Skill Writer

Create and optimize Claude Skills following best practices and validation rules.

Core Principles

1. Context is scarce — Claude is already smart. Only add what it doesn't know. Challenge each paragraph: "Does this justify its token cost?"
2. Description is the trigger — Description determines WHEN a skill loads. It MUST describe triggering conditions, NEVER summarize workflow.
3. Progressive disclosure — Three-level loading: metadata (~100 tokens, always loaded) → SKILL.md body (<5000 tokens, on trigger) → references/scripts/assets (on demand, unlimited).
4. English-first — Write descriptions in English for reliable LLM matching. Body can be bilingual.
5. Directive-style — Use MUST, NEVER, CRITICAL RULE for hard constraints (not just "should" or "try to").
6. No duplication — Information lives in EITHER SKILL.md or references, never both. Keep core workflow in SKILL.md; move detailed reference material out.

Skill Types