skill-maker

Create agent skills using an eval-driven lifecycle by drafting a SKILL.md, running repeatable eval loops with subagents, and refining behavior through automated grading signals. This structured workflow ensures production-grade quality from intent capture to final deployment.

An eval loop for refining AI agent skills is a repeatable workflow phase where subagents execute tasks, grade.ts evaluates outputs, and aggregate-benchmark.ts tracks performance. It provides deterministic grading signals to iteratively improve skill behavior until plateau detection.

Optimize agent skill trigger accuracy by refining the skill description during the final workflow phase. The skill-maker workflow analyzes eval loop results and benchmarking artifacts to adjust triggers, ensuring the skill activates correctly across intended prompts and tasks.

Benchmark agent skills for production readiness by running repeatable eval loops with bundled scripts like grade.ts, aggregate-benchmark.ts, and detect-plateau.ts. This generates validated timing and benchmark artifacts to verify deterministic behavior before distribution.

Yes, skill-maker works without external dependencies. It operates environment-agnostic using bundled scripts and references like validate-skill.ts, generating production-ready artifacts including SKILL.md, evals, and benchmarks for repeatable distribution across agent environments.

Stop refining an agent skill during an eval loop when detect-plateau.ts identifies performance stabilization. The workflow uses aggregate-benchmark.ts to track grading signals across iterations, indicating diminishing returns when benchmark metrics cease meaningful improvement.