The Four Steps

# The Four Steps

This is a free kit — two short markdown files with the exact prompts to paste — for anyone who asks an AI for something and gets back confident-sounding garbage. It is not a tool, an app, or a subscription. It is four things you do in a different order, and they work on an email or a business plan just as well as on code.

The problem usually is not the model. It is the order you asked in.

## Step 1 — ask for the plan, not the thing

Do not ask for the finished output first. Ask what it intends to produce, read that, fix it, and only then let it build.

> Before you write anything, give me your plan for this in 5 bullets: what you're going to produce, the structure, and any assumption you're making about what I want. Do not start until I say go.

The decisions that ruin an output get made silently in the first few seconds. A plan makes them visible while they are still free to change. Rewriting one bullet costs nothing. Rewriting a finished draft costs you the hour.

If you agree with all five bullets immediately, the plan is too vague. Ask it to get specific about the part you actually care about.

This is the step almost nobody does.

> Which parts of this are you least confident about, and why? Be specific and rank them.

It will tell you. A model does not experience doubt, but it does carry a real signal about where it is interpolating rather than standing on solid ground — and ordinary conversation never asks for that signal, so the same confident tone ends up covering a well-known fact and a total guess.

Every low-confidence spot it names is where the garbage was going to come from. You are reading the error report before the error.

For anything it ranks high, either supply the missing fact or tell it explicitly what to assume. Do not let a guess stay a guess just because it was disclosed.

Not the first twenty percent of the whole thing. A whole thing, at twenty percent size.

Writing a report? One complete section, intro to conclusion, before the rest. Building a page? One working path end to end, ugly, before any styling. Writing an email sequence? One complete email, not five rough openings.

A small complete version exposes the problems that only show up at the joins — the parts that have to connect to each other. A half-built big thing hides them until you have written too much to want to start over.

Take what you got, open a different model, and paste it in.

> Here is something another AI produced for this task: [paste]. What is wrong with it? Be specific and harsh. What would you have done differently, and why?

A model is a poor reviewer of its own output, because the same assumptions that produced the mistake also make it invisible. A different model has different blind spots, so the two do not overlap. In the interview these steps come from, running two different models over one project surfaced ten problems, and all ten were fixed.

ChatGPT, Claude, Gemini and Copilot all have free tiers. Any two of them is enough. You do not need a paid plan on either one.

It is slower at the start — about two extra minutes, every time. On a throwaway one-line request that is not worth it. Use this when the output actually matters.

Step 2 is not a lie detector. It surfaces what a model is *uncertain* about, which is a strictly smaller set than what it is *wrong* about. It can still be fluently, confidently wrong about something it never flagged. Treat the ranking as a place to look, never as a score.

The vocabulary around this is also contested, and practitioners genuinely disagree about how much a person should do versus hand to the agent. Treat these four steps as one working practice, not a settled standard.

Two ways to get it, and they are not the same. Comment WORKFLOW on the reel and it comes straight back to you, no email, nothing to fill in. Or download it from the button on this page, which asks for an email first. Same two files either way.

• THE-WORKFLOW.md — all four steps, each with the exact prompt to paste, what it is actually doing, what a good answer looks like, and the honest limits of each one.
• THE-QUESTION.md — step 2 on its own, with five variants aimed at vague briefs, factual claims, plans, code, and a model that is agreeing with you too easily.

## Credit

The four steps come from Dex Horthy of HumanLayer, interviewed by David Ondrej. His *12-Factor Agents* paper carries the earliest documented use of the phrase "context engineering" that we could find — though he does not claim sole credit for it, calling it possibly a Newton-Leibniz moment, and neither do we.

The step-2 question is Victor Taelin's, credited in that interview. It is not ours. The five variants in THE-QUESTION.md are.