Lock the meaning
Ask the model to list every claim, assumption and number in the draft before touching the wording. Fix anything unsupported now.
The three passes
Meaning, then tone, then QA.
Each pass has one job and one prompt. Do them in order.
Fix the tone
Give the audience, the channel and the level of directness. A board memo, a support email and a LinkedIn post need different language.
QA line by line
Compare original and rewrite for missing caveats, softened warnings and invented details. Fix by hand, not with another prompt.
Prompts
One prompt per pass.
- Pass 1 — audit: “List every factual claim, number and assumption in this text. Mark any that are vague, unsupported, or stronger than the evidence given. Do not rewrite anything yet. Text: [paste]”
- Pass 2 — rewrite: “Rewrite this for [audience] on [channel]. Keep every claim from the audit exactly. Remove generic AI phrasing, shorten long sentences, and make the next step for the reader explicit. Text: [paste]”
- Pass 3 — diff: “Compare the original and the rewrite sentence by sentence. List every place where the rewrite makes a claim stronger, weaker, or different. Original: [paste] Rewrite: [paste]”
Example
The same paragraph through all three passes.
Original output
Our solution significantly improves team efficiency and has been shown to deliver outstanding results for organizations of all sizes.
After the passes
In the March pilot, the support team closed tickets 22% faster. We have not yet tested this outside teams of 5–20 people.
Pass 1 caught two unsupported claims (“significantly”, “all sizes”). Pass 2 replaced them with the pilot numbers. Pass 3 confirmed the limitation stayed in.
Why it works
What each pass catches that the others miss.
The passes are not redundancy — each one looks at the text through a different failure lens.
| Pass | Catches | Typical find |
|---|---|---|
| 1 — Meaning audit | Unsupported claims, inflated scope, missing sources. | “Studies show” with no study; “all users” meaning “the three users we asked”. |
| 2 — Tone rewrite | Wrong register, buried requests, filler structure. | A support reply written like a press release; the ask in paragraph four. |
| 3 — Line diff | Drift introduced by pass 2 itself. | A softened warning: “do not deploy on Friday” became “consider deployment timing”. |
Long documents
Scaling the passes to multi-page output.
For anything longer than a page, run pass 1 once over the whole document and turn its result into a locked-claims list — a short, numbered inventory of every fact, figure, promise and caveat. Then run pass 2 section by section, pasting the locked-claims list into every prompt. This solves the two problems long rewrites always hit: the model forgetting constraints established three sections ago, and the human reviewer losing track of which claims were where. The claims list is also reusable — the next time you regenerate that document, pass 1 is already done.
Pass 3 on long documents is best done backwards: diff the last section first. Quality of AI output degrades toward the end of long generations, and reviewers' attention degrades in the same direction — so the worst text reliably meets the least attention. Reversing the reading order puts your freshest scrutiny on the weakest material.
A note on tooling: none of this requires anything fancier than a chat window and a text editor with two panes. If you run the same rewrite frequently — a weekly report, a recurring client letter — save the three prompts with your locked-claims list as a template. The workflow's value compounds when the passes stop being typed from scratch.
Failure modes
Where the workflow breaks down — and the guard for each.
“The draft looks fine”
Confident prose hides unsupported claims better than clumsy prose does. The audit pass exists precisely because fluency and accuracy are uncorrelated in model output.
“Fix meaning and tone together”
One prompt doing both jobs lets tone preferences vote on factual content. Keep the passes separate even when you are in a hurry — especially when you are in a hurry.
“Ask the model if it changed anything”
A model grading its own rewrite is a conflicted witness. Pass 3 can be model-assisted, but the final diff read is yours — it is five minutes, and it is the whole point.
FAQ
Rewriting ChatGPT output: common questions.
Why three passes instead of one prompt?
Because meaning and tone pull in different directions. A model asked to fix both at once tends to smooth over claims. Separating the passes keeps each one checkable.
Which pass matters most?
The QA pass. Most people skip it, and it is where softened warnings, dropped caveats and invented details get caught before they reach a reader.
Does this work with Claude and Gemini too?
Yes. The passes are model-independent. Some people run pass 2 in a different model than pass 1 to get a second opinion on tone.
How long does this take?
About ten minutes for a one-page text. Far less than the time you lose when a client or teacher finds an error the rewrite introduced.
Related