Optimizes agent skills by running evals, mutating prompts, and keeping only validated gains.
npx skills add https://github.com/exiao/meta-skills --skill skill-improver-exiao
Please help me install this Agent Skill. Skill: skill-improver Source: https://github.com/exiao/meta-skills/tree/main/skill-improver Command: npx skills add https://github.com/exiao/meta-skills --skill skill-improver-exiao
High-intent search queries and answers about installing and using this skill.
Define 8-12 test inputs and 3-6 binary pass/fail eval criteria, then run the loop: establish a baseline score, propose structured edits from failure traces, and keep changes only when the PACE validation gate confirms a statistically real gain.
A good eval is a yes/no question two different scorers would answer identically, such as checking for banned phrases or required sections. Avoid 1-10 scales, which compound variance, and avoid checks so narrow the skill games them.
No, the default uses a different architecture for the optimizer than the target so it can catch systematic blind spots the target cannot see in itself. A same-model config is available for quick or cost-sensitive runs.
No, it never edits the original SKILL.md. All mutations happen on a working copy in an autoresearch directory, and the improved version is delivered separately for the user to adopt.
Validation scores can drift upward over many experiments because the gate sees them repeatedly. The held-out test set, scored once at the end, reveals the honest improvement; a large val-test gap indicates overfitting.
8-12 inputs support a clean 50/25/25 train/validation/test split, and the PACE gate needs at least 8 paired validation instances to auto-commit. With 4 or fewer inputs no split is possible and overfitting risk is high.
engineai-nzRun autonomous mutation experiments to improve skill prompt pass rates.
hiddink-aiMine recurring failure patterns from multi-session agent transcripts to generate skill candidates.
githubValidates coordinator and agent template changes through end-to-end squad sessions.
photonics-dhlBuild AI pipelines by declaratively programming LM workflows with typed signatures.
