Overview
How might we design repair contract so people can trust and act on AI output?
When to use
- Important for high-stakes chat workflows where users iterate on drafts, analysis, or decisions and need predictable, explainable retries instead of random variation.
When to skip
- First-turn prompts where there is nothing to repair yet.
- Creative brainstorming where tight constraints would kill useful variance.
- Fully autonomous agents that already self-correct with clear receipts.
Rules
Retry buttons that resubmit the identical prompt with no delta.
Repair UIs that hide the constraints the model will follow.
Forcing users to retype the entire original request to fix one line.
Accepting a repair without showing what will stay unchanged.
Evidence
| Product | Implementation |
|---|---|
| Cursor | Follow-ups that cite the prior diff and ask for a targeted fix. |
| ChatGPT | Rewrite and continue prompts that scope the next answer. |
| Claude | Retry with user-stated constraints in the same thread. |
| GitHub Copilot | Inline fix prompts that reference the rejected suggestion. |