skill-creator

To create and benchmark a Claude Code skill, you define the intended workflow, draft SKILL.md content with YAML frontmatter, generate eval prompts, and run trigger or evaluation comparisons against baselines to produce benchmark-driven improvements.

Trigger optimization for Claude skills improves the rate at which Claude invokes the correct skill by using a trigger evaluation set and iterative rewriting, needed when your skill fails to activate reliably in real user scenarios.

You iterate Claude skill performance using scripted workflows by repeatedly rewriting SKILL.md based on quantitative pass or fail assertions and qualitative feedback loops until the skill stably achieves the desired benchmark results.

Yes, building skills requires YAML frontmatter with name and description fields in SKILL.md, and the environment needs the pyyaml Python dependency to parse configurations and run scripted evaluation workflows.

Benchmark-driven iteration for skill creation relies on deterministic testing with predefined assertions, meaning it cannot assess subjective quality outside the defined evaluation set and requires manual viewer review for qualitative blind A/B analysis.