Design, scaffold, and validate new AI agent skills. Use when the user wants to create a new capability, codify a workflow, or fix a broken skill. Triggers: "create a skill", "new skill", "write skill documentation", "define workflow".
Defines the strict architectural standards for creating high-fidelity Claude Code skills.
Before generating any skill content, you MUST read the following architecture definitions to ground your reasoning:
Read("reference/directory-structure.md")Read("reference/anti-patterns.md")Read("templates/skill-body.md")You must distinguish between Timeless Principles and Volatile Syntax.
reference/WebSearch if you are not 100% certain of the current version.
reference/ file.Default assumption: Claude is already very smart.
Before adding content, ask:
Prefer:
Avoid:
managing-d1-migrationsd1-managerDecide where the information lives based on the Timeless vs. Volatile rule:
SKILL.md: Navigation logic, decision trees, and triggers ONLY.reference/: Detailed documentation, examples, and timeless concepts.scripts/: Deterministic logic (Python/Bash) to avoid LLM hallucination in complex calculations.mkdir -p .claude/skills/[skill-name]/{reference,scripts,templates}.SKILL.md using the frontmatter from templates/skill-body.md.ALWAYS:
description that includes trigger phrases (WHAT + WHEN).reference/ or scripts/.Read("reference/...") commands in the SKILL.md to force context loading (Progressive Disclosure).NEVER:
SKILL.md files (> 150 lines is a warning, > 500 is illegal).