Evaluate and improve an existing Agent Skill with darwin-skill

Who it is for

This skill is for people who already have an Agent Skill and want to evaluate and improve it through an experiment loop. Officially: optimize your Agent Skills the way you train a model. It is inspired by Karpathy's autoresearch and moves an autonomous experiment loop from model training onto Skill optimization.

Inspiration:

https://github.com/karpathy/autoresearch

It does not teach you to write a first skill from scratch, and it is not generic prompt polishing. A better starting point is: you already have a runnable skill file, you want to see which dimension is weak, then decide whether to change it. If you have no skill to evaluate yet, the page does not give another full tutorial on "create a skill first".

Official notes:

https://github.com/alchaincyf/darwin-skill

First try in 3 minutes

The official install command is only this one:

npx skills add alchaincyf/darwin-skill

This writes the skill into the current assistant environment. Confirm you want it installed before running. Do not add a second install command.

After install, do a low-risk action: pick an existing skill, have the assistant evaluate it with the Darwin flow first, do not change files directly, and do not run deletes or hard resets during evaluation. The page says the full loop includes evaluate, improve, test, and keep or roll back. The first time, stop at the evaluation result and see clearly whether the next stage will rewrite the skill.

If the assistant enters Phase 2 single-dimension optimization, it must stop at CHECKPOINT and wait for your confirmation. Success signal: you see the evaluation notes, and any rewrite happens only after you agree. If it keeps changing files without a checkpoint, stop immediately.

Usage examples

Example 1: Phase 2 single-dimension optimization with confirmation at the checkpoint. The evidence says this stage forces a pause and waits for user confirmation. Give one skill to the assistant, allow only a single-dimension change proposal, read the proposed change at CHECKPOINT, then decide whether to continue. Do not let it keep a new version automatically before confirmation.

Example 2: check the high-risk action blocklist. The evidence says the ninth dimension requires the skill to list forbidden destructive operations explicitly, for example rm, git reset --hard, force push. The first time, only check whether the target skill wrote this blocklist; do not actually run those commands. If the list is missing, stop at "point out the gap". The page does not write a command that auto-patches the blocklist.

Limits

The optimization loop will change your skill files and includes keep or roll back. The first time, treat CHECKPOINT as a hard stop, not a skippable hint. High-risk commands are covered by this standard only when the skill under evaluation lists them as forbidden; the evidence does not say Darwin will intercept every rm for you.

It handles Skill optimization, not model training itself. autoresearch is only the inspiration. The materials do not write a specific slash command that starts evaluation, and they do not write a test-set format. Stop missing steps at the existing notes; do not install another optimizer.