I Stopped Writing Giant Prompts. This 3-Step Pattern Fixed My AI Output

For months, my prompts kept getting longer. More context. More rules. More "please make sure to" stacked on top of the last failure. The outputs kept landing in the same place: fluent, confident, and about 70% usable. The rewrite took longer than writing the thing myself would have.

The fix was not a better mega-prompt. It was splitting one big ask into three small ones: Gather, Draft, Verify. Each step has one job, and I can see exactly where things go wrong.

A single giant prompt asks the model to do four jobs at once: find the relevant facts, decide what matters, write in the right shape, and check its own work. It does all four in one pass, in that order, and it is worst at the last two. The facts it skipped in step one never come back. The errors it made in step three get polished in step four instead of caught.

Three short prompts beat one long one because I sit between the steps. Thirty seconds of my attention at the joints saves a full rewrite at the end.

The chat interface at chat.openai.com — the pattern works here because each step is its own short exchange.

Goal: facts and raw material only. No writing yet.

"List the facts I need for [task]. For each one, give the source or mark it as unknown. Do not write the [document] yet."

Worked example. The task: turn last week's merged pull requests into release notes for users. My old approach pasted 14 PR titles into one prompt and asked for finished notes. The output invented a feature name and described one fix backwards.

"Here are 14 pull request titles and descriptions. [paste] Group them into: new features, fixes, and internal-only changes. Flag any where the description is too vague to explain to a user. Do not write release notes yet."

The model returned a clean grouping and flagged two vague PRs. I checked those two against the actual code changes. Total time: about two minutes, most of it mine.

Goal: write only from the gathered material, in a shape I specify.

"Using only the groups above, write release notes. One line per change, plain language, present tense, no marketing words. Internal-only changes are left out. If a change is unclear, write [CHECK] next to it instead of guessing."

Two things do the work here. "Using only the groups above" fences the model in; it stops reaching for plausible extras. And the [CHECK] instruction gives it a legal way to admit uncertainty. Models guess because most prompts offer no other move. Give them an escape hatch and they use it.

The chat interface at claude.ai — the same three-step pattern runs unchanged across chat tools.

"Read the draft against the grouped list from step 1. Report: (1) any claim in the draft that is not in the list, (2) any user-facing change in the list that is missing from the draft, (3) any line a user could misunderstand. List problems only. Do not rewrite yet."

This is the step the giant prompt could never do, because the same pass that wrote the text was grading it. A fresh pass with a checklist catches the backwards fix and the missing feature. Then I fix the flagged lines myself, or send one short prompt for just those lines.

On the release notes, Verify caught one invented detail and one omitted fix. The draft shipped after two small edits instead of a full redo.

Three honest limits before you try it: it costs three exchanges instead of one, so save it for work that matters. The Gather step is only as good as the material you feed it — paste real source text, not a summary of a summary. And Verify finds mismatches between draft and source; it cannot tell you the source itself is wrong. That check is still yours.

But for anything longer than a paragraph that other people will read, this split has ended my rewrite loop. The prompts got shorter. The output got better. That trade is available to anyone willing to sit between the steps for thirty seconds at a time.