Skip to content

Vol. V · The Harness · V.2

Skills and Progressive Disclosure

A skill is a folder with a SKILL.md. Metadata always sits in context; the body loads on trigger; scripts and references load on demand. This is how you scale expertise without flooding the window.

8 min read

Doctrine

  • name + description must say what it does and when to use it. Under-triggering is the default failure.
  • Keep SKILL.md under ~500 lines. Push depth to references/ and deterministic work to scripts/.
  • Write in imperative form. Explain why. Do not staple ALWAYS/NEVER to every line.

The three layers

Anthropic's skill-creator and public skills repo define the loading system. Layer 1: frontmatter in the system prompt (~100 words each). Layer 2: body when the description matches. Layer 3: bundled files when the body points at them. Addy Osmani's agent-skills and the context-engineering pack show what 'production-grade' content looks like at layers 2–3.

Skills as the encyclopedia format

This encyclopedia is itself a skill body: the atlas is layer 1 (what exists), the article is layer 2 (when to apply it), the repo READMEs are layer 3 (load on demand). Agents should not ingest forty READMEs up front. They should load Canon, then open one volume.

Anti-patterns

  • A 4,000-line SKILL.md that is always loaded.
  • Descriptions that only name the skill ('use for frontend') without trigger contexts.
  • Skills that duplicate each other so the model picks none.

Related