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

SKILL.md

GitHub

Safe 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

  1. Record the evidence.
  2. Identify the failure or improvement pattern.
  3. Check for counterexamples.
  4. Define scope.
  5. Propose a lesson.
  6. 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.