skill-creator

Start by defining what the skill should do, when it should trigger, and its output format. Then write a SKILL.md with YAML frontmatter containing a name and description, followed by markdown instructions, and optionally add scripts, references, or assets directories.

Run each test prompt twice in parallel: once with the skill and once without as a baseline. Grade both outputs against objective assertions, then aggregate pass rates, timing, and token usage into a benchmark to compare configurations.

Create 20 realistic eval queries mixing should-trigger and should-not-trigger cases, then run the automated optimization loop. It evaluates candidate descriptions on a train split and selects the best by held-out test score to avoid overfitting.

A good description states both what the skill does and specific contexts when to use it, since the description is the primary triggering mechanism. Slightly assertive phrasing helps because models tend to undertrigger skills.

Yes, but with adaptations: run test cases sequentially yourself instead of in parallel, skip baseline runs and quantitative benchmarking, and present outputs inline for qualitative feedback instead of using the browser-based review viewer.

Bundle a script when test runs show subagents repeatedly writing the same helper code across cases, or when a task is deterministic and repetitive. Scripts save every future invocation from reinventing the same logic.