← Back to cards
Reference Card

The Skill Spectrum

Same mechanism, wildly different scale.

Context Ladder · rung 3 — Skill
Step 1 · Floor

Well-described markdown

A trigger and a body. That's it.

Anatomy
  • name — kebab-case, matches the folder
  • Tight description — what it does + when to use it
  • Markdown body with the instructions
  • No other frontmatter needed

Example — a tight convention skill that triggers on a phrase and outputs a checklist.

Step 2 · Mid

Procedural + bundled refs

Structured steps with a small kit alongside.

Adds
  • Trigger / Procedure / Evidence / Exit structure
  • Bundled reference files in the skill directory
  • Anchored Interview — questions anchored to the corpus

Example — · scans the repo, asks anchored questions, writes a structured spec.

Step 3 · Ceiling

Orchestrated

Scripts, scoped tools, fork into a subagent, override the model.

Adds
  • Bundled scripts & !`cmd` injection
  • Scoped allowed-tools
  • context: fork — whole skill in a subagent
  • model / effort overrides

Example — a deep-research skill that forks a subagent, runs scripts, and returns only the result to the main thread.

Frontmatter capabilities — at a glance

Field What it does When to reach for it
allowed-tools: Read Grep Pre-approves these tools while the skill is active — no per-use permission prompt Skills you want running without interrupts
model · effort Override the model and/or effort level for this skill's turn (otherwise inherits the session) Reasoning-heavy work (spec, design review)
context: fork + agent: Explore Run the whole skill in a forked subagent of the named type; main thread only sees the return Long exploratory skills (deep-research)
paths: ["**/*.test.ts"] Glob patterns that limit auto-activation to matching files Path-scoped skills — though a folder CLAUDE.md often fits better
hooks: Hooks scoped to this skill's lifecycle Composing skills into pipelines
disable-model-invocation: true Prevents Claude from auto-invoking — user-only via /skill-name Workflows with side effects (deploy, commit, send-message)
user-invocable: false Hides from the / menu — Claude-only Background-knowledge skills users shouldn't run directly