skill-builder

Run init_skill.py with a kebab-case name and target path to scaffold the directory, then edit SKILL.md to complete the description and workflow sections. Finish by running quick_validate.py to confirm the skill passes all format rules.

The official spec requires name and description, while this environment also mandates version and updated dates. Optional fields include argument-hint, allowed-tools, context, model, effort, watches, and lifecycle tracking fields like last-synced-epic.

Run quick_validate.py against the skill directory to check over 20 rules covering frontmatter legality, kebab-case naming, description length and trigger quality, file references, and token budget levels. package_skill.py runs the same validation automatically before creating the .skill archive.

disable-model-invocation set to true means only the user can trigger the skill, giving zero context cost until called. user-invocable set to false hides the skill from the slash menu while still letting Claude load it automatically as background knowledge.

Use watches for skills tied to evolving source code, such as SaaS module skills, so staleness is detected when tracked files change. Each watches entry needs a non-empty glob pattern pointing at the monitored code paths.

Warnings fire when the description exceeds its token budget level, lacks a trigger phrase like Use when or Triggers in the first 150 characters, uses first or second person pronouns, or summarizes workflow steps. Rewrite it in third person focused on when the skill should activate.