safe-learning-governance
Use when a user asks an AI agent to learn, improve continuously, capture repeated mistakes, update rubrics, promote lessons, review stale rules, or build a self-improvement loop without broad memory r
Files1
View on GitHubSKILL.mdskills/safe-learning-governance/SKILL.md
GitHubSafe Learning Governance
Improve from evidence without turning every observation into a permanent rule.
Evidence Tiers
observation: one event or screenshot.pattern: repeated evidence across runs or tasks.lesson: validated pattern with scope and counterexamples.stable rule: only after repeated success and explicit promotion.
Learning Steps
- Record the evidence.
- Identify the failure or improvement pattern.
- Check for counterexamples.
- Define scope.
- Propose a lesson.
- Verify on a future task before promoting.
Stop Rules
Stop before broad memory rewrite, deleting old lessons, overriding project truth, or changing runtime behavior from a lesson alone.
Output
Return:
- evidence;
- pattern;
- proposed lesson;
- confidence;
- scope limit;
- next verification.