Skill Creator

Problem

Building an agent skill is not only about writing a SKILL.md. The hard parts include unclear trigger conditions, bloated body text, test prompts that do not sound like real users, and a lack of evidence that the new version beats the baseline.

How It Works

ac runs a draft-test-evaluate-rewrite loop. It first clarifies what the skill should do, when it should trigger, what output format is expected, and whether objective tests are useful. It then writes SKILL.md, with a trigger-focused description and a body that stays progressive rather than dumping every detail at once. For verifiable tasks, it creates test prompts in evals/evals.json, launches with-skill and baseline subagents in the same turn, drafts quantitative assertions while runs are in progress, captures timing and token data, grades outputs, aggregates benchmark stats, and opens a viewer that shows both qualitative results and quantitative metrics. It also revises the skill description to reduce undertriggering.

Boundaries

It fits workflows with stable inputs and checkable outputs: file transformation, data extraction, code generation, and fixed procedures. Subjective tasks such as writing style or visual design can be tested, but forcing brittle assertions on them is unhelpful. The workflow assumes subagents, scripts, and an evaluation viewer; in headless or display-less environments, use the static HTML output path.