Self-Correction

Problem It Addresses

When users reply with no, not quite, wait, or you missed that, models often patch the previous answer, over-apologize, or keep expanding from a mistaken premise. Self-Correction treats these signals as intent drift: return to the original first-turn request instead of repairing the prior response.

How It Works

• Trigger detection: covers explicit negation, challenge, premise rejection, result rejection, rephrasing, and omission cues such as not this way, you misunderstood, or I mean....
• Backtrack to source: analyzes the user's initial prompt, constraints, stack, format, and implied context rather than the last AI turn.
• Rebuild the response: the new answer should be substantively different, may list likely interpretations for confirmation, and avoids asking the user to explain what went wrong.
• Keep overhead low: minimal apology, no long correction narrative, no repeated wrong assumptions.

Scope And Limits

Useful for multi-turn technical conversations where requirements, frameworks, output formats, or resource constraints are easy to misread. It does not replace tool execution, tests, or missing external context, and user additions should not be treated as negations unless they actually contradict the prior request.