skill-creator-normal

Start by defining what the skill should do, when it should trigger, and its expected output format. The skill then guides you through drafting SKILL.md, planning the directory structure, writing test prompts, and running evaluations to verify it works.

Run each test prompt twice in parallel: once with the skill and once without it as a baseline. The aggregate_benchmark.py script then computes pass rates, execution time, and token usage with mean and standard deviation for both configurations.

Generate about 20 realistic trigger eval queries split into should-trigger and should-not-trigger sets, then run the description optimization loop. It tests description variants against train and held-out test queries and applies the best-performing version.

Beyond SKILL.md, use scripts/ for deterministic repetitive tasks, references/ for domain knowledge too large to inline, assets/ for output templates, and agents/ for subagent prompt files. SKILL.md should point to each bundled file with guidance on when to read it.

On Windows, running npx skills find directly can silently produce empty output due to an npx compatibility issue. Wrap the command in powershell -Command to get correct results before concluding no community skills exist.

Use blind comparison when you need a rigorous judgment of whether a new skill version is actually better than the old one. An independent agent compares both outputs without knowing which skill produced them, then an analyzer explains why the winner won.