skill-architect

Invoke the build phase with a request like "create a skill". The workflow classifies your skill into one of six archetypes, generates taxonomy-compliant naming and frontmatter, then scaffolds the directory using scripts/scaffold.py with starter references.

Ask to review or audit the skill. The process runs scripts/validate.py first, then applies a checklist covering frontmatter fields, body structure, description quality, and trigger collisions, classifying each finding as an error, warning, or suggestion.

A reliable description is 150-400 characters, opens with a differentiator from the nearest neighbor skill, and includes 3-6 quoted user-language phrases plus a WHEN clause. Triggers should be 6-10 verb-object phrases with no duplicates of description text.

Yes. The metadata phase treats description, name, license, compatibility, and allowed-tools as protected fields. It only adds or refines metadata fields like domain, triggers, role, scope, and output-format, each grounded in actual skill content.

Wrong activation usually comes from description collisions with neighbor skills or overly broad triggers. The review phase runs a collision scan against the 2-3 nearest neighbor skills and resolves overlap with leading qualifiers, anti-triggers, or priority settings.