skill-creator

Define the task, target user, tools, inputs, and outputs first, then choose an architecture pattern (Linear, Router, Methodology, Widget, or API Wrapper). Write YAML frontmatter with name and description, add numbered workflow steps, and defer deep content to reference files.

Score the skill against a 10-dimension rubric covering trigger quality, defaults coverage, instruction design, reference strategy, dynamic calling, output contract, error handling, code quality, conciseness, and domain accuracy. Scores of 70+ indicate production quality.

A good description names what the skill does, the categories of requests it serves, and the distinctive vocabulary users will use, all under 1024 characters. Avoid listing near-synonymous phrasings and point to sibling skills for neighboring requests.

Evaluation and planning work on Claude.ai, but creating skill files requires a CLI-based agent such as Claude Code. The skill detects available tools at runtime and adapts with fallback paths.

Keep SKILL.md under roughly 250 lines with workflow, decision tables, defaults, and the output contract. Move detailed API documentation, code templates over 20 lines, formulas, troubleshooting, and dated datasets into reference files.

Over-triggering usually comes from a generic description or capitalized MUST/NEVER warnings that current models over-apply. State each constraint once in plain words with its reason, and rewrite the description around intent categories.