Microsoft SkillOpt: Auto-Improving AI Agent Skills Docs Without Retraining (41% → 80%)

Exact instruction

1. 1Conceptual research paper — not directly implementable yet as a consumer tool. Key insight to apply now:
2. 2Treat your Claude system prompt / CLAUDE.md as a living skills doc, not a set-it-and-forget-it file.
3. 3After each session where Claude made mistakes, identify which instruction was missing or vague.
4. 4Make 1–4 targeted edits (add a rule, sharpen a vague instruction, remove a bad one).
5. 5Observe whether next session's output is better — keep changes that help, revert ones that don't.

Suggested prompt

Adopt?

Yes: The manual version of SkillOpt is exactly what good Claude Code practitioners already do — iterate on CLAUDE.md after each session. The paper validates this approach scientifically. The 1-4 targeted edits per session finding is directly actionable.

Find the resource

Search 'SkillOpt Microsoft paper' or 'SkillOpt agent skill document optimization' — the paper is publicly available on arXiv or Microsoft Research.