Use incidents, false alarms, and near misses as feedback. Review detection, alert value and volume, failure behavior, tests, spec clarity, agent context, and architecture. Traceably turn confirmed learning into specs, tests, monitoring, and working agreements for the next Build Loop.

Support the Learning Loop for [Project]. Refer to [Spec and version] and approved team observations.

Example observation: several reviews reported stale totals after quantity changes. Supporting evidence must be included.

Separate observations from hypotheses about causes. Add plausible alternative explanations. Do not evaluate individuals.

Design exactly one bounded experiment: cover every state transition with a before-and-after test case on suitable upcoming changes.

Specify expected value, human owner, budget, AI skill version, context selection, comparison cases, and review date. Test AI skill changes and context size separately. Compare quality, rework, review time, and AI costs. Evaluate an ongoing experiment before starting another.

After a human decision, turn proven learning into a regression test or versioned working agreement. Set another review date.

Confidential retrospective contributions stay within the agreed group. Do not turn local success into an organisation-wide requirement.

Add task boundaries before use:
Human owner: [Name]
Approved sources/files/tools: [List]
Selected AI skills and versions: [List]
Context: [Relevant files/examples; size; fetching and trimming]
Investment boundary and review point: [Team effort, AI costs, date]
Time and cost limit: [Boundary]
Test command or evidence: [Specific]
Escalation for open rules or limits: [Contact]