correction-learning-loop

A skill that auto-fires the moment I correct a mistake or flag a missed expectation. It logs the correction in the session's mistakes table — what happened, what type, what rule would have prevented it — before work continues, so lessons survive time pressure.

Corrections evaporate if they only live in the conversation where they happened. Logging each one with its preventing rule, at the moment it happens, is how the same feedback stops being needed twice.

Non-trivial or recurring corrections escalate: a persistent memory file, and for patterns, a candidate rule in the promotion queue. The loop is why the standing instruction file keeps growing more specific instead of the same mistakes cycling.

The whole method is downloadable and adaptable — skills/correction-learning-loop/SKILL.md inside the zip.