AI Self-Learning Meta Skill

Problem: debugging knowledge is hard to retain

In real engineering work, many working paths come from repeated trial and error: finding the right environment variables, respecting an implicit deployment order, and knowing which failing command should not be retried. Once the session ends, that context is easy to lose, so the next session may search again, guess again, and hit the same traps.

How the skill works

This skill converts recently completed work into a structured SKILL.md rather than saving one-off answers. It first checks whether the material qualifies: the path has been validated, its failure modes can be named, and at least one dead end was tried and ruled out. It then chooses project-level or global scope, de-duplicates existing skills, and extracts the golden path, including exact commands, file paths, environment variable names, required order, and approaches to avoid. Finally, it writes the skill to the target directory and reports a one-line summary to the user.

Boundaries

It is intended for multi-step, reusable workflows that are likely to recur, not single facts or one-time fixes. Project-specific paths should default to project scope; global scope is better for cross-repository habits. Secret values must not be written into skill files; only their locations should be recorded. Names must follow lowercase letters, digits, and hyphens, and the main file should remain concise.