The Coordinator Pattern

The Context Problem

• **The problem** — Multi-step pipelines (fetch, analyze, plan, implement) accumulate context. By step 4, the AI carries the weight of steps 1-3 in its context window. Quality degrades as context fills up.
• **The solution** — A 'coordinator' skill with `disable-model-invocation: true`. It doesn't reason — it just dispatches subagents. Each step gets a fresh context window. Step 3 doesn't carry the weight of step 1's output.
• **Why disable-model-invocation** — Without it, the coordinator would try to 'think' about the instructions, wasting tokens on reasoning that isn't needed. With it, the skill is a pure dispatcher.

Build a Coordinator

1. 1 **Real example** — `/dx-figma-all` orchestrates 3 sub-skills: extract Figma design, generate prototype, verify against reference. Each runs as a separate agent. Total pipeline: zero context bloat.
2. 2 Create a coordinator skill
   • › Set `disable-model-invocation: true` in frontmatter
   • › Each step invokes a sub-skill via `Skill("dx-sub-skill")`
   • › Sub-skills communicate through spec files, not shared context
3. 3 **When to use it** — If your multi-step skill's steps can run independently with just file I/O between them, make it a coordinator. Keep monolithic skills only when steps truly depend on shared in-memory context.
4. 4 **Test sub-skills first** — Each sub-skill must work standalone before wiring them into a coordinator. Independent testability is the whole point of the pattern.