Learning Loop

---
name: learning-loop
listing_tier: name-led
description: Turn failures and user corrections into tested workflow improvements. Use for lessons and retrospectives.
---
# Learning Loop

Improve a decision process after a material failure, user correction or contradictory result. Use across coding, research, design and operations; an ordinary successful task or a spelling fix does not require a retrospective. The outcome can be a narrower rule, a test, a corrected measurement, or retirement of unhelpful guidance.

## Learn from evidence

Read [the correction procedure](references/correction-loop.md) for a substantive correction. Preserve what happened and the exact artifact versions. Separate the observed discrepancy from its proposed cause. Reproduce or qualify that cause, find the smallest useful correction, and check both the failure and a legitimate case the correction must still allow. A tabletop review is not a measured improvement in real performance.

Put the correction where it belongs: code, data, measurement, a project decision, a skill or memory. Search existing guidance first. Keep one canonical explanation and update its retrieval descriptions and pointers; keywords appearing in a store do not prove the correction changed behavior. Preserve superseded evidence privately instead of erasing the history.

Apply supported changes within the current task's authorization. The request to improve a workflow can already authorize its local edits; do not repeatedly ask for that same permission. This skill does not authorize unrelated memory edits, changing an owner's policy, publishing future releases, or scheduled background work.

Distinguish a user's policy choice from a claimed empirical result. A user can choose a conservative restriction despite limited evidence; record its scope as a decision, without pretending one incident proves a universal mechanism. Explain the tradeoff if useful, but do not use this skill as a veto over an explicit, permissible instruction. An authorized publication still needs appropriate handling of private material; evidence uncertainty alone does not cancel publication authority.

## Close the event

Record the correction as incorporated and checked, rejected with evidence, or pending a specific missing input. Preserve mixed outcomes. Revisit when later evidence challenges the rule; narrow, replace or remove it if its cost exceeds its demonstrated benefit. Do not recursively run self-improvement because this skill describes it, or require a review panel for every change.

For an authorized memory audit, read [memory-hygiene integration](references/memory-hygiene.md) and the installed audit skill's own instructions. The integration is optional. For multiple research lanes, use [Research Garden](../research-lane-coordinator/SKILL.md); for shared worktrees and PRs, use the applicable traffic-control procedure. Do not load these other workflows just because this skill mentions them.