Learning Loop · LL

The goal is better work next time, with fewer interventions from the user: a subsequent agent finds and correctly applies the decision without the user repeating it. Assess correctness and repeated intervention together in later comparable work; completing a retrospective or writing guidance does not establish this outcome.

What did we learn? Why was it not already carried forward? What smallest change would carry it into the next relevant task?

Run this global procedure in the current work's scope. The working project owns its decisions; the skill's global availability grants no authority to edit other repositories or global policy. Follow the user's existing authorization: assessment requests stay read-only; requests to apply improvements proceed through the authorized repairs.

1. Bound the evidence

Resolve the working owner and its instructions. Use the requested date, commit range or task; state the boundary when inferred. Read the relevant worklog entries, actual diffs, source evidence and current code, standards and checks. Distinguish committed, staged and unstaged work. Reconcile superseded steers against explicit acceptance and current implementation.

Finish with a bounded source set and allowed destinations. Preserve unrelated work.

2. Find the learning and the gap

Look for successful decisions as well as repeated or costly corrections. For each useful finding, identify the evidence, the decision worth retaining, and what the next agent should do differently. Check whether the current owner already carries it.

Classify the carry-forward gap before choosing a repair: