Repair AI Output with a Repair Prompt
Don't re-roll the whole response — send back a surgical prompt that fixes the violations and keeps everything that was right.
Overview
When output breaks the rules, the reflex is to regenerate — and lose everything that was correct in the re-roll. The repair prompt is the better move: it restates the original requirement, lists each violation with its specific fix, and instructs the model to return only the corrected output, changing nothing else. This setup demonstrates it on a classification that invented a label ("Very Urgent") outside the defined set — one of the violations where a surgical fix beats a fresh attempt.
Workflow
-
Validate first
The repair prompt is generated from real findings — no findings, no repair, no busywork.
-
Send it in the same thread
The model repairs its own response best when that response is still in context.
-
Re-validate the repair
Paste the corrected output back in — PASS closes the loop; anything else gets a fresh repair prompt.
Why This Works
- "Fix only the problems listed" constrains the model exactly where regeneration doesn't
- Restating the original requirement re-anchors the contract
- The validate → repair → re-validate loop converges in one or two rounds
Best for
- Responses that are 90% right — most of them
- Long outputs where regeneration costs real tokens
- Conversations where the model's context is still warm
Not for
- Content that is wrong rather than misformatted — bad facts need a better prompt, not a repair
- Defining label sets up front — that's the Data Classification Prompt
Use cases
- Fixing one wrong label without re-rolling the analysis
- Correcting format violations in long responses cheaply
- Standardizing the repair flow across a team