Darwin Skill
Meta skill that optimizes your other skills: it takes an existing skill, runs it on cases, evaluates the result, proposes an improved version, tests it, then keeps the improvement or rolls back if it regresses. An evaluate → improve → test → keep-or-revert loop, inspired by Karpathy's autoresearch (an agent that self-improves in a loop). Its v2.0 (May 28, 2026) adopts SkillOpt's validation framework (Microsoft Research), and the SkillOpt repo lists Darwin among its integrations. The 'it optimizes your skills while you sleep' pitch no longer holds in v2.0: every phase stops at a checkpoint and waits for your approval before moving on.
Strengths
- Evolutionary loop with rollback: only versions proven better are kept
- Adopts the validation framework of SkillOpt (Microsoft Research), which lists Darwin among its integrations: a change only lands if it beats the previous version
- Automates a tedious tuning job nobody does by hand on their skills
- An explicit 100-point rubric across 9 dimensions: 59 for structure, read statically, 35 for effect, including 23 for results measured by running test prompts with and without the skill, and 6 for the list of things not to do. Independent sub-agents score every new version, so the model that edits is not the one that judges. A skill that shines in review but fails in use does not get through
Limitations
- Without a representative evaluation set, it optimizes for the wrong metric: the test harness is your responsibility
- Recent and niche: little production feedback, be cautious before automating it on critical skills
Best for
- Skill builders who already have an internal library and want to harden it without manual tuning
- AI platform teams industrializing the maintenance of a shared skill catalog
- Tech leads curious about the self-improving agent pattern wanting a concrete case to experiment with