AI Strategy · LeverageAI · The Outbound Mirror
Newsjacking with a Canon: Commentary at the Speed of the Feed
Everyone replying to that tweet is deriving their take right now — hot, thin, sourceless. Yours was compiled months ago, and the agent just looks it up. Point a body of work at other people’s posts and you get a combination nobody can match without the substrate: depth at the speed of the feed, and every post doubling as a receipt.
Scott Farrell — LeverageAI · A field guide for builders with a second brain · ~13 min read
TL;DR
- The move: commentary on an influential post beats self-sourced content because it inherits attention that’s already flowing. You surf a wave instead of making one. Newsjacking — but with a canon behind it, which changes what the newsjack can be.
- The edge: everyone else replying is deriving their take at tweet-time. Yours was compiled months ago and the agent retrieves it — “here’s what this means, here’s where I published the same mechanism in January with receipts, here’s where it breaks, here’s the rung he didn’t take” — within the hour, grounded in five hundred pages. Unmatchable without the substrate, and visibly so.
- The build: it’s the inbound curator re-terminated — the four diff classes become four content formats, the briefing queue becomes a draft-post queue, and one discipline is non-negotiable: a named human at the gate reading every word. That’s the whole line between a canon and a slopcannon.
In the previous piece I built a filter that reads the AI discourse for me — a curator that diffs every incoming tweet against a wiki of what I actually think and, most of the time, stays silent. It points inward: the world arrives, my worldview ranks it, and a briefing lands on my phone instead of a feed swallowing my evening. This piece is about turning that same engine around. Because the instant you have a machine that can say “here’s what Karpathy’s tweet means in your world, here’s where you already argued it, here’s the extension he didn’t take,” you have not just a reader. You have a publisher — and the thing it publishes is better than what everyone else is posting about the same tweet, for a reason that is structural rather than a matter of trying harder.
Here is the problem it solves, stated the way it actually shows up. I’ve written something like seventy articles and a stack of ebooks in eight months. Almost nobody reads an ebook. An excerpt and a nice image do a little better. But the honest gap in all of it — the thing my own assessment kept naming — was never the quality of the ideas. It was distribution: strong corpus, weak reach, and no appetite to become a full-time poster to fix it. “Post more original content” is the standard advice, and it’s exactly wrong, because it asks the scarcest input — my attention — to manufacture the second-scarcest — yours. The move that breaks the trade is to stop manufacturing attention and start borrowing it — the old media move called newsjacking,1 now with a canon behind it.
Your take was written in January
When Karpathy posts something and the timeline fills with replies, look at what everyone in that thread is actually doing: deriving a take, live, in the same ten minutes, from the same tweet, with nothing behind it. Hot, thin, sourceless — not because they’re lazy, but because a tweet is all they’ve got to work from. Their take is exactly as deep as the time between reading the post and hitting reply.
Mine isn’t, and the difference isn’t effort — it’s inventory. My take was compiled months ago and written down. The agent doesn’t reason a response to the tweet; it retrieves the position I already hold and diffs the tweet against it. This is the index-is-the-data argument pointed at content instead of retrieval: the value isn’t in generating a hot response on demand, it’s in having the compiled understanding sitting there, pre-computed, waiting to be looked up. What lands within the hour is a whole structured reaction — here’s what this means, here’s where I published the same mechanism in January with a timestamp, here’s the point where it breaks, here’s the next step the author didn’t take — grounded in five hundred pages that no one else in the thread can reach.
Depth at the speed of the feed is a combination nobody can match without the substrate — and, crucially, it’s visible that they can’t.
That visibility is the part worth dwelling on. Anyone can post a confident opinion. What no one can fake in real time is “I wrote the mechanism you’re describing, here, on this date, and here’s the part you haven’t got to yet.” A receipt with a timestamp is unforgeable by improvisation. So every post the machine drafts is doing two jobs at once: it’s commentary on the news, and it’s evidence of intellectual priority — a public, dated marker that you were early to the idea now trending. The distribution engine and the credibility engine turn out to be the same engine. You newsjack the attention and bank the receipt in one motion.
The diff classes were a content taxonomy all along
The curator sorts every incoming item into one of four buckets by asking one question — how does this differ from what I already believe? Those four diff classes — contradicts the canon, independently converges, genuinely novel, already known — were defined for inbound triage, to decide what deserves an interrupt. Turn the engine outward and something clean happens: the same four buckets become a content-format taxonomy. Each class of diff has a natural shape of post, because the relationship between the tweet and your canon is the angle. You don’t choose a format; the diff hands you one.
| Diff class (from the curator) | Becomes this post format | A worked item off the feed |
|---|---|---|
| Independently converges | The receipts post. “Karpathy just described the janitor pattern; here’s mine, published in June, timestamped.” Confirmation from someone credible, plus proof you were early. | He posts that agents should file their own answers back into a wiki as new pages. You wrote that months ago. |
| Contradicts the canon | The respectful-disagreement post. Reliably the highest-engagement format, because it’s a genuine stake in the ground — a real position, not a nod. | A respected voice argues maintained knowledge-graphs are a dead end and long-context plus plain RAG wins. You think that’s wrong, and you can say why, with structure. |
| Genuinely novel | “One rung further.” Extension, not restatement: “he stopped one rung short — here’s the next rung,” read through your frameworks. | A thread shows a trick you have no page for. You extend it into your own model and hand back the step they didn’t take. |
| Already known | Nothing. Silence. No post. This is most of the feed, and not-posting is the discipline that protects every post you do ship. | The daily thread re-explaining what RAG is. You have nothing to add, so you add nothing. |
Read that top-to-bottom and notice the ordering is doing work. The two classes that make the strongest posts — convergence and contradiction — are the two where your canon has the most to say, because you’ve already staked a position the tweet either confirms or challenges. The respectful-disagreement post in particular is the one I’d bet on: a genuine, reasoned stake in the ground reliably outdraws agreement, and the canon is what lets you disagree with substance instead of contrarian noise — you’re not being difficult, you’re pointing at the page where you already worked out why. And the bottom row — “already known,” the overwhelming majority of any feed — produces the most important output of all, which is nothing. We’ll come back to why the silence is load-bearing.
One worked example, end to end
Abstractions are cheap; here is the actual path a single item walks, from a tweet scrolling past to a post I approve from my phone. This is a convergence-class item — the receipts format — and I’m showing the whole pipeline because the human gate in the middle is the part that matters most.
# DRAFT QUEUE — Tue 09:41 · item 1 of 2 this week # SOURCE: @karpathy, thread (link) DIFF CLASS: converges-on-canon 1 · TWEET "You shouldn't dump raw docs into context. An agent should read each source and fold it into a maintained, interlinked wiki it writes and curates itself." 2 · DIFF matches: framework.wiki-graph, concept.write-back, post.index-is-the-data (published 14 Jun, timestamped). relation = CONVERGES. seam = he keeps the *answer*; your scout loop also keeps the *path*. that half is still yours. 3 · DRAFT format: RECEIPTS POST (quote-tweet, attribution native) ---------------------------------------------------------- "Strong agree — and there's a receipt. I made this exact argument in June: the index IS the data, the wiki out- thinks RAG for agents. [link, 14 Jun] One rung further than the thread goes: don't just file the *answer* back — keep the *exploration path* too. The dead ends are telemetry about your map's missing edges." ---------------------------------------------------------- 4 · GATE -> Scott. approve · edit · kill. [waiting] ✓ on approve: post; log take + engagement to signals; file as derived page.
Walk the five steps and notice what each stage is and isn’t doing. The tweet is just a source, exactly as it is for the inbound curator. The diff is the only expensive cognition, and it’s the step no competitor can run, because it requires my whole corpus to compute — it finds the matching pages, the publication date, and the seam (the rung the author didn’t take). The draft is written to the format the diff class dictated, quote-tweeting natively so the attribution is structural rather than something I have to remember to add. Then the gate: it stops, and waits for me. And on approve, the loop closes — the post ships, and the take plus whatever engagement it draws get logged back. The machine did everything except the one thing that’s mine to do.
Four disciplines, all cheap, all load-bearing
Left unconstrained, this becomes exactly the thing it’s trying to beat: a firehose of confident, forgettable takes. Four disciplines keep it on the right side of that line. None of them cost much; all of them are the difference between the tool being an asset and being an embarrassment.
1. You are the terminal judgment — permanently, not just at first
Every draft lands in a queue. You approve, edit, or kill it from your phone, and nothing ships without that tap. This is not caution about “automation maturity,” a phase you graduate out of once the model gets good enough. It’s permanent, because your taste is the product and the un-fakeable ingredient. The canon supplies the substance; your judgment supplies the reason anyone should trust the account attached to it. Automate the drafting all the way to the edge and keep the last inch human forever — that inch is the whole brand.
2. Deconstruct what was said; never construct what wasn’t
Quote-tweet natively, so the source is structurally present and the reader can always see the original words next to your reading of them. The take is always your interpretation of their post — never a paraphrase that could be mistaken for a claim they made, never words put in their mouth. This is the same North Star my quote-card pipeline already runs on: get behind what it means; don’t repeat what the reader can already see. Restating the tweet is worthless — they can read the tweet. The value is entirely in the delta you add, and keeping the original visible is what makes the delta legible as yours.
3. A post budget, exactly like the interrupt budget
The moment this runs, the temptation is volume — and volume is precisely wrong. Selectivity is the brand. Most tweets score “already known” and should die silently; the account’s signal-to-noise ratio is its reputation. Give it a hard cap — two or three posts a week — and the scarcity forces real ranking, the same way the curator’s interrupt budget forces it to decide what a genuine contradiction looks like. Two or three excellent takes a week beats two a day, forever. An account that posts only when it has a receipt or a real disagreement teaches people that its posts are worth stopping for.
4. Decay physics — the one place a slow loop kills the value
Everywhere else in this stack, patience is free: a wiki page is just as true next week, so the janitor can run on a lazy cron and nothing is lost. Commentary is the exception. Its value decays in hours — a perfect take on a two-day-old tweet is worthless, because the discourse has moved and the attention you were borrowing has dispersed. So this is the single component in the whole architecture where the loop has to run tight: the draft must be in your queue while the conversation is still live. Get the physics wrong here and a brilliant post arrives at an empty room.
The pattern under all four
Speed comes from the machine (canon + fast draft, on a tight loop). Trust comes from you (a named human at the gate, ruthlessly selective). Automate the perishable half; keep the compounding half human. That split is the entire operating manual.
The one rule that separates a canon from a slopcannon
There’s a fair objection sitting under all of this, and it deserves to be met head-on: isn’t an AI that auto-drafts commentary on trending posts just a slop cannon — the exact machinery flooding every timeline with generated filler? The contempt is real and it’s earned; the discourse around AI “slop” is full of people rightly sneering at accounts that point a generator at whatever’s trending and fire.2 If that’s what this were, it would deserve the sneer. It isn’t, and the difference reduces to two conditions you can actually test.
A real canon underneath, and a named human on top. Those aren’t vibes; they’re falsifiable. State the boundary as rules and you can check any account against them:
Slopcannon (fails the tests)
Reacts to the trend with no prior position — the “source” is the tweet itself. No dated receipts, because there’s nothing that pre-existed the tweet. Optimises for volume and reach. No human reads it before it ships. Restates or vaguely agrees. Attribution is decorative.
Canon-grounded commentary (passes)
Every take diffs against a written body of work that predates the tweet. Convergence posts carry a checkable timestamp. Optimises for selectivity — most items produce silence. A named person approves every word. Adds a delta the reader couldn’t get from the source. Attribution is structural (native quote-tweet).
The tests are sharp on purpose. Does a take trace to a page that existed before the news? Is there a real name accountable for what shipped? Does the account post rarely, and only to add something? A slopcannon fails all three by construction — it has no canon to trace to, no one signs it, and its entire logic is volume. Canon-grounded commentary passes all three, and passing them isn’t a matter of tone or disclaimers. It’s a matter of whether the substrate and the gate are actually there. Which is the reassuring part: the thing that keeps you from becoming the slopcannon isn’t restraint or good taste in the abstract — it’s two structural facts about your setup that anyone can audit.
It’s the same engine, pointed outward
None of this is a new system, and that’s the whole reason it’s worth building. The inbound curator and this outbound commentator are one pipeline with two terminal tools. The scout still walks the wikis anchored on an incoming tweet; the senior still diffs it against the canon and finds the seam. The only thing that changes is the last step. Inbound, the senior emits a briefing and the terminal tool is a push notification to my phone. Outbound, the senior emits a draft and the terminal tool is a post to a queue. Same conversation-handoff loop, same map, same diff. The curator’s briefing queue, re-terminated as a draft-post queue — and my approval of the briefing simply becomes the publish gate. The marginal cost of the entire distribution capability is one North Star and one output tool.
And the drafting-into-fragments half is already a solved, running system: I’d earlier built a pipeline that breaks my ebooks into quotable pieces and renders them as images — the proof that depth can be industrialised into shippable fragments, though the image mechanics are their own story.3 The commentary engine slots straight onto that: the take gets written, the quote-card machinery makes it a visual, and the same human gate approves both. Every part was already on the shelf; this just wires them into a new destination.
This is the last bolt in a flywheel I’ve been describing for a while. I used to think I was writing for an audience; then I realised the audience was me and the next ebook, and each article was really a compression pass on my own worldview.4 The gap that reframe left open was distribution — all that compiled understanding, under-read. This closes it, using nothing but parts already built, running on other people’s reach, and publishing evidence of the canon’s priority as a side effect. The corpus starts doing its own marketing.
The loop closes: engagement is signal, filed back
There’s a final turn that makes this compound instead of merely repeat. When a post ships, its afterlife is data. Which takes landed and which fell flat; which frameworks resonate with outsiders rather than just with me; which objections keep recurring in the replies. All of that gets harvested and filed back into the wiki as derived pages — so the canon learns what the world finds interesting about itself. The same compression that built the worldview from my own writing now also runs on the world’s reaction to it, which is recursive worldview compression fed by an outside signal for the first time.4
Notice what that does to the next post. The curator now knows not only what I believe, but which of my beliefs the market keeps arguing with, which receipts drew the most nods, which extensions people asked me to take further. So it drafts better — not because the model improved, but because the substrate learned. The account gets sharper the longer it runs, in the direction the audience keeps pointing it. Inbound, the feed got quieter as it got smarter. Outbound, the posts get better-aimed as the canon absorbs its own reception.
I didn’t set out to build a distribution engine. I set out to stop drowning in the feed, kept a wiki of what I actually think, and pointed it inward to read the world for me. Turning it around cost almost nothing — one North Star, one output tool — and it turned five hundred pages of quiet corpus into a thing that publishes depth at the speed of the feed, banks a timestamped receipt every time an influential voice catches up, and gets sharper each week from the arguing. The audience stopped being just me and the next ebook again. Only this time, the machine writes the first draft, and the one thing it never does is press send.
The reach was never the problem. The substrate was.
If your organisation is sitting on real expertise that nobody reads — whitepapers, case studies, hard-won internal know-how decaying unread — that’s not a “post more” problem, it’s a substrate problem. At LeverageAI we compile the body of work into a worldview first, then point cheap agents at it: a curator that reads the discourse, and a commentator that publishes your take on it — grounded, selective, and always gated by a human whose name is on it. Talk to us about building the canon before the cannon.
References
- [1]David Meerman Scott. Newsjacking: How to Inject Your Ideas into a Breaking News Story and Generate Tons of Media Coverage (Wiley, 2011). — Coined “newsjacking”: injecting your angle into a story whose attention is already surging, to inherit reach you didn’t have to manufacture. The canon-grounded version inverts the usual gimmick — the take pre-exists the story rather than being improvised for it. newsjacking.com
- [2]Andrej Karpathy. “LLM Wiki” (gist, April 2026), and its comment thread. — The archetypal influential source whose posts this engine comments on; also the substrate concept (an agent maintains an interlinked wiki, filing answers back so exploration compounds). The comments capture the wider contempt for “slop” — generators pointed at whatever is trending — which is exactly the boundary a real canon plus a named human gate is built to clear. gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
- [3]LeverageAI — the quote-card pipeline (ebooks → quotable fragments → images) is the proven industrialisation of depth-into-fragments that this commentary engine reuses; its image mechanics are a separate field note. leverageai.com.au
- [4]LeverageAI — related canon (context, not statistics): A Newsfeed That Hunts Its Own Blind Spots (the inbound sibling — where the four diff classes are defined), The Index Is the Data (the compiled worldview a take is retrieved from), The AI Learning Flywheel (audience-is-me-and-the-next-ebook), Worldview Recursive Compression (the loop the engagement harvest feeds), and The Proposal Compiler (marketplace-of-one economics). leverageai.com.au
Discover more from Leverage AI for your business
Subscribe to get the latest posts sent to your email.