Today's AI Tip
Want feedback, not a rewrite? Tell AI not to touch your draft.
What most of us ask
Tested in GPT-5.5 (OpenAI API), Oct 2026
We paste a draft and type this. In our test with Sam's fictional cover note, GPT-5.5 took it as permission to rewrite everything.
It rewrote the best line
Real output · GPT-5.5 (OpenAI API) · Oct 2026
The line about the late bus became "plans change, a bus is late, or a customer needs a quick, calm solution." That's not how Sam wrote it.
Smoother, but no longer Sam's
Only 1 of 5 body sentences survived unchanged. It also added things Sam never wrote, like "I was excited" and "to cancellations".
Source: Our test, GPT-5.5 (OpenAI API), Oct 2026
Ask for feedback, not a rewrite
Tested in GPT-5.5 (OpenAI API), Oct 2026
The model does what your words ask for. Saying "feedback" and "do not rewrite" keeps the draft in your hands.
Every quote matched the note
Real output · GPT-5.5 (OpenAI API) · Oct 2026
We checked all 5 quotes against Sam's note, and each matched word for word. Nothing was rewritten. Here are two of the five suggestions.
It's not perfect, so you decide
Suggestion 1 pushes toward a "more polished tone", the same push that made the first rewrite generic. You still choose: keep the bus line, and weigh cutting "not a natural salesperson".
Make every change yourself
Paste your next draft under the prompt and make every change yourself. Remove names, account numbers and other personal details first, and check the tool's data policy or your employer's approved tools.
Try it: Run it on an email you sent this week; count the suggestions you'd take.
Sources and assumptions
- Prompting best practices, Claude Platform Docs (Anthropic): Exact quote: "Claude's latest models are trained for precise instruction following and benefit from explicit direction to use specific tools. If you say \"can you suggest some changes,\" Claude will sometimes provide suggestions rather than implementing them, even if making changes might be what you intended." Also: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result." Scope: this is about Claude, in a section on tool use and taking action. It is context for 'wording decides suggest vs. rewrite', not evidence about GPT-5.5. (checked 2026-10-13)
- Our own test run, engine/ai-test.js (record.tests ids 'before' and 'after', tool chatgpt): Both GPT-5.5 (OpenAI API) outputs on the fictional cover note, run 2026-10-14T02:00 UTC (Oct 2026). 'Can you improve this?' returned a full rewrite with a subject line. The feedback prompt returned 5 numbered suggestions, each quoting the note verbatim, with no rewrite. (checked 2026-10-13)
Assumptions:
- The cover note, Sam Ortiz, Dana, Brightwell Tours and Harbor Kayak Co. are fictional, written for this test.
- Only GPT-5.5 via the OpenAI API was tested (one run per prompt, default settings, no system prompt). The Claude runs failed to launch in the research environment (output null in record.tests), so the slides must not claim any Claude result, and prompt_tools names GPT-5.5 (OpenAI API) only.
- The ChatGPT app may answer differently from the API, and repeat runs vary. The post shows one real run per prompt, not a guarantee.
- '1 of 5 body sentences unchanged' counts the five body sentences of the original (greeting and sign-off excluded) against the rewrite, by hand: only 'My CV is attached.' survives word for word.
- The 'added things Sam never wrote' list comes from comparing the rewrite with the original by hand: 'I was excited', 'help keep daily operations running smoothly', 'instead of losing them to cancellations', 'when plans change', 'a customer needs a quick, calm solution', and a subject line.
- All 5 quoted phrases in the feedback output were checked by hand against the original note and match word for word (curly quotes aside).
The short version
- "Improve this" got a full rewrite.
- Ask for feedback, and say not to rewrite.
- Have it quote the exact words.
- You make every change yourself.
