."⁹

Read that carefully, because it cuts both ways. A criterion is a production input, not only a test. Stating what the judge will look for changes what gets made before any judging happens. That's the power of putting judgment on both sides. It's also the danger: whatever quirks the criterion's wording carries, the builder inherits them. Chapter 11 is about that danger.

A practitioner who builds LLM judges for a living, Hamel Husain, puts the prerequisite bluntly: "You cannot write a good judge prompt until you've seen the data." And when domain experts are made to decide pass or fail on real outputs, "they are compelled to articulate their expectations clearly. This process uncovers nuances and unspoken assumptions."¹⁹ That's Chapter 2 again, from inside the harness: the test's criteria are formed by judging attempts, not written first.

A judge for comparison, a stranger for truth

In Demonstrate Over Tell I drew a line between two instruments that people confuse. A fixed judge, holding the same criteria while you vary the work, is genuinely useful for comparing versions. But its score "is not evidence that a reader understood anything. It tells you what the judge rewards." Hence the rule: "Keep the judge for comparison. Add a comprehension test for truth."

For taste, both instruments should be fed from the same distinction. The judge compares versions against it. The cold viewer's questions test whether it landed on someone who knew nothing. One source, two instruments, which is Chapter 9's no-duplication rule again, in the harness this time.

Shape Four rows from the week, abbreviated. "None" in the last column is common and honest: most of fit can't be measured by code, only asked about.

Put judgment on both sides of production and something happens that feels like success at first. The tests start saying no. Often. On things no linter would ever catch. Films get stuck. Is that the harness working, or the harness going wrong?

Two things I said at the end of the week, a minute apart, sum up where this chapter starts.

The first is delight: a test suite blocking on taste, not bugs. The second has the right instinct buried in its last two words. "I guess." A stuck film might be the harness doing its job. It might also be the harness quietly teaching the builder the wrong thing. From the outside, the two look identical.

When the bottleneck moves

Something changes once a system can make a plausible artefact faster than it can establish that the artefact deserves to ship. The hard question stops being "can we manufacture this?" and becomes "have we made the right experience?" At that point the harness isn't a unit-test suite any more. It's a creative-review process, with everything that implies: judgment calls, taste disagreements, the risk of a reviewer whose preferences quietly become everyone's.

That's the right place for it to be. It's also why the rest of this chapter matters.

Three reasons repeated rejection can mean nothing

Evaluators steer

The third reason is the one that deserves the most suspicion, because it looks like progress. Scores go up. Rounds get shorter. The work gets steadily more like one particular thing.

Anthropic's design-harness write-up reported exactly this: the wording of its criteria, phrases like "the best designs are museum quality", "pushed designs toward a particular visual convergence."⁹ I made a meal of that line in Demonstrate Over Tell, so I'll leave it as one clause here. The general pattern has been measured more formally. Optimise hard enough against an imperfect proxy for human preference and "optimizing its value too much can hinder ground truth performance, in accordance with Goodhart's law."²⁰ And creative work done with AI help has been found to score better individually while losing variety collectively: stories written with AI ideas "are more similar to each other than stories by humans alone… an increase in individual creativity at the risk of losing collective novelty."²¹ One artist who tested generic AI aesthetic judges against her own taste found they "felt overly general rather than tailored to their aesthetic sensibilities."²²

For most work, homogenisation is a side effect. For a brand, it's the whole failure. A brand's job is to be this company. A harness that nudges every client's films toward the same evaluator-pleasing look has made each film better and every brand worse.

The opposite failure exists too, and it's worth one sentence so nobody reads this chapter as an argument for lenient judges. The same Anthropic author, early on, "watched it identify legitimate issues, then talk itself into deciding they weren't a big deal and approve the work anyway."⁹ Strictness isn't the enemy. Undiagnosed strictness is.

Who guards the judges?

Two of my own frameworks already had pieces of the answer.

In The Exemplar Ratchet I named critic-taste drift: "Promotion to exemplar stays with a human. If the loop's own critics decide what counts as greatness, the ratchet drifts toward the critics' taste. That is specification gaming with a friendly face." And in the same book: "If you optimise against a twelve-point rubric, you will get pages that score well on every point and are bland. Use the rubric to diagnose. Decide between pages with a single question: which would you rather receive?"

In Your AI Partner Is the Challenger I made the rule about roles: "Challenger is a feature. Arbiter is a procedure." And: "Enthusiasm is not ship. Scepticism is not kill. Both are prompts to instrument." An enthusiastic judge's high score and a harsh judge's low one are both challenges. Neither is the verdict. Chapter 16 shows a real pair: one judge called a film a masterclass, another scored it six, and the useful thing was the reason attached to the six.

The triage question preserves the value of strict review without enthroning the reviewer.

Asking "is this a defect in the artefact, in our perception, or in our account of what we want?" treats every rejection as evidence to be located, not an order to be obeyed. A harsh judge gets taken seriously and still gets checked.

There's an obvious objection, and I'd rather meet it than have you think it. Isn't a harness that rejects on taste just your taste, enforced? It would be, if verdicts skipped triage and conditions. Under this method, my own verdicts don't become edicts. They become test labels: a judge worth keeping should agree with them, and when it doesn't, that's a finding about the judge or about me. Even the author is a challenger.

The superstition trap

Now the failure that worries me most, because it's the one that hides inside success.

In 1948, B. F. Skinner fed pigeons at fixed intervals regardless of what they did, and watched them develop rituals. One turned anticlockwise; another thrust its head into a corner. Whatever each bird happened to be doing when food arrived, it did more of. "The bird behaves as if there were a causal relation between its behavior and the presentation of food, although such a relation is lacking. … A few accidental connections between a ritual and favorable consequences suffice to set up and maintain the behavior in spite of many unreinforced instances."²³

Skills do exactly this. One voice gets rejected, and the lesson written down is "no voices like that", so every similar voice is banned forever. One pause lands beautifully before a promise, and pauses become compulsory everywhere. One film works with a note filling in on screen, and on-screen notes become house style. Each lesson was true once. None of them carried the condition that made it true, so each one became a ritual the next builder obeys without knowing why.

I've met this failure in other forms and written about it each time. In The Wiki Playbook it's calcified lore: "A conditional pattern hardens into an unconditional rule… The tell: a claim with no conditions attached that everyone obeys." In Institutional Memory Not Cognition it's what you get when you skip the stale-context check, asking what conditions made the original judgment true and whether they still hold: "Without this check, you do not get institutional cognition. You get institutional superstition with better search." In Made From You, about a founder's judgment: "We need to know when a judgement applies and what evidence would overturn it. Otherwise a sensible lesson becomes a rule applied everywhere." And in Demonstrate Over Tell, the first warning: a picked reason is "evidence about this moment, not a permanent preference."

The guard is conditions, not force

The usual instinct, when a lesson seems important, is to state it more strongly. Always. Never. Must. That's exactly backwards. In The Perturbation Review I built a ladder for how far a lesson from one engagement is allowed to climb, and the lowest rung carries the most important sentence: a local lesson is "true here, stays here. A success, not a shortfall." Promotion needs evidence from elsewhere: "Recurrence nominates. It does not promote." And the rule for a single striking case: "n=1 may climb, and it arrives wearing its ceiling."

A principle becomes more useful as its conditions become clearer, not simply as its wording becomes stronger.

That's Part III. Three questions, triage, three carriers, judgment on both sides, and conditions on everything. One thing it doesn't explain is the thing that started this book: why the agent that watched all of this wrote the best account of what the work was for. That needs a different kind of chapter.

Ask the question twice. "Why are we making these films?" Once at the start of the work, once at the end. Same agent, same instructions. Look at what each answer has to work with.

At the start, the question invites a textbook answer, and a good model gives a good textbook answer. By the end, the textbook answer is visibly inadequate. It can't explain why the beautiful film failed, or why the voice was wrong. An explanation has to account for those contrasts or it's no explanation at all. So the answer gets better, not because the writer improved, but because the problem did.

My thinking partner's phrase for what the supervisor had acquired is the best one I've found: a pattern of decisions that needed a coherent explanation.

Theory formation from practice

Put that as a cycle and it becomes something you can run on purpose.

Making improves the explanation; the explanation improves the next making.

I described it at the time without any of that vocabulary, with a fair bit of surprise in my voice:

Why ask the top of the loop, not a worker?

There were several agents involved that week: builders, one per film; judges; a cold viewer. The one I asked for the page was the supervisor, the agent that dispatched the builders, read the judges, relayed my verdicts and ran the fix rounds. That turned out to matter.

A local worker knows the repair it's making. It knows its own film had music too loud under the voice and it turned the music down. The supervisor has access to something no worker has: the relationships among repairs, the original purpose, my interventions and the critics' disagreements. Its valuable job isn't to combine the workers' reports into a summary. It's to notice that several apparently different problems are one failure. Three films marked down for different reasons (posed tools, a static end card, generic imagery) might all be failing the same principle: nothing on screen is about the viewer's own situation.

The practical rule follows directly. If you want the account, ask the agent that held the whole run, at the end of the run. Don't hand a summary to a fresh writer and ask it to explain. A summary is exactly the thing that throws the contrasts away.

The same model, a different problem

This connects to an argument I made in The Inference Field about why two context windows of the same size can be "cognitively nothing alike". One built by retrieval is a pile of plausibly relevant material, with the work of finding structure left to the model at the last moment. One built by walking is "the serialised trace of a system progressively constructing its own understanding", where "every step in that trace is still in the room, still conditioning the next one." The supervisor's context at the end was that second kind. Every build, every rejection and every repair was still in the room when I asked the question.

And it connects to a distinction from Experience Is Compressed Priors: "Domain priors change what you try. Process priors change how you try." I described process priors as "how you get better at asking that second question with an AI that can supply historical structure, while you supply direction, taste, and the recognition of what is still wrong." The page the supervisor wrote is a domain prior: what these films are for. The way it got written, asking the informed supervisor to account for the whole run, is a process prior. Keep both. The second is the one that will help on the next job.

The research cousin

There's a close technical parallel in GEPA, the prompt optimiser from Chapter 7. It "samples trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons." When traces are "paired with the final outcome of the system (e.g., success or failure), they provide substantial diagnostic value, allowing practitioners to trace errors or successes back to specific decisions." Language models, it says, can use these traces "to perform implicit credit assignment, attributing responsibility for the final outcome to the relevant modules."¹⁴

That's the supervisor's job, described for a benchmark. I want to be precise about what the parallel does and doesn't show. It shows that detailed experience, reflected on in language, supports better updates than a bare score. It doesn't show that anything has solved taste: GEPA's tasks have scores you can compute, and fit, mostly, doesn't. And the learning-science reason the account lands after the contrasts is the one from Chapter 5. Contrasting cases create "a time for telling".

How to ask for the account

If the supervisor's explanation is this valuable, don't leave it to a lucky request at the end of a week. Ask for it deliberately, with questions that make it account for the contrasts rather than summarise the activity.

The page I was handed that week is shown in full in Chapter 15, with a timeline of what the supervisor had seen by the time it wrote it. And Chapter 18 shows what happened when that page's account met the development records: they didn't entirely agree.

Which points at the limit of everything in this chapter. The supervisor explains best because it saw everything. That same fact makes it the worst possible judge of whether the finished work makes sense to someone who saw nothing.

Ask the supervisor whether an opening connects, and it will answer from inside everything it knows about what the opening was meant to do. It can't help it.

It knows the brief. It knows which moment the viewer is supposed to recognise from their own life, because it wrote that moment into the storyboard. It knows the offer, so a shot of the offer itself reads, to it, as obviously relevant. Asked "would a stranger understand this?", it simulates a stranger who, somehow, knows all of that too. Economists named this failure decades before anyone built an AI.

"Better-informed agents are unable to ignore private information even when it is in their interest to do so; more information is not always better."

— Camerer, Loewenstein and Weber, on the curse of knowledge

The study that named the curse of knowledge found that better-informed people systematically failed to predict what less-informed people would do, even when they were paid to get it right. When the informed traded in a market with real stakes, market forces "reduce the curse by approximately 50 percent but do not eliminate it."²⁴ Money couldn't buy an informed person their ignorance back. Neither can a prompt.

Two roles, two questions

"Can't you just tell the supervisor to read it fresh?"

It's the obvious shortcut, and it doesn't work. The curse of knowledge is precisely the inability to set private information aside on request. Tell an informed agent to "imagine you've never seen this before" and you get an informed agent's idea of a newcomer: someone who happens to find all the right things obvious. The simulation is contaminated by the thing it's trying to remove.

You need a different agent. Different context, different time, different job. That's not a luxury. It's the only arrangement that actually tests what you care about.

I had half of this already

In The Exemplar Ratchet I described the cold-arrival reader, who "arrives with no context on purpose, often in a customer's voice", spends about twenty seconds with the page, and answers two questions: what does this company do for you, and what would you do next? The point of it: "That is evidence about a reader, not your opinion. It depersonalises critique." In Demonstrate Over Tell I added two cautions: test "people who were there, and people who weren't", and treat a fresh AI reader as "a screen, not a verdict. A model reading a page is not a customer reading a page."

That was the cold half. What the films added was the informed half, and the rule that holds the pair together: the informed role forms, the cold role tests, and neither is allowed to do the other's job. A cold reviewer asked to explain a failure will invent a plausible reason, because it has no access to the real one. An informed supervisor asked whether the work lands will say yes, because to it, it does.

What actually makes a reviewer cold?

"Cold" turns out to be harder to achieve than it sounds. Most "cold" reviewers are lukewarm: they've been told something that warms them up without anyone noticing. Here's the checklist I now use.

The second bullet on that list isn't hypothetical. In the week this book comes from, a viewer that had been told it had already seen the ads passed an opening I'd scored six out of ten, and called its moment of recognition "specific". A viewer that had never heard of the business was harder on it, and in one run said my verdict back to me in its own words: it would never have been on the offer page; it would just be scrolling. One clause in a persona moved the harness towards my verdict. Not in every run, and Chapter 19 shows all of them, with both personas side by side.

Keep them apart in practice

• Separate context. The cold reviewer gets the artefact and a persona. Nothing from the supervisor's history leaks in, including file names and captions that describe intent.
• Separate time. The supervisor writes its account after the run. The cold reviewer tests the next artefact made under that account, not the account itself.
• Separate authority. The supervisor's explanation is tested by what the cold reviewer finds in the next piece of work, never by the supervisor's own assessment of whether its explanation was right.

Anthropic's design-harness write-up reached the same arrangement from engineering pragmatism: "Separating the agent doing the work from the agent judging it proves to be a strong lever"; "tuning a standalone evaluator to be skeptical turns out to be far more tractable than making a generator critical of its own work."⁹ The deeper reason models are poor judges of their own work, including their tendency to prefer what they produced, is covered in the sibling book; I won't repeat it.

So the artefact now has its test: a reviewer who knows nothing. But the supervisor produced something else as well, the eloquent account of what the work was for. What tests that?

A good explanation of a finished job has one characteristic flaw. It makes the previous case sound inevitable.

Every choice fits. Every rejection was obviously right. The principle was there all along, waiting to be noticed. Reading it, you feel the satisfying click of a story that accounts for everything. That feeling is what makes the explanation persuasive, and it's evidence of nothing. A story that accounts for everything that happened is easy to write after the fact. The question is whether it accounts for what hasn't happened yet.

An explanation earns its place in the skill when it helps with a new case, not merely when it makes the previous case sound inevitable.

Kahneman and Klein put the underlying problem in one line: "Subjective experience is not a reliable indicator of judgment accuracy."¹⁰ That's true of the reader's feeling that the account is right. It's equally true of the writer's.

The records kick the story

If an explanation needs something outside itself to test it, the obvious candidate is a new case, and Chapter 26 designs that test. But there's a second check that's available immediately, and it's easy to skip because it feels like bookkeeping: the records of what actually happened.

I noticed this in the conversation after the week, when my thinking partner started reading the development records alongside my account of what had happened:

It's the mechanism from Chapter 2, one level up. Attempts kick a builder's understanding of the work. Records kick an explainer's understanding of the attempts. And the reason they're useful is the reason they're uncomfortable: they can contradict the neat story. The records weren't written to support anyone's account. They're the one witness that didn't know what story was going to be told.

There are two kinds of contradiction worth looking for every time.

Both contradictions turned up in my week, and neither was flattering to the neat story. Chapter 18 shows exactly what the records said. For now, the doctrine: keep the records, and read them before you promote an explanation into the skill.

Four columns that keep an account honest

The simplest discipline I know is to make every claim in an explanation declare which kind of claim it is. My thinking partner suggested four kinds, and I've used them since.

Every claim goes in exactly one column. An idea that sits in "supplied judgment" can't also be credited as a discovery. A test that sits in "proposed" can't be quoted as evidence. And a blank "observed" column is a finding, not an embarrassment; it tells the reader exactly how far the account can be trusted.

Four agreements, one observation

There's a subtler trap. After a good week you can end up with four things that all agree with each other: the brief, the artefacts, the supervisor's page and your own discussion of it. It feels like four confirmations. It's one.

The page was written from the brief and the films. The discussion was written from the page. Of course they agree; they're one lineage. In Designing Loops, Not Prompts I gave the rule for this in a different setting: "derived output ranks below source evidence. A hard-won synthesis can be filed back, but typed as derived — cache, not evidence; subordinate to its supports; stale when they change… Skip that and the system starts citing its own echoes, confidently, and you will not notice for months." The supervisor's page is derived output. The development records are source evidence. When they disagree, the records win.

The failure has a tell, which I described in The Wiki Playbook for a knowledge graph that consolidates too eagerly: "A janitor with a vague directive merges two distinct old ideas into one false claim, silently erasing a real difference. The tell: a page that reads more confidently than any of its sources." An after-the-loop explanation that reads more confidently than any single record it came from is showing exactly that tell.

The same thing has been measured for AI judges, in a narrower form: a panel of judges fed the same inputs "never significantly beats its best member", however their verdicts are combined.²⁵ Agreement among things that share their inputs isn't independent evidence. (What that means for building judge panels is Code as Proxy's territory.)

Where the explanation goes, and how it's filed

None of this means throwing the explanation away. In Replay-Driven Design Evolution the list of what should survive a burned-down build included a folder of rejections, each kept with its reopening conditions, and the conclusion was that the durable outputs are "the harness and the history", not the first codebase. The explanation joins that history. It goes in dated, marked as derived, with its claims sorted into the four columns and its principles carrying their conditions and ceilings (Chapter 11). What it doesn't do is become the new source of truth just because it's the best-written thing in the folder.

And its ceiling is set by the rule from The Perturbation Review: "n=1 may climb, and it arrives wearing its ceiling." An explanation of one day's films is an n of one, however eloquent.

The records, my thinking partner said, "are therefore not merely memory of the work. They are material for reasoning about how the work learns." That's the reason to keep the raw sessions, not just the tidy summary of them, however beautiful the summary is. The summary tells you what the work concluded. The records let you check whether it was entitled to.

Three checks, three sources of independence

Part IV has built three checks, and it's worth seeing them together, because each one gets its independence from a different place.

That's the doctrine. The next part is the evidence it came from: one day of films, from the page that started this book to a single rejection walked through all three of its lessons, in the order the records give it, with the parts that corrected the story left in.

Parts I to IV made the argument without the evidence. This part is the evidence, in the order the development records give it, with the parts that make the setup look less clever left in. Almost all of it happened in one day.

The day began with films for businesses that exist only on my own platform.

There's a follow-up service for people home from intensive care, a farm spray-contracting service, and ten others, each with a website an agent had built and each getting a one-minute film. Partway through, I added a job that was different in kind: a short series for a small cosmetic-dental clinic. A real business with a real offer and, as I'd expected, no brief. So I supplied one myself, generated with AI beforehand from my own frameworks and what we'd already learned on the concept films.

Four films were built for the clinic, judged, and sent back for repair. While the repairs ran, I asked the agent that had supervised all of it for something else.

Chapter 1 made the claim in the abstract. This is the page.

What the page said

It opened on an idea none of us had put in those words.

"People don't arrive with dental words. They arrive with one tooth."

— The supervising agent's page

The paragraph under it did the work a brief usually fails to do. Nobody books a consultation thinking "lateral incisor proportion", it said. They've noticed something in photos for years: "one tooth a bit shorter, a chip, an edge worn flat." And then a sentence I'd have been proud to write: "What they're unsure about is whether it can change without everything changing, and whether they'll be talked into more than they wanted."

So, the page went on, the films don't claim the clinic listens. "They let the viewer watch it happen: someone points at the part they mean, says what they'd change and what they'd keep, and those words visibly carry through into the plan." Hold on to that phrase, what they'd change and what they'd keep. Chapter 18 goes looking for where it came from, and the answer isn't the one the page implies.

From there it set out the design of the films as four principles, each one about the viewer rather than the clinic.

Quoted from the page, with the clinic, its people and its offer removed.

It insisted the voices sound local, "not like an American ad". And it gave each film "one job, for one moment in someone's decision". I won't walk the films one by one here. What matters is that the page thought in decisions, not deliverables.

Read those lines together and one thing stands out. Every one is organised around the viewer's concern and the viewer's language, not the clinic's features. My brief had asked for that in principle. The page said it more sharply, and in a form a builder could act on.

What it had seen by the time it wrote

Chapter 12 argued that the supervisor explained best because, by the end, it had a pattern of decisions that needed explaining. Here's that pattern, as it stood on the day.

Set that list beside the page and you can watch the page answering it. Why did a beautiful film fail? Because craft that doesn't touch what the viewer has already noticed is a still life; hence "start in their world". Why did the tools have to work on screen? Because "a viewer takes on trust what they're told, but they believe what they watch." Why was a voice wrong? Because a speaker has to sound as if they belong where the viewer is. Generic marketing language couldn't have accounted for any of it. The page had to, so it did.

The build agent's own account, when I later asked what we'd learned, was more modest. Explaining each film under the same four headings (its job, its story, how it links to the viewer, where to use it) "was clearer than any brief we wrote." Those headings then became a note each builder writes before it builds, which the harness checks against what a cold viewer finds. That part belongs to Code as Proxy.

Then the page went back into production

A good document would have been the end of it. What makes this a loop is what happened next. Once all four films had cleared review, I asked what we'd learned from the reviews and the page, and specifically "what extra ideas need to go into the original prompting (not just for this [clinic] sub project)". Ten additions came back for the general brand-film skill. Two came straight off the page.

The first is the device three separate builders had found on their own, which only an agent that had seen all three could name as one thing. The second turns a constraint into a principle, and it's the one I'd most want a builder working for any business that can't promise outcomes to have read.

Chapter 12's line, "making improves the explanation; the explanation improves the next making", is a claim. This is the receipt. The explanation, written from the making, became two lines of the skill that makes the next film. Chapter 18 follows those two, and the other eight, into the skill, and finds that the story of where they came from is less tidy than it looks.

What the page is evidence of

One more fact, because it shows where the gate actually was. When I asked for the page, I'd told the agent to upload the films with it and add the rest as they finished. Then I stopped the upload.

So the page went live with no films on it, each marked "Coming soon". A film was added only after its repair round and a fresh round of judging, against a bar of 7 or more from every judge and no honesty gap: nothing a viewer came away believing that the film had no right to promise. The page explained the work. It didn't approve it. The review loop did that.

The page's sharpest idea, that craft has to connect to something the viewer has already noticed, didn't arrive as an idea. It arrived as a film that one judge loved and another turned down.

The film's job was to stop someone scrolling. About twenty seconds, tall, built for phones, and in the page's words, "beauty first, so it works with the sound off."

The first cut delivered exactly that. Macro stills of the material being worked (resin, shape, contour, shade tabs, polish), one word on screen per cut. Then a teaching model with proportion lines, a single spoken line, "Shaped around what you'd like to keep", and an end card with the offer. It was the first of the four clinic films to finish building, and the judging ran straight away.

Three verdicts on one cut

"The craft shots never connect to what a viewer notices, such as a chipped or short edge."

— The harsher judge, on the first cut

It had other notes: the music sat almost as loud as the voice, two dissolves left ghosted text, and the end card needed a web address a phone could read. The build agent's verdict on the verdicts took one line: "Opus's criticisms are fair, so they're the fix round."

Neither judge was wrong

Chapter 4 put the two meanings of "on brand" into two sentences. Here they are with a film under each.

The enthusiastic judge wasn't malfunctioning. It was answering the first question: is this beautiful, premium, polished? Yes. The harsher judge was answering the second: does this connect to the person watching? No. Both were real readings. Only one was the question the film existed to answer.

In the terms from Chapter 4, the craft family won and the doctrine family lost. The film was superbly made and didn't carry what this business actually offers, which is to start from what you've noticed. Chapter 4 gave the rule from The Exemplar Ratchet that decides it: a page has to win across the families, not average them. Average 8.8, 8 and 6 and you get a comfortable pass. The bar the films had to clear was 7 or more from every judge, and that one word, "every", is what stopped a beautiful film that didn't fit from shipping.

One inference of my own, labelled as that. The first cut used the clinic's own colours and type, so a brand check built on palette and typography would almost certainly have passed it. It was on brand in the shallow sense by construction. The fit question is what failed it.

Three rounds, three lessons

The repair took three rounds, not one, and the interesting thing is that each round's rejection was about something different.

Reasons are the judge's own words or the revision instructions that quoted them. Scores are given only where the story turns on them.

Round two is the one to dwell on. The first repair had done what it was asked: each beat now named something a viewer might notice. But the tools were, as the supervising agent put it in the next round's instructions, "posed, not working." A tool held beside a finished surface tells you what the service does. It doesn't show you. The instruction for round two ended with the sentence I'd put on the wall of any studio that makes things with AI:

"This round is about making the change happen on screen. That's what makes it amazing rather than handsome."

— The supervising agent, instructing the second repair

That second rejection is where the definition of "demonstrate" in Chapter 6's table came from: a change the viewer watches inside the shot, before, the service acts, after. The first rejection had said connect. The second said connect by showing a change. The distinction got sharper because the first repair satisfied the words of the first rejection without satisfying its point.

Round three looks like housekeeping, and it's easy to skip. But the film was running well past its intended length because a static end card had been allowed to sit, and that turned into a rule for every future film: a length is a budget, not a suggestion.

The card, filled in

What happened to the judge that loved it

By the end of the reviews, the judge that called the first cut a masterclass had scored every film between 8.5 and 9.7. The build agent's summary was fair to it: it "confirms the sound but can't separate a weak film from a strong one", and the recommendation was to use it only for the listening report. Why it behaved that way belongs to the sibling book.

The taste lesson isn't about that judge, though. What made the 6 valuable wasn't that it was low. It was that it came with a reason specific enough to act on, three times running, and each reason was about fit rather than finish.

The craft film's rejection was clean: one question, three rounds, a named reason each time, each ending up in a carrier. The voice rejection earlier that day was nothing like as clean, and it's the better lesson for it.

Here's the whole of the rejection, as I typed it.

Paddock Relay is a farm spray-contracting concept, and its film opened on a farm manager's voicemail. That's taste at its rawest: no principle, no vocabulary. "Crappy." "Older tech." And the most honest part, "or something", which admits I couldn't say what was wrong, only that something was. Everything in this chapter is where those words went.

Leg one: the voice itself

The cause took minutes to find. The voicemail used a local text-to-speech voice that the engine's own authors grade D+; the narrator's voice was graded A. The builder hadn't processed it any differently. The voice itself was weak. So only voices graded B- or better were allowed from then on, and anything else was refused with the reason and the alternatives. That exposed a catch: the engine had no male voice above C+.

So the farmer was recast with a different engine, given a spoken description of who he is, where he is, what he's doing and how he talks, and the best of several takes was kept. I chose that voice myself, by ear. The skill still carries my verdict on it: "perfection", "amazing".

That closed the artefact leg, and it would have been easy to stop there. A weak thing was found and replaced. Most feedback loops end at exactly this point.

Leg two: the checks that couldn't hear it

Why had the builder picked a D+ voice in the first place? Not carelessness. The skill had told builders to audition voices by comparing their pitch and checking that the transcript came out clean.

"Neither measure catches quality, so the builder picked a bad voice in good faith."

— The build agent, diagnosing the voice

Later that day the judges were calibrated against my own verdicts, and this voice was one of the test pairs: the rejected voice against the one I'd picked. The judge that read a transcript preferred the rejected voice in four runs out of four. The transcriber had misheard the gruff replacement, and the judge marked the mishearing down as bad diction. The judge that could listen agreed with me in six runs out of six, and named the thing I hadn't been able to: "an incongruous British accent".

Why judges miss what they miss is Code as Proxy's subject, and it's told there properly. For taste, the instrument leg taught something narrower, and the skill now says it twice. Once as a warning: "Pitch and a clean transcript don't measure quality, so don't audition your way around the catalogue." And once, more bluntly, as a limit: "No model can judge taste in voice or music yet. Say what you chose and why; a human listens."

Leg three: from quality to place

Look again at what the listening judge named. Not a low grade. Not "older tech". An accent, and an accent that didn't belong. Without anyone deciding it, the diagnosis had moved from how good the voice was to where it was from, and who it was talking to.

The move finished a few hours later, on a different job. When the clinic series started, the supervising agent didn't reach for the platform's default narrator. Unprompted, it cast locally, with a narrator whose accent matched the clinic's area "rather than our American house voice". It tried twelve regional voices and let a listening judge pick the narrator and the patient's voice. That's the moment I describe at the start of Chapter 10, when I found my judgment turning up in the testing.

Then, when the day's lessons were harvested, the principle got its name. It was the last of the ten additions to the general skill: "Cast the voice from the audience's place." It went into the skill with an example, and the example is the interesting part.

The fault I'd heard was a British-accented farmer. The example the skill carries is an American narrator. The principle had outgrown its origin. It isn't "no British voices"; it's "nothing that sounds imported to the people listening". That's what a rejection looks like once it has travelled all three legs: the thing it protects has a name, and the name reaches cases the original rejection never touched. Chapter 10 unpacked what the principle does and doesn't say, and Chapter 11 filled in its conditions.

One guard held, and it's worth pointing at because it's exactly where superstition would have crept in. The ban was scoped to the engine's own quality grades, not to an accent. The catalogue still includes a British-accented voice that passed on quality. The rejection was about a weak voice and then about belonging. It was never about Britishness.

The neat story and the record

Here's how I first described the whole thing, a little after the fact:

That's not wrong. But when the development records were read against it, the story came out in more pieces.

Four events in one day, three kinds of agent and me. Not one remark, generalised.

Why tell it this way, when the neat version flatters me more? Because the messy version is the one the method predicts. Chapter 2 argued that the distinctions taste depends on are revealed by consequential attempts, plural, not extracted from one remark. And the neat version would teach the wrong lesson: that a single well-put "I don't like that" is enough to form a principle. It wasn't. It started one.

The part I haven't done

There's one more fact, and it belongs in any honest version. The clinic's voices were chosen by a listening judge. When the films went live, the build record listed what was still for me to check, and the first item was the voices: "A … listening judge chose them; you haven't heard them yet."

That's where the record leaves it. So the principle stands, and its origin is messier than the tidy version. Its application in that series hasn't been approved by the person whose ear started it. The skill's own rule says a human listens. That human is me.

By the end of the day, lessons like this were scattered across a chat, several rounds of repairs, a calibration report and a page. I asked for them to be gathered into one place. The next chapter is what came back, and what the records said when we checked it against them.

Chapter 15 quoted the question I asked once the films had cleared review. The answer opened with a sentence I'd like every retrospective to start with.

"Your review loop worked: every film improved each round, and the notes that drove the improvement point to a handful of general fixes. Nothing below is applied yet."

— The build agent, answering the harvest request

Nothing below is applied yet. A harvest is a proposal. It becomes a change only when someone accepts it, and this one knew that about itself.

What the reviews showed

It started from the evidence. The harsher judge had given seventeen verdicts across all the rounds of all four films. The agent counted which criteria it had scored under 7, and went back to what had actually been wrong each time.

As recorded in the harvest, lightly generalised.

Read that table by kind, not by count, and it sorts itself. Pacing and sound are craft, and code would later measure both. Distinctiveness, demonstrate and the hook are failures of fit that no palette check could see: the films looked right and still didn't say anything only this business could say, or show anything happening. And connection isn't a failure of the films at all. It's a defect in the instrument's vocabulary, one word doing the work of two, which is how Chapter 10's split into recognition and the aha came about.

Two notes about the instruments, one line each. The cold viewer caught a commercial gap the judges hadn't: two films left viewers not knowing the price ("I'd ask about pricing on the call"). And the listening judge, as Chapter 16 showed, couldn't gate.

Ten additions, and where each came from

Then the proposals for the general skill, the one every future film on the platform is built from, written, as Chapter 9 described, as a principle, a why and a test rather than a rule.

1. Show the change, don't pose it. From the craft film's second round (Chapter 16).
2. Their words travel. "The strongest connection device we have." From the page, and from three builders who found it independently (Chapter 15).
3. Say and show the offer. The price and the next step on screen for muted viewers, and said aloud for everyone else. From the cold viewer's "I'd ask about pricing on the call".
4. Length is a budget. From every builder running long, and the craft film's third round.
5. Size the example to the offer. "If the example is smaller than what's sold… bridge the gap honestly on screen." See the records, below.
6. When you can't show a result, show being understood. From the page (Chapter 15); the constraint behind it is Chapter 23's.
7. One thing only this company has. "Generated images carry mood; distinctiveness comes from the company itself." From the distinctiveness row.
8. In a series, give each film one job in the buyer's decision, with the same next step at the end of all of them. From the page.
9. Write the marketing note first. "It sharpens the build, and the test harness can check the note against the cold viewer's answers." From the page's four headings (Chapter 15).
10. Cast the voice from the audience's place. From the voice (Chapter 17).

Every one of them has a traceable origin in a rejection, a comparison or the page. None of them is a rule about a clinic. Each is a statement about fit that should hold for the next business, and each carries its own reason, so a builder can tell when it doesn't apply.

And for the harness

The same answer sorted the harness changes by what they'd change. Most of them are told properly elsewhere, so here they are as a map.

A harvest with a kill switch

I said yes. Not unconditionally.

Two things in that are doing more work than they look. The first is the length lesson, which is the craft film's three rounds turned into an instruction: the skill now says 30 to 60 seconds is the range that works, and that under about 25 is hard to land. The second is the conditions. Chapter 14 said an eloquent account is a candidate until it helps with a new case. I didn't have that sentence then, but the authorisation does the same job: it names the new case (a fifth film), says what "helping" means for each carrier, and says what happens if it doesn't.

Then the records contradicted the neat story

Afterwards, my thinking partner read the development records alongside my account of the day, and the account changed. Chapter 14 named the two contradictions to look for, approval and provenance. Both turned up.

Approval was the voices, and Chapter 17 has told it: the clinic's local voices were chosen by a listening judge, and the record says I hadn't heard them. A high score from a listening judge isn't my yes, however well it hears.

Provenance is the bigger one. Chapter 15 asked you to hold on to a phrase from the page: what they'd change and what they'd keep. The neat story, and the one the page implies, is that the agents found it. The record of my original request says otherwise. I'd supplied a research brief with that request, and I said so plainly when the question came up:

And that brief, read now, already held most of what I'd been admiring in the page.

Look at the second item on the left. The page's founding line, "People don't arrive with dental words", was in the brief as an audience hypothesis, in the imagined voice of a patient. The attempts didn't originate the central ideas. They sharpened them, bounded them and made them testable.

The provenance makes the achievement clearer, not smaller.

— My thinking partner, after reading the records

I agree with that. The neat story, where the AI found the strategy in a clinic's website, is more impressive and less useful. What actually happened is a capability you can plan around: a professional's framework-informed brief carried into a new client's world, and made sharper and more concrete by production than it was when it went in. The three sources that made the films, and why it pays to keep them apart, are Chapter 20's subject.

And here is Chapter 14's point about connected derivatives, in practice. The brief, the films, the page and our discussion all agreed with each other, because each was made from the one before. The page wasn't wrong. It just didn't say where its best idea came from, and nothing in that lineage could have told us. Only the records could.

So the harvest was written, authorised with a kill switch, and corrected by its own records. Then the fifth film ran, built with the new prompting and judged with the new harness. It cleared the judges' bar on its first build. I thought it was fantastic. And I gave its opening six out of ten.

Both of those things were true at once. The film cleared every judge. I thought it was fantastic. And two parts of it didn't work for me. The distance between those statements is this chapter.

The fifth film in the clinic series explained why a treatment plan looks at more than the one tooth you've noticed. It was the first film built with the ten additions from Chapter 18 and judged with the new harness, under the kill switch I'd written into the authorisation. The first data point looked good. The first four films had each needed at least one repair round to clear the bar. The fifth cleared it on its first build: every judge 7 or more, the cold viewer 9 out of 10, no honesty gap. That's one film, and its ceiling is "a hopeful sign that the new prompting helped early", not a finding.

The verdict

I'd already called the film fantastic and asked for it to be published as it was, with one note about the end card: "it could do with more composing IMHO — but isn't terrible". Later the same evening I went back to the opening.

The build agent's response was the first move of the method, and it's easy to miss because it looks like note-taking: "Recorded. Your full verdict on E gives us a clean human reference to test the judges against." A table followed: opening about 6, middle awesome, end card needs composing. Then the standard any judge would now be held to: "A judge worth keeping should find that same shape: praise the middle, flag both ends."

Then I narrowed it. "The image connects, just not the iPhone/web page. The inset image." The opening was a man on his back steps at dusk, with one line about only ever noticing one tooth. Over its first five and a half seconds, an inset phone showed the clinic's offer page. The scene worked. The inset didn't. So the label became precise: "A good judge should flag the inset, not the scene or the words."

That's the first field of Chapter 8's triage card, the rejection in the rejector's own words, filled in precisely enough to test against. Everything after this depends on that precision. "The opening's a bit off" couldn't have told a judge that got it right from one that got lucky.

What the judges said

The judges agreed with me where the film was strong and missed both of its weak ends. A harness that only praises correctly is half an instrument.

Notice the order, because it's the doctrine doing its job. The verdict was recorded and the judges were compared before anything was repaired. Fix the opening first and the evidence of why the harness had forgiven it would have vanished with it.

Leg one: the film

The revision brief kept almost everything. The man, the steps and his lines stayed. The middle stayed "exactly as it is". Only the inset was to go, replaced "with something from his world, not the clinic's: something a stranger scrolling Instagram, who has never seen [the clinic], recognises as their own." The end card, whose smallest text code had measured below the legibility floor for a phone, was to be composed properly. And then three words that matter more than they look: "Your call how."

The builder's call was a photo message from the man's sister: "Look at you two laughing!" It's him, laughing, with his daughter on his shoulders. The picture closes in on his laugh, and a ring settles on the one short tooth. Everyone else sees the laugh. He sees the tooth.

The tall cut was another matter. It had put the phone, photo and all, over the man's head.

So the artefact leg had found a second defect, and this one didn't stop at the film. (The shape of that revision brief, keeping everything except one element and composing the end card from its measurements, is Code as Proxy's example of a well-made descendant brief.)

Leg two: how we'd been telling

Three things in the harness had forgiven the opening, and a fourth missed the covered head. All four came out of that evening's records.

The judge brief. The build agent traced part of the miss to itself, and said so without being asked.

"Part of the miss is my fault: the judge brief I wrote for E says 'Where it's seen: retargeting', so the judges forgave an opening that assumes the viewer has seen the offer. Your point is that any ad gets met cold on Instagram. So the next brief change is that every film must work for someone who has never seen the offer, whatever its intended placement."

— The build agent

I'd hold that paragraph up as the instrument leg done honestly. It names its own contribution to the miss, says what my rejection actually meant, and turns it into a rule about every film rather than a patch on this one.

The persona. The cold viewer had been given a persona, and it wasn't cold. The fix changed one clause.

Here's what the two personas said about the two openings, every run, in the order they ran. I'm showing all seven because the tidy version is in the build agent's summary, and the runs are not tidy.

From the judging records; place names and treatment names removed. The question and its grading were being reworked while these ran, so read the words, not only the grades.

Read the answers rather than the grades and the second defect in the harness is obvious. Five of the seven are about whether the person on screen looked like the viewer. That's the question's fault. It asked whether there was a moment "you thought 'that's me' or 'someone like me'", and a woman in her forties, shown a man, answered the question she was asked.

The question. So the question was rewritten to ask about the viewer's own life, "even if the person on screen isn't like you", and the grading was rewritten with it. The new grading rule is worth quoting whole, because of where its example came from:

That last example reads like run six's sentence, lightly generalised. As far as I can tell from the records, a viewer's words became the grading rule. Now read the table again under it. Runs two to five are look-alike complaints, not misses. Run six is the only answer about the situation, and it's my verdict, in a stranger's voice. Run seven recognises the new opening as a situation that was hers. And run one, the primed viewer, had no reason to find the offer page foreign, because she'd been told she'd seen it.

"Instead of enjoying it you zoom in on your own mouth. I have absolutely done that."

— A truly cold viewer, on the new opening

One honest note about the record itself. The build agent's summary, and the skill's own account of the case, tell this more cleanly than the runs do: one truly cold viewer, one verdict per opening, the primed viewer passing both. The skill even attaches "I was relating to the line, not the person" to the old opening; in the runs, it was said about the new one. Chapter 14's rule applies to the harness's own history as much as to anyone's story. The records win. The summary's direction holds; its tidiness doesn't.

The rubric. The fourth was the covered head. Neither main judge had named the phone over the man's head in the tall cut. After the rubric was tightened to cover "a person's face or head, even seen from behind", the judge found a real flaw on the revision's tall cut, a label running across the top of his head at 0:39, which the builder's tall-only fix had already moved. On the opening, it now said the cut "layers what a wide frame would… then a card over the photo", which is my "needs to attempt less" in its words. It still never said in so many words that the phone covered his head. The build agent wrote that down as "the one miss left". Meanwhile the main judges' recognition scores for the opening rose from 7 to between 8 and 9 on the new cut: the instrument and the artefact moving together, and measured.

Leg three: the skill had caused it

When the build agent traced the covered head, the trail didn't end at the builder. It ended at the skill. The skill's advice for tall cuts had said "the card sits over the quiet part of the scene", "which is how E's phone landed on his head." The principle was wrong, not just the film.

The same was true, more quietly, of the other two instruments. The question's "someone like me" had a principle built into it, that recognition means resemblance, and the brief's placement note had one too, that a viewer who's been retargeted can be assumed to know things. Both were wrong in a way that let a disconnected opening through.

The answer was Chapter 9's three carriers, described from practice rather than theory; I quoted it there. Here is what it actually changed.

Quoted from the brand-film skill and the judge as committed that evening; clinic details removed.

Look at the conditions in the new text, because that's where Chapter 11's guard shows up. Tall cuts attempt less than the wide one, not less in general. The company's own material isn't banned from the opening; it comes "second, and light". The lesson carries its scope with it.

Where it went

By the end of the evening, everything was committed and pushed, with nothing forced, and the brand-film skill was on all twelve company boxes. The film agents inside them pick it up on their next film task, with no restart. The lesson now sits where a builder working for a different company, in a different world, will read it before it builds. Whether it helps there is the successor test, and that's Chapter 26.

The case, read against the doctrine

Why this is the evidence the book rests on

One human verdict, a few sentences about an opening, became three things. A better film, which I approved. A harness that now asks the right question of a truly cold viewer, reads its answers by situation rather than resemblance, and has a written note of the one thing it still misses. And principles stated precisely enough, with their conditions, to change the next film for a different company. All of it is dated, all of it is in the record, and the skill text is quoted before and after. If this book's claim is true anywhere, it's true here, and you can check it.

That's the method working for a business whose owner never had to explain their brand. They couldn't have, and nobody asked them to. The next part takes that case head on: what to do when the client can't brief you at all.

Looking back on the clinic films, one thing struck me that had nothing to do with film.

That isn't a criticism of anyone. It's the normal case. The people running a small practice are good at the practice. They've never had to write down who their customer is, what they promise them, or how they sound, because nobody has ever asked, and the business works anyway. Chapter 4 quoted a large vendor's admission that even brand guidelines running to hundreds of pages "never go far enough". The businesses that most need help don't have the hundreds of pages. They don't have the history of approvals and rejections those systems learn from, either.

What the usual workflow assumes

The conventional creative workflow starts from a brief the client supplies: the audience, the brand promise, the tone of voice, what makes them different, what the campaign is for. My thinking partner put the problem with that plainly when we talked it through: "Asking the owner to complete a sophisticated brand questionnaire may amount to asking them to perform the professional work they need help with."

I'd made the same point from the other side in The Manila Folder, about helping people discover what they might want to do: "We don't need to understand someone completely before showing them something — showing them something can help us understand them." And its warning about the alternative: "Questionnaires produce false clarity: neat fields, wrong fit."

Requirements engineering has had a name for this for a quarter of a century. "I'll Know It When I See It (IKIWISI). Requirements often emerge only after users have had an opportunity to view and provide feedback on models or prototypes."³ And practitioners building AI evaluations say the same about domain experts, whose expectations are ones "they might not be able to fully articulate upfront" until they're judging real outputs.¹⁹ If people with a profession can't state their standard in advance, it's unreasonable to expect it of a business owner who has never been asked.

Instead of requiring the client to supply a finished brand strategy, the system uses informed judgement to develop a concrete candidate they can recognise, correct and approve.

That's the relationship change, and every chapter in this part hangs off it. The client stops being the source of the brief and becomes the judge of a candidate. That's a job they can actually do.

Three sources, kept apart

If the client isn't supplying the brief, something else is. In the clinic's case there were three sources, and the records only made sense once we separated them.

The middle column is my thinking partner's wording, after reading the development records.

Keeping them apart isn't bookkeeping for its own sake. It "prevents a misleading account in which the AI somehow discovered the whole strategy in the website". That's exactly the account I'd have told if the records hadn't been read against it: the central idea looked like a discovery of the third source, and it came from the second.

What transferred

Here's the part I believe most strongly, and it's the reason this book is about taste rather than templates.

"What transferred was your way of thinking about communication — not just a successful video template."

— My thinking partner, on what carried into the new client

The follow-on sentence is the distinction: "A template carries a previous answer. Your contribution helps the system decide which questions a different client needs answered." A template from the concept films would have given the clinic a film shaped like the concept films. What actually carried was a set of questions: what does this viewer already handle, what would they recognise, what can be shown rather than claimed, what can't be promised. Those questions produced different answers for a different business, and that's the point of them.

Two passes, and the one that can't be loaded

In Worldview Recursive Compression I described two stages in making anything with AI that sounds like you: "There are TWO compilation passes, not one. Skip the first, and you're stuck nursing outputs." The first pass compiles the professional into the builder: your frameworks, constraints, voice and standards. The second applies that builder to a client's context to produce the work.

The clinic case adds something to the second pass. Its input, "client/project context", was written as though it were a document you load. When the client can't brief you, it isn't. As my thinking partner put it, "the client-specific context is not fully available at the beginning. Some of it has to be developed through the work." The second pass has an input that production has to write. That's the next chapter.

"They should just hire a brand strategist"

Some should. Many won't, and some can't. And even a good strategist does better work with something to react to than with a blank sheet. In Demonstrate Over Tell I argued that an open "why?" is a recall task and a set of offered options turns it into recognition, which people can do in seconds. The line I'd give any strategist: "A system that can't take a 'no' cheaply can't learn what 'yes' means." A concrete candidate is the cheapest way to get a client's "no", and their "no" is where their brand starts to become visible.

Tag every claim with its source

The operating habit that makes all this safe is small. Every claim in the working brief gets a source and a status. It's the told-versus-revealed record from Chapter 3, with one more column, for the client.

Shape Four rows from the clinic case, generalised. Date every row when you use it. "Not yet" is the most important entry in the table: it's the list of things to show the client.

Two things fall out of keeping this table. The first is honesty about provenance, the thing Chapter 18 needed the records for. The second is a to-do list. Every "not yet" is a claim about the client's business that nobody from the client has seen. Chapter 23 turns that list into the thing you show them.

But "production can write it" is only true if production is looking. If some of the client's context has to be developed through the work, the work needs a way of developing it on purpose rather than by accident.

"We provide personalised care." Every service business says it. It can sit on a website for ten years without anyone asking what it means.

Try to film it and the sentence stops working. A camera can't point at "personalised". It has to point at something, and so the production question arrives, whether anyone planned to ask it or not.

What would the viewer actually see happening that makes this care personal?

Nobody can answer that with a synonym. The answer is a scene: a person says what they'd like to keep, and it turns up later in the plan. Or the answer is a silence, and the silence is information. My thinking partner listed where it points: the question "can expose a hole in the script, a hole in the brief — or a hole in the service's own account of what it does." Three holes, three owners. The builder owns the script, the professional owns the brief, and only the client can say what the service actually does.

"Filling gaps" is three different jobs

Once production starts exposing holes, it's tempting to fill them all the same way: write something plausible and move on. That's where most of the damage in this kind of work gets done, because the holes aren't the same kind of thing.

The three kinds and their examples are my thinking partner's, from our discussion of the clinic records.

The rule that matters most is in the last row. Hypotheses about the audience "need testing rather than being promoted into facts because the story sounds convincing." That's the characteristic mistake of a persuasive account, and a supervising agent that explains beautifully (Chapter 12) is exactly the kind of writer that makes it. "A good system can work with all three. It should not treat them as interchangeable."

That's why I'd call this brand development through production, not brand extraction. Extraction assumes the brand is in there somewhere, waiting to be found in the website and the owner's head. Development admits that some of it doesn't exist yet: choices nobody has made, and guesses about the audience nobody has tested. The verb matters, because "extract" licenses you to present what you wrote as what you found.

Why the owner can't just tell you

The first kind of gap is the one that looks like it should be easy. It's already true. Just ask. The sociologist Harry Collins explains why asking doesn't work, in a taxonomy of the ways tacit knowledge stays tacit. A review of his work summarises the relevant kind as knowledge that "could be made explicit… but is not made explicit", and names its sub-forms.²⁶

And the second kind of gap, the choice not yet made, is Schön's point from Chapter 2 applied to the client: "An evolving intention is one of the outputs" of designing.² A client often doesn't know which part of their service to lead with until they see it led with.

Filling is fine. Knowing what you're filling is the job.

I'd noticed the system doing gap work before I had any of this vocabulary.

Both of those are true, and the second is still how I feel about it. But the discipline is in the word "filling". Filling a hole in the script is the builder's job and nobody minds. Filling a hole in the service's own account, and presenting the filling as fact, is how a film ends up promising something the business doesn't do. The records from that day have a specimen of each kind handled well.

• A hypothesis, labelled as one from the start. The brief I supplied called its audience ideas "creative hypotheses to test, not findings from patient research" (Chapter 18). That label is the third row of the table, correctly filled in before a single film existed.
• A choice, designed through production. The same brief warned against jumping from one small concern to a much larger package. After review, that became a principle for the skill, "size the example to the offer", with the instruction to "bridge the gap honestly on screen". The choice of how to bridge it was made in the films, where it can be seen and judged rather than argued about.
• Something already true, made visible. The fifth film's middle, the part I called "awesome", explains why you can't treat just one tooth. That reason was already true clinically. The film's job was to make it visible to someone who'd never been told (Chapter 19).

The method: tag every claim before anyone shows it

The tag also tells you which of Chapter 20's three sources a gap needs. Already-true gaps need the client's knowledge. Choices need the professional's design and the client's decision. Hypotheses need the audience, which is the one source no amount of production can stand in for.

All of this assumes you can see the gaps in the first place. Most of the real work is noticing them, and what gets noticed as missing depends on whose judgment is doing the looking.

Here's the sentence from our discussion of the records that I'd put at the centre of this part of the book.

Your expertise is influencing more than the answers. It is influencing what counts as an unanswered question.

— My thinking partner, after reading the development records

Picture two readers going through the same client's material: the website, the offer page, the enquiry form, the practitioner's photo. The first has no particular framework for communication. They "see enough to begin: offer, price, logo, colours, contact form." Everything a production needs is there. The second reader carries a way of thinking about how a film connects to a person, and sees something else entirely: a set of questions the material doesn't answer.

None of the second reader's questions can be answered by finding a document. That's the point, and it's worth stating exactly.

That reframes what the autonomy I was so struck by actually was. The system wasn't only "completing assigned production steps". It was "identifying prerequisites for doing the work well and proposing ways to satisfy them." Where the prerequisites came from is the question, and the honest answer is: from the judgment it was carrying. "The AI filled in the gaps" is the flattering description. "The professional's judgment decided which gaps existed" is the accurate one.

The absence nobody had named

The clearest example from the day is one Chapter 18's table already showed. In ten of the harsher judge's seventeen verdicts, distinctiveness scored under 7. The generated imagery was, in the judge's phrase, "the stock vocabulary of premium cosmetic dentistry". The distinctive parts were the clinic's own things: a real portrait, the real enquiry form, the clinic's own lines.

That's an absence nobody had written down: nothing in these films is only theirs. It didn't show up as a missing field. The films had a logo, a palette, an offer. It showed up only when a judge was asked a question about distinctiveness, repeatedly, and kept failing the same thing. What it teaches is simple once it's said. Generated material supplies the mood of a category. Distinctiveness has to be found in the company itself.

My thinking partner was careful to call that "a recorded interpretation of the review rounds" that "became a proposed general principle", and that's the right status. It's a candidate, supported by one series. And the further claim, that viewers would actually understand or respond better to the distinctive version, "still requires audience evidence." The absence was noticed. Whether filling it helps is a separate question, with a separate kind of answer.

Remove the contribution, and see what disappears

There's a test I wrote for a different purpose that turns out to be exactly the right test here. In Made From You, about companies built around one person's expertise, the claim was that "your understanding affects what the company notices, which questions it asks, how it interprets a request and when it refuses an easy answer." And the test for that claim: "Changing the colours, biography and tone of a standard offer does not establish that claim… The useful question is what would change in an actual piece of work if your contribution were removed."

Run it in both directions. For the client: hide the logo, and what remains that's only theirs? For the professional: remove my brief, and what would the system no longer have noticed? From the records, the honest answer is the central question (what would you change, and what would you keep?) and the demand that a person's preferences survive into the planning. Those were mine, and the films were organised around them (Chapter 18). Without them, my guess is that the builders would still have made competent films, and more of them would have looked like Chapter 16's first cut: craft displayed rather than connected. As I put it in the same book, "your practical distinctions shape its decisions." Here the distinctions shaped what counted as missing.

This is also where Chapter 21's mismatched saliences get resolved. An owner is unlikely to see their enquiry form as distinctive; to them it's a form. The professional can. Noticing absence is resolving a mismatched salience in the client's favour: seeing that the ordinary thing they do is the thing only they have.

Absence prompts

If the professional's contribution is a way of noticing what's missing, it can be written as questions any system can ask of any client's material. These are mine, from the films. Yours will differ, and should.

Run two of them against the clinic's own material and you can see how they work. Prompt four first: what here could any competitor show? A macro shot of resin, shade tabs, a gloved hand holding an instrument, a polished surface. Strike all of that and very little of the generated imagery survives, which is exactly what the harsher judge kept finding. Now prompt five: what's the one thing only this company has? The clinic's real enquiry form was one answer. Its question, "What would you most like to discuss?", is the business speaking in its own voice, at the exact moment a hesitant person decides whether to ask. A film that opens on that form starts in the viewer's world and in the clinic's at once. Nobody briefed it as a brand asset, and an owner is unlikely to think of their enquiry form as one. The prompt is what turns it from a field into a relationship.

One rule goes with them. When a prompt finds nothing, record that as a finding, not as a gap to paper over. "We found no moment where the customer's words change anything the service does" is a fact about the material, and a useful one. It goes to the client as a question: is this something you do and haven't said, or something you don't do? It doesn't go onto the screen as an invention.

That last sentence is harder to keep than it sounds, because noticing an absence creates a temptation. A system that can see a consultation that listens is missing can also draw one, beautifully, whether or not the clinic's consultations work that way.

The sentence that draws the line is short, and I'd like every agency using AI to have it pinned somewhere.

The same goes for every service moment a film can draw. A call-back sequence, a planning note, a clear explanation of fees: each "may be an excellent communication design. Whether it faithfully describes the clinic's actual process remains a separate question." And the line gets more important, not less, as the tools improve: "The more capable the system becomes at filling gaps, the more important this distinction becomes." Inventing the business isn't the failure of a weak system. It's the characteristic failure of a good one.

A demonstration explains

The rule I use for this is from Demonstrate Over Tell, and it was already in the brief I supplied for the clinic, before any film existed: "A demonstration explains. Evidence substantiates." The book spells out what that means: "A demonstrated moment shows what something is like. It is not proof that it happens reliably, or at all." And the version I'd put in front of anyone tempted to film a service moment the client hasn't confirmed: "A demonstration of a feature that doesn't exist is a misrepresentation with good typography."

Applied to generated media, the consequence is blunt: "A generated demonstration is not evidence of reliable service delivery." A beautifully listening consultation, drawn by a model, is evidence that the model can draw listening. It says nothing about the clinic.

Not a reason to retreat

The obvious overreaction is to stop showing service moments at all and go back to the generic ad: logo, smiling person, offer, phone number. That throws away the most useful thing the process makes. The proposed film has a second use, and it's the one this part of the book has been building towards.

Show the clinic the proposed experience and ask which parts accurately represent its practice, which need correction, and which it would like to adopt.

"Now the client does not have to manufacture a brand guide from a blank page. It can exercise judgement over a concrete account of its service." That's the questionnaire problem from Chapter 20, solved from the other end. The owner who couldn't tell you their brand promise can certainly tell you, watching a scene, "we don't do that", or "yes, that's exactly what we do", or "we don't, but we should".

In Specimen, Not Prescription I drew the line between a demo that closes listening and one that opens it. The specimen line says: "Here is a working specimen of the consulting model I'm describing. It proves the method exists. Discovery determines whether, where and how it applies to you." Its value is the resistance it invites: "Adversarial contact with a concrete object is higher-resolution than polite agreement with a vague sentence." The proposed film is a specimen of the business, shown to the business. Every "we'd never say it like that" is information about the service itself, the kind a questionnaire never gets.

The client review sheet

How to run it matters as much as the sheet.

• Walk the moments in order, with the draft playing. The owner is reacting to a scene, not recalling a policy.
• Ask for a mark per moment, not an overall opinion. "Looks great" is the answer a sheet exists to prevent.
• Record corrections verbatim. They're the most valuable output of the session. They are the brand, arriving.
• Don't let "adopt" become "publish". A commitment the business likes and hasn't implemented is a plan, not a fact.

There's a family resemblance here to a method I described in Experience Is the Unit, where you write the review you'd be proud to earn and work backwards from it. The warning there applies exactly: the target is not a script. "Their actual feedback must be free to contradict the design target." And "the most useful review is not necessarily the most enthusiastic." A review sheet that comes back with every moment marked "represents our practice" and no corrections hasn't been read. It's been signed.

Both directions

My thinking partner's summary of what that means: "That is brand work becoming connected to service design." The valuable outcome "is therefore not simply a better marketing surface. It is a clearer agreement about what the business intends customers to experience." I think that's the most underrated thing in this book. A small business that can't write down its brand promise can end up, via a film it never asked to be the source of, with its brand promise written down, corrected by the owner and true.

When you can't show a result, show being understood

The clinic case came with a hard version of the honesty boundary, and it's worth telling because the constraint made the films better.

In Australia, advertising for higher-risk cosmetic procedures is tightly regulated. The national guidelines say such advertising "must not use testimonials due to their potential to create unrealistic expectations of beneficial treatment." Before-and-after images "may create unrealistic expectations", and where they're used at all they must be genuine images of actual patients treated by that practitioner, unedited, and carry a prominent warning.²⁷ A generated film can't show a result at all. There's no real patient in it to show.

The brief I supplied had already flagged that direction, and the films took it seriously. Chapter 15 quoted the page's answer: no results, no reveal, because such shots "also make weaker films, because they promise an outcome nobody can promise." The films showed the care and the process, and ended "with the person just as they began." At harvest the constraint became a principle for every future film, whatever the industry: "When you can't show a result, show being understood."

One boundary on my own claim. I'm not saying any of these films was reviewed for regulatory compliance; that's a professional judgment I'm not qualified to make, and the review page itself carried a checklist to work through before any ad ran. The point is narrower. An honesty constraint pushed the work away from outcomes and towards the process and the person's agency, and that turned out to be exactly where the brand's real consistency lived.

Look at what survived every honesty check in the series. Not the look, and not a result. What survived was the person's own words carrying through the service, which is exactly the thing the next part examines: what being heard actually requires.

Here's the one film detail this chapter needs. It's from the page, and it's the whole idea in two lines.

The page's name for the device was "Their words travel", and Chapter 15 showed it running through the whole series: what the person says at the start turns up later, in the notes, on the call-back card, in the plan. Three builders found it on their own. It was the strongest thing in the films.

But "travel" undersells it, and my thinking partner put the sharpening in one line.

Their words do not merely reappear. They change what happens.

"That is the difference between remembering someone's words and demonstrating that they were heard."

Theatre and consequence

It's easy to fake the first half. "A system could copy 'keep the colour' onto three screens and still propose exactly the same treatment it would have proposed without that preference. Repetition alone would be theatrical personalisation." The words travel. Nothing changes. The viewer sees their words everywhere and, if they're paying attention, notices that none of it mattered.

The stronger version is harder to fake because it has to show a consequence: "the preference visibly constrains the response: what is discussed, what is preserved, which options remain open, and what the person is — or is not — being asked to commit to." "Keep the colour" is heard when the plan keeps the colour and the options that would have changed it disappear.

Illustrative Generic service moments, not from any client.

The research agrees, from two directions

Listening research found the same split between seeming to listen and being understood. In an experiment with 180 participants in peer interviews, paraphrasing people's opinions back to them was "associated with the social attractiveness of the listener but [was] not associated with participants' conversational satisfaction or perceptions of feeling understood by the listener."²⁸ Repeating back made the listener likeable. It didn't make the speaker feel understood. (Those were student interviews, not marketing, so it's an analogue, not proof; but it's the right analogue.)

Information-systems researchers have a blunter name for the marketing version. Services that "actively use a user's name" without "offering personalized features other than using one's name" are "fake personalization, in other words, it is actually not personalization at all."²⁹

I'd said a version of it myself, about customer records, in The Manila Folder: "Personalised wording around a funnel is still a funnel. The show-it test asks whether you changed the object." Changed the object is the test here too.

The brand's real consistency

This is where being heard stops being a film technique and becomes a statement about the brand. Read the clinic's four films together and what they share isn't the look. My thinking partner's reading: "The family resemblance is deeper than [the palette]. It is customer agency carried through the experience."

"Different films can look and sound different while preserving the same relationship with the customer. Conversely, four beautifully colour-matched films can contradict one another about that relationship."

That "gives 'brand consistency' a much more interesting meaning": not the same hex codes on every asset, but the same thing happening to the customer's contribution every time. Read that way, the series is four expressions of one proposed relationship: different surfaces, the same promise about what happens to what you say. (The films one by one belong elsewhere.)

It's the patient-side version of something I'd written about founders in Made From You: "Your practical distinctions shape its decisions. You can correct what it understands." "Keep the colour" constraining a treatment plan is the same promise, made to a customer instead of a founder.

Designing for consequence, not echo

The being-heard test is easy to agree with and easy to fail, because echo is cheap and consequence isn't. An echo needs a text field. A consequence needs the builder to know what the service would have done otherwise, and to show the difference. So here's the working rule I'd give any builder, human or AI, working on anything that claims to listen.

None of this is specific to film. Chapter 8's illustrative proposal card is the same test in another medium: the buyer said the budget was tight, and the draft led with the larger package anyway. The fix wasn't a warmer sentence. It was changing what got recommended first. Wherever a person tells you something, the question is the same: what's different now because they said it?

In the harness, and at its limit

Chapter 6 gave being heard its test in the harness: "Where did the person's words change what happened?" The viewer has to be able to point to the moment. Chapter 10 showed that test sitting beside the skill that asks for it. This chapter is the principle behind both. The sibling book puts the ranking in two short sentences, "Being recognised is good. Being heard is better", and tells how the harness came to check for it.

And the limit, stated in the same breath as the principle, because it's the same limit this book has kept all along. The reading of the four films as one relationship was, in my thinking partner's own words, "my reading of the supplied explanation, not a measured audience finding." Whether viewers actually feel heard by a film that shows someone else being heard is a question about people, and no harness answers it. What the harness can do is make sure the film shows a consequence rather than an echo. Whether the consequence lands is for an audience to say.

"Their words change what happens" is a test for a patient in a film. It turns out to be exactly the same test for the person who made the film, and for anyone writing an account of the work.

My thinking partner noticed a symmetry after reading the records, and once you see it you can't stop seeing it.

The film, the process and the story about the process are all asking one question of three different contributors: did what you gave us change what we did next?

A contribution has been understood when it changes the subsequent work, not merely when it has been repeated back.

"That is a better test of 'being heard' than verbal acknowledgement." It's also the sentence I'd keep if I could keep only one from this book, because it's the one that tells you whether any of the rest happened.

Three scales, one test

The last row is the one that surprised me. "For this conversation, it is whether reading the records changes the account we would otherwise have written. And it does change the account." The discussion of the work passed the same test the films were trying to pass. If it hadn't, if the records had only decorated a story we'd already decided on, this book would be the theatrical version of itself.

A receipt for contributions

I'd built the institutional version of this test before, for organisations trying to prove their accumulated knowledge was doing any work. In Institutional Memory, Not Cognition the participation receipt had four fields: the prior claim that entered, why it was reached, its effect on the work, and who accepted it. The line that matters: "Without effect, there was no cognition — only display." And: "'The wiki helped' is not a test."

Taste work needs the same receipt, shaped for a rejection instead of a wiki page.

Notice what the third row asks for. Not "the skill now mentions it". The next artefact is different. That's a harder bar, and it's the right one, because a skill file can fill up with a person's words while every film it produces stays the same.

Filled in for the voice, honestly, it reads like this. The contribution: "That first voice is crappy compared to the rest — seems like older tech or something", mine, on the day. Where it landed: the skill's voice section (the ban, the audition warning, "cast from the audience's place" with its example), the harness (trust the listening judge on voices, not the transcript reader), and a contrasting pair (the weak voice against the recast one). What changed in the next artefact: every later build is refused a weak voice and told why, and that's recorded. What isn't recorded is that my remark caused the clinic's local casting; the agent did that on its own initiative. Who accepted: me, for the ban and the recast; for the clinic's voices, nobody yet: the record says I hadn't heard them. Reopen if: a local audience responds better to a voice that isn't local. Half of that receipt is clean, and the other half is an honest "not yet". Both halves are useful.

Ornamental memory

I've met that failure under other names. In Made From You the standard for a founder's corrections was that "if a correction matters, it should alter a proposal, a responsibility, a boundary or the decision to proceed", and the temptation it guards against was spelled out: "The easy response would be to acknowledge the preference and keep presenting the same founder-led service." The chapter's one-line version: "A correction that changes nothing is not participation."

And in Like Their Only One, written as commitments a service company makes to its customers, the learning commitment was deliberately demanding: "identify something we learnt, then show the later decision it improved. A growing library by itself proves very little." On expert judgment: "We have to demonstrate that transfer; recording an expert conversation does not establish it." And the sentence I keep coming back to: "Remembering a customer without adapting to them makes the memory ornamental."

The taste harness has its own version of ornamental memory: a skill file full of my verdicts, quoted beautifully, that no builder's output is different because of. It looks like a harvest. It's a scrapbook. The receipt catches it, because the third row stays empty.

What the autonomy is for

This changes how I think about the autonomy that impressed me so much that day. It's tempting to value it as absence: the system did all that without me. My thinking partner's reframing is the one I'd keep.

"The autonomy is not valuable because it removes you from the picture. It is valuable because your judgement can have consequences in places where you have not personally specified every decision."

That's the receipt's third row, turned into a purpose. Every film the platform's builders make now draws only on voices that survived a ban I asked for in one sentence about a farmer's voicemail. Every tall cut they make will be told to attempt less in its opening, because of two sentences I said about a clinic film. I won't see most of those films. That's what I mean by leverage, and it's the only kind worth having: my contribution changing work I never touched, in a direction I'd have chosen.

Every scale in the table above has evidence except one. My contribution changed the films I watched, and the instruments I corrected, and the account I read. It hasn't yet been shown to change the work of a builder that never met me, on a company it has never seen. That's the harvest's real test, and it's designed in the next chapter.

Chapter 3 previewed the standard and Chapter 14 demoted eloquence. Here's the standard in full, because everything in this book is answerable to it.

The proof is not how impressive the resulting essay sounds. It is whether a fresh successor, using the harvested skill and evidence, can avoid a seductive wrong answer and recognise a genuinely good unfamiliar one, without requiring you to repeat the whole apprenticeship.

Every phrase in that sentence is a design constraint. Fresh successor: no access to the sessions that formed the harvest. Using the harvested skill and evidence: it gets the carriers, not the history. A seductive wrong answer: something attractive enough to pass a shallow check. A genuinely good unfamiliar one: a right answer that doesn't look like the last right answer. Without repeating the apprenticeship: I don't get to steer it through the same rejections again. If I have to, the harvest didn't carry.

Test with resources, on a new case

Learning science has a precise version of why this test has to be designed carefully. Schwartz and Martin compared students taught by "tell-and-practice" with students who had first tried to invent their own approaches. On an ordinary test the two groups looked the same. But "the invention condition did better than the tell-and-practice condition when there was a learning resource."³⁰ Most assessments, they point out, are "sequestered problem solving": students "are shielded from contaminating sources of information that could help them learn".

That's the trap for testing a harvest. A recipe and a formed judgment can look identical on a familiar case. They separate only on a new case, and only when the tester is allowed to use the resources it was given. So the successor gets the skill, the contrasting cases and the harness, and a company none of them has seen.

What taste adds to a transfer test

I've designed transfer tests before. In Generative Design Patterns the regenerate test asks: "can an agent, given only the card plus a local domain sketch, produce something a third party recognises as a family member?" Chapter 6 used its sibling, the family test, on the week's vocabulary. And Chapter 9 used The Exemplar Ratchet's version: did you extract the shape, or only the wallpaper?

The successor test keeps all of that and adds the two things taste needs. Refusal: the successor has to decline an answer that would pass a check on polish, palette and plausibility. Recognition: it has to produce, or accept, a good answer that looks nothing like the first series. A family test might pass a successor that cloned the consultation note into a plumber's film. The successor test fails it.

Two rules from earlier books decide the order of operations. From Made From You: "State the contribution before seeing the output… Agree what would count as success and failure before producing an attractive result." And from Experience Is the Unit, the target must stay "free to contradict" what comes back. A pass line written after you've seen a lovely film is a description of the lovely film.

The sheet

Reading the result

The result has two halves, the builder and the harness, and they can fail independently. That's useful, because each combination points somewhere different.

And there's a fifth outcome that sits across all four. If my own verdict disagrees with a pass, that's a new test label, and the apprenticeship isn't over. It never quite is. The point of the harness was never to replace the verdict. It was to need it less often, and for harder cases.

How far a lesson may climb

This is where Chapter 11's ceilings get their evidence. Every lesson in the harvest arrived as a candidate from one series, and The Perturbation Review's rule applies: "n=1 may climb, and it arrives wearing its ceiling." A passed successor test is what lets a lesson move from "true for this clinic" to "a transfer hypothesis" to "a principle we expect under these named conditions". And some won't climb. The same book is blunt about that: "Most material should stop at levels one or two. That is not failed compounding. It is correct discrimination."

That's also the answer to the strongest objection to this book: taste formed on one client is local. Yes, some of it. The successor test is how you find out which parts, and the answer comes from running it, not from arguing about it.

Where it stands

Here's the evidence as it stands, without rounding up. An earlier version of the brand-film skill, harvested after the very first film, was tested once on a very different company, a farm spray-contracting concept, with a control and a transfer build. Its builder found its own spine, its own recognition moment and its own devices without copying the first film. That record is the sibling book's, and it's told there. The harvested skill with this book's lessons in it is now on all twelve company boxes, and their film agents pick it up on their next film task.

The successor test as designed here hasn't been run on it. Until it has, the harvest is a well-formed candidate, not a proven capability. The first sheet should be written before that next film starts, which is the only way it can be an honest test.

The test is designed around films. Nothing in it needs film. A proposal, a website, a report, a design: anything judged by fit has a seductive wrong answer and an unfamiliar right one, and the same harness can be pointed at it.

Strip the case away and here's what's left.

Taste, meaning judgment about fit between an expression, an audience, a situation and a purpose, can't be fully briefed. It forms through consequential attempts and diagnosed rejections. It travels only when it's carried three ways: a skill that makes, a harness that challenges, and contrasting cases that anchor the words. And the client who can't brief you generalises to any principal who can't specify what they need, which is most of them.

That describes most knowledge work that faces a person. A code review is judged by fit to a codebase and a team. A contract is judged by fit to parties who'll have to act on it. A policy explainer is judged by fit to a member of the public on a bad day. None of them has a spec that settles whether it's good. All of them have people who can say "no, not that" and struggle to say why.

What travels unchanged, and what doesn't

The method, across seven kinds of work

Here's the same machinery pointed at seven kinds of work. Every row is a sketch of the shape, not a result: none of them comes from this book's records, and the specific distinctions are examples of the form a distinction should take, not recommendations for your team.

Illustrative Each row is a shape to adapt. The last row is Chapter 8's worked card; the others are drawn the same way.

Read down the last column. Every one of those seductive answers would pass a shallow check: it's tidy, it's plausible, it reads well, and it's exactly what a capable model produces when it's been told about the task and not formed on it. That column is the reason this book exists. The generic default isn't bad work. It's work that fits nobody in particular.

Two rows, walked

In both, the most valuable leg is the one people skip. The artefact gets fixed either way. What stops the same failure recurring is the instrument leg, the reviewer that passed it, and the principle, stated with the condition under which it's wrong.

The principal who can't brief you

Part VI's client isn't only small businesses. A board that can't say what good governance reporting looks like. A service team that can't specify its ideal onboarding. The users of an internal tool who can't describe their own workflow, because they've never had to. The move is the same in each. Keep the three sources apart: what they've given you, what your professional judgment brings, and what production discovers (Chapter 20). Triage the gaps into already true, choices and hypotheses (Chapter 21). Notice the absences your framework makes visible (Chapter 22). And show them a concrete proposal to mark as represents, needs correction or adopt (Chapter 23). Never a questionnaire.

When the real verdict is slow

Films get a cold viewer's verdict in minutes. Some work doesn't. A contract's real test may be a dispute years later. A course's real test is what learners can do months after it ends. The harness still forms taste on the fast verdicts it can get: cold reviewers, counterparts, a colleague new to the codebase. What it mustn't do is fill the slow verdict's slot with a guess. Keep the slot open and labelled, the way this book has kept "no audience has been measured" in view throughout. The sibling book's habit of keeping intended, delivered and observed accounts side by side is the right shape for it, with one rule I'd copy exactly: "Mark 'not yet measured' until it's measured. A blank is information. A plausible guess in that cell destroys it."

Knowing the method travels is one thing. Having the forms on the desk on Monday morning is another.

You don't need a film pipeline to run any of this. You need the next three rejections you receive or give, and the discipline not to fix anything until you've written them down.

The Monday method

One habit runs underneath all eight: keep the rejections, not just the fixes, each with the condition that would reopen it. In Replay-Driven Design Evolution a folder of rejections "with reopening conditions" was on the short list of what should still exist if the first build burned down, because "the durable outputs… are the harness and the history". A rejection you've fixed and forgotten teaches the next builder nothing. A rejection you've kept, triaged and conditioned is the cheapest training data you'll ever own.

The forms, and where each is defined

How taste work fails: a field guide

Most of the failures below happened in the records this book comes from; the rest are the ones the research and my earlier books warn about. The tell is the thing you'll notice first; the leg is where to look.

How this book could be wrong

A doctrine that can't be wrong can't be tested, and Chapter 14 said an explanation that predicts nothing is a description. So here are the observations that would show this one wrong, stated with the same conviction as the rest.

Two products

Chapter 1 started with a page that explained the work better than I could, and a question: why would the explanation get better after the making? The answer took twenty-seven chapters, and it fits in the sentence my thinking partner wrote before any of them.

"The video is one product of the work. A more precise, transferable understanding of good video is the other."

Swap "video" for whatever you make. The artefact gets spent on the job it was made for. The understanding is the only part that compounds, and only if you keep it somewhere a successor can use: in a skill that makes, a harness that challenges, and cases that show what the words mean, each with the conditions under which it's wrong.

You know the test for this book, because it's the one Chapter 25 applied to everything else. Not whether it read well. Whether your next rejection changes something besides the artefact.