The Tuesday Afternoon That Changed Everything

It was 2:47 PM on a Tuesday when Sarah, the tech lead at a mid-sized fintech, realized they’d been doing it wrong for six months.

Her team had been wrestling with API latency issues. Customer complaints were stacking up. The engineering squad had tried everything: Redis caching layers, database query optimization, load balancer tuning, even a complete rewrite of their authentication middleware. Twelve different “promising approaches” over six weeks. Each one delivered marginal gains—3% here, 8% there. The cumulative improvement: 12%.

Not enough. Not even close.

Sarah had just finished yet another standup where the team debated whether to try horizontal scaling or invest in a CDN. The energy in the room was… depleted. They were competent engineers chasing symptoms, not solving the core problem.

That evening, she stumbled across a blog post about “Pre-Thinking Prompting” while searching for system design patterns. The opening line stopped her cold:

“You’re asking your AI—and your team—to do two jobs at once: understand the problem space AND solve the problem. No wonder you’re getting mediocre results.”

Two weeks later, her team had shipped a solution that reduced latency by 58% for 73% of queries. More importantly, they’d found a repeatable process for tackling complex problems.

This is that story. And the framework behind it.

Chapter 1: The Context Starvation Problem

Let’s start with why most AI interactions—and most team brainstorms—produce shallow outputs.

When you send a prompt to Claude, GPT-4, or any large language model, you’re typically doing something like this:

Seems reasonable, right? You’ve specified the domain, the tech stack, even framed it as an expert consultation.

But here’s what’s actually happening under the hood: the model has to simultaneously:

• Infer the problem space — What kind of latency? Network? Database? CPU? Authentication logic?
• Guess the constraints — Can we change the database? Is horizontal scaling an option? What’s the budget?
• Assume the context — Is this a read-heavy or write-heavy workload? Are there compliance requirements? What’s the current architecture?
• Generate solutions — All while burning tokens on exploration instead of quality reasoning

The model—and your team in that standup—is context-starved. It’s like asking someone to navigate a city without giving them a map, then wondering why they took wrong turns.

The Two-Job Trap

This is what I call the Two-Job Trap: collapsing problem understanding and problem solving into a single pass.

Think about how your best brainstorming sessions actually work. The productive ones don’t start with “How do we solve X?” They start with:

• “What exactly is X? Let’s define it.”
• “What have we tried? What did we learn?”
• “What are the hard constraints we can’t change?”
• “What’s the tension here? What are we trading off?”

Only after that groundwork do you generate solutions. The context-building phase is the scaffolding that makes good ideas possible.

Sarah’s team had been skipping that phase entirely. So had their LLM prompts.

Chapter 2: Enter Pre-Thinking Prompting

Pre-Thinking Prompting (PTP) is stupidly simple in concept: generate structured context about the problem before you ask anyone—human or AI—to solve it.

It’s not a longer prompt. It’s not a better preamble. It’s a separate, disciplined pass that fabricates the missing scaffolding on purpose.

When Sarah ran her first PTP session with the team, it took two hours. They didn’t write a single line of code. They didn’t even propose solutions. They just… thought. Systematically.

The Four Moves of PTP

Here’s the framework Sarah’s team used, and the one that’s now become my default for any complex problem:

Time investment: 10-15 minutes for simple problems, 1-2 hours for gnarly ones. The payoff? You eliminate 3-5x more wasted cycles downstream.

What Sarah’s Team Discovered

During their PTP session, Sarah’s team filled a whiteboard. Here’s what emerged:

Problem statement (one sentence): “Authentication API P99 latency exceeds 800ms, causing checkout abandonment, primarily due to synchronous validation of user credentials against legacy LDAP during peak hours.”

That alone was clarifying. They’d been saying “API is slow” for weeks. Now they had specificity: P99 latency (not median), authentication step (not database), peak hours (timing matters), LDAP validation (the actual bottleneck).

The contradiction: Speed↑ vs Reliability↑

• Why speed matters: Every 100ms of latency costs 2.3% conversion (they had data)
• Why reliability matters: Financial regulations require real-time credential validation; no caching user auth state
• Current sacrifice: They were prioritizing reliability, accepting slow auth

Reframing that unlocked the solution: “What if we pre-compute high-confidence validations and only synchronously validate low-confidence or flagged accounts?”

This reframing came from a constraint flip: instead of “all validations must be real-time,” they asked “what if only risky validations are real-time?”

Chapter 3: The PTP-to-TTC Connection

Here’s where it gets interesting for AI practitioners.

Modern reasoning models—OpenAI’s o1, Anthropic’s extended thinking, Google’s chain-of-thought—use something called test-time compute (TTC). Instead of just generating one response, they spend extra inference cycles to “think longer”: sampling multiple reasoning paths, searching over intermediate thoughts, running internal deliberation before finalizing an answer.

The research is clear: performance improves when models think longer, not just when they’re trained bigger.

But here’s the nuance most people miss: TTC is a search procedure. Without structure, it’s blind search.

PTP transforms blind search into guided exploration.

How PTP Shapes the TTC Search Space

In practice, this means:

• When you run a PTP pass and feed the output to a reasoning model, each sampled chain can explore a different reframing
• Contradictions become explicit axes the model searches across
• Assumptions become checkpoints for reflection and validation

The model doesn’t just “think harder.” It thinks across structured dimensions.

Chapter 4: Making It Practical

Theory is nice. Execution is what matters. Here’s how to actually use PTP in your workflow.

When to Trigger PTP

Not every problem needs a full PTP pass. Use it when:

• Input is thin or ambiguous — You know there’s a problem, but the contours are fuzzy
• Multiple stakeholders with conflicting needs — Marketing wants speed, Legal wants compliance, Engineering wants simplicity
• History of “lots of ideas, weak outcomes” — You’ve been throwing spaghetti at the wall
• High stakes or irreversible decisions — Architectural choices, regulatory commitments, major refactors

The 15-Minute PTP Recipe

For smaller problems, you can run a lightweight PTP in 15 minutes:

1. Set a timer. Seriously. PTP is not analysis paralysis—it’s disciplined pre-work.
2. Write the one-sentence problem statement. If you can’t, the problem isn’t clear yet.
3. List 3 knowns, 3 assumptions, 2 unknowns. Tag each explicitly.
4. Identify 1-2 contradictions. “We need X↑ but also Y↑, and they fight because…”
5. Generate 3 reframings. Constraint flip, inversion, or level shift.
6. Define 1 first-signal test. What could you measure in ≤2 weeks that would change your approach?

That’s it. No special tools. Just structured thinking.

Automating PTP with Agents

The beauty of PTP: it’s ideal for agentic workflows.

You can create a “Breakdown Agent” that:

• Takes a messy problem statement as input
• Runs the four PTP moves (Strip, Stretch, Stress, Stage)
• Emits 3-8 mini-briefs for downstream agents (Thinker, Implementer, Reviewer)
• Maintains an Assumption Ledger that updates as signals arrive

This is how multi-agent systems avoid context thrashing. Each agent gets a clean, scoped brief instead of the original chaos.

Tools You Don’t Need

PTP works with:

• A text file and 15 minutes
• A whiteboard and a team
• A Markdown template (I’ll share one at the end)

You don’t need special software, frameworks, or certifications. Just discipline.

Chapter 5: Why “More Domain Words” Helps (But Only If Structured)

There’s a common heuristic in prompt engineering: “Add more domain-specific context.”

It works, but not for the reason most people think.

Dumping extra domain tokens can help by activating relevant regions of the model’s latent space—the manifolds associated with that domain’s concepts. But unguided verbosity can anchor the model incorrectly.

PTP solves this by organizing extra words into axes: constraints, contradictions, metrics, reframings. The model doesn’t just see more tokens—it sees structure it can systematically traverse.

Think of it like this:

Both have domain terms. Only the second gives the model navigable structure.

Chapter 6: Common Pitfalls and How to Avoid Them

Pitfall 1: Analysis Paralysis

Symptom: PTP session drags on for hours; no output, just endless refinement.

Fix: Time-box each move. Strip (5 min), Stretch (7 min), Stress (5 min), Stage (8 min). Use a timer. Good-enough beats perfect.

Pitfall 2: Reframings That Only Swap Nouns

Bad reframing: “How do we reduce latency?” → “How do we improve response time?”

Good reframing: “How do we reduce latency?” → “What if we accepted higher latency for low-priority users and allocated resources to high-value transactions?” (constraint flip + segmentation)

Fix: Each reframing should open a genuinely different solution space. Ask: “Would this lead me to explore different approaches?”

Pitfall 3: Assumptions Disguised as Facts

Symptom: “We know users want faster load times” (tagged as Known)

Reality: That’s an assumption unless you have data.

Fix: Be honest. Tag it Assumed, add it to To-Validate, and define a test. Bounded speculation is fine; silent guesses are not.

Pitfall 4: Skipping the First-Signal Test

Symptom: You run PTP, generate ideas, then… build for weeks without validation.

Fix: Always define a ≤2-week test. “Shadow endpoint for 5% of traffic” or “Manual validation with 10 users.” Get signal fast.

Chapter 7: The Bigger Pattern

PTP isn’t just a technique. It’s an instance of a broader pattern I call Speculative Briefing.

The meta-move:

When input is thin and stakes are high, run a structured pass that (1) decomposes the problem, (2) fills gaps with labeled assumptions, and (3) emits falsifiable tests. This brief becomes the “fuel” for downstream work.

This pattern shows up everywhere:

• Engineering: Reverse PRD from a one-line feature request
• Research: Proto-hypothesis deck from a vague question
• Design: Reverse design brief from a “vibe”
• Operations: Draft runbook from a class of incidents
• Sales/Marketing: Reverse campaign brief from a one-paragraph goal

The common thread: manufacture missing context on purpose, tag guesses clearly, validate fast.

Sarah’s team now runs a 15-minute PTP before every sprint planning session. They call it “pre-flight.” It’s become as routine as standup.

Conclusion: Pre-Context the Context

Three months after that Tuesday afternoon, Sarah’s team shipped their best quarter ever. Not because they suddenly got smarter or hired more engineers. Because they stopped asking themselves—and their AI tools—to do two jobs at once.

They separated understanding the problem from solving the problem. They gave structure to ambiguity. They made guesses explicit and tested them fast.

The one-liner that stuck on their team room wall:

“Don’t write the task. Write the context that makes any task sane.”

That’s Pre-Thinking Prompting.

It’s not magic. It’s not a silver bullet. It’s just disciplined pre-work that turns vague requests into navigable landscapes, scattered constraints into explicit contradictions, and blind exploration into guided search.

When your AI outputs—or your team brainstorms—feel mediocre, the problem usually isn’t the model or the people. It’s context starvation.

Feed the context first. Everything else gets easier.