Skill Creator

Problem

Building AI agent skills often stops at a draft: the SKILL.md exists, but there are no test prompts, no baseline, and no way to compare versions, so triggering is unstable and output quality is hard to judge. This skill turns skill development into a repeatable engineering loop: clarify intent, draft the skill, run tests against baselines, and iterate based on scored feedback.

How It Works and Where It Fits

The workflow centers on a closed loop around SKILL.md:
- Capture intent: define what the skill does, when it should trigger, and what output format is expected.
- Draft and research: fill in edge cases, dependencies, MCP tools, and examples before overfitting to prompt text.
- Run evaluations: launch with-skill and baseline subagents for each test prompt in the same iteration, saving structured metadata.
- Score quantitatively: draft verifiable assertions, generate benchmark data, and review pass rate, latency, and token usage.
- Iterate: revise the skill based on user feedback and benchmark failures, then expand the test set.

It is best suited to file transforms, data extraction, code generation, and fixed workflows where output can be checked. Subjective writing or design tasks may be better judged qualitatively. The skill also enforces safety boundaries and should not be used to create misleading, malicious, or unauthorized-access content.