You ask AI to make an event invitation email “more polished.” The second version changes the headline, the third drops the registration details, and the fourth turns the opening you liked back into press-release language. Every version looks like work was done, yet no one can say which one is closer to finished.
This back-and-forth usually has nothing to do with how clever the prompt is. The cause is feedback that shows no clear difference and doesn’t say what must stay untouched. “More natural,” “more professional,” and “warmer” each have ten interpretations, so AI changes tone, length, structure, and wording all at once, and the next round loses its point of comparison. To make revisions converge, give only three things each round: what to keep, one visible change, and what not to touch. Then have AI deliver two versions to compare.
More adjectives, more things AI changes
When AI hasn’t got it right, it’s natural to add more adjectives: “a bit shorter, a bit warmer, but professional.” The trouble is that more adjectives mean more directions to interpret, and AI changes more things each round.
OpenAI’s guide to evaluating AI output makes two points. First, models sometimes produce different output from the same input, so you can’t judge from a single result. Second, models are better at distinguishing between options, so evaluations should use pairwise comparisons, classification, or scoring against specific criteria rather than open-ended questions about whether something is good. Revising works the same way. Instead of asking AI “is this polished?”, have it produce two versions and point out which is closer and what the difference is.
A feedback card with three entries
Keep: what’s already right. For example, the date, location, registration link, and opening sentence stay.
Change: only one difference you can see this round. For example, “cut the middle section from 90 words to 45, and replace three slogans with one situation an attendee will actually face.”
Boundaries: what must not be sacrificed for the change. For example, no new promises about the event, no change to the fee, and no unconfirmed speaker credentials presented as fact.
These three entries add one key thing that “make it a bit shorter” lacks: AI knows when the change is done, and which edits would actually be a step backward. Anthropic’s documentation on evaluation makes a similar point: good success criteria are specific and measurable, so instead of “good performance,” spell out what should be achieved.
Explain your choice of version the same way. Don’t say “B feels warmer.” Say: “Choose B, because its first two sentences say directly what attendees will take away, and it stays within 45 words; keep A’s date sentence. Both versions present the free consultation as a guarantee, so don’t use that line.” Now the next round has a condition it can act on, and the reasons for rejection are on record.
Classify disagreements instead of asking AI to split the difference
When two colleagues disagree about the same sentence, a common move is to ask AI to “find a middle ground.” The result usually satisfies neither. Google’s People + AI Research team, which studies AI interaction design, cautions in its guidebook that disagreements between labelers shouldn’t simply be discarded as noise; they may reveal problems with instructions, workflows, or the data itself.
So first ask what kind of disagreement it is. If one person cares about brand voice and the other about regulatory disclosure, this isn’t a matter of word choice but two different responsibilities. Make the regulatory content a boundary that can’t be touched, and let the brand owner handle tone. If it’s only a matter of how enthusiastic to be, have each person pick one acceptable and one unacceptable example sentence and spell out the difference.
Knowing when to stop
Changing one aspect at a time doesn’t mean other qualities don’t matter; it avoids changing five things at once and not knowing where an improvement came from. If two conditions conflict, such as adding full background while also halving the word count, have a person set the priority first rather than letting AI guess.
Also know when to stop. If this round’s condition is met and nothing on the keep list was broken, stop there. Don’t treat “there’s a new version” as progress just because AI can produce a fifth one. When source facts conflict, when brand and regulatory requirements have no agreed priority, or when a change would create an external commitment such as a price, contract, or hiring promise, stop and hand it to the person responsible.
The next time AI revisions start going in circles, save the current version first, then rewrite your feedback into those three entries. When the finished piece goes to a colleague, include which version was chosen and why; see A Tidy AI Summary Does Not Mean a Teammate Can Take Over. If it’s code being revised, for who signs off after the first version, see AI’s First Version of a Code Change Is a Price Check; the Real Bill Is Review and Maintenance.
AI handoff card
I've asked AI to revise a document (or slides, or a design brief), and after several rounds it keeps going in circles. I want it to converge.
The method comes from this article: https://boosterminiclass.com/en/posts/ai-revision-feedback-compare/
Its view: revisions go in circles because the feedback shows no clear difference. Each round, give only three things: what to keep, one visible change, and what not to touch. Then have AI deliver two versions to compare, and keep every version.
Ask me 3 questions first:
1. Which file is the latest version and which is the previous one, and what am I least happy with right now?
2. Which content is already right and must not change (for example dates, prices, links, a particular sentence)?
3. Are there facts, commitments, or regulatory requirements that must not be sacrificed for the change?
Based on my answers, give me:
- A comparison of the latest and previous versions: intended changes, unresolved problems, and changes that shouldn't have happened, each with a quoted snippet;
- A feedback card: keep / this round's single change (specific enough to see the difference) / boundaries not to touch;
- Versions A and B that edit only the specified area, explaining how each meets the condition and what it may cost;
- A stop condition for this round.
Don't choose a version for me, don't merge, and don't widen the scope of edits; wait for my decision before continuing.
References
- OpenAI Developers: Evaluation best practices — https://developers.openai.com/api/docs/guides/evaluation-best-practices [accessed: 2026-10-11]
- Anthropic (Claude Platform Docs): Define success criteria and build evaluations — https://platform.claude.com/docs/en/test-and-evaluate/develop-tests [accessed: 2026-10-11]
- Google People + AI Research: People + AI Guidebook — Patterns — https://pair.withgoogle.com/guidebook-v2/patterns [accessed: 2026-10-11]
