Test binaire skill vs agent
Heuristic taught by Willy (Gouvernance IA) to decide whether a Claude Code artifact should be a skill or an agent: if it must arbitrate (decide, say no, escalate) → agent; if it only informs → skill. Completed by 3 agent criteria: written mandate, decisions alone within scope, clear escalation out of scope. The test avoids error #1 across nearly a million skills shipped: confusing the two and giving decision power to a procedure book.
Strengths
- One closed question to ask before every creation, no heavy framework, no matrix to fill
- The 3 agent criteria (mandate, scope, escalation) become an operational checklist to frame a sub-agent
- Compatible with the Anthropic orchestrator + sub-agents pattern (sub-agent applies skills, orchestrator arbitrates)
Limitations
- Author heuristic (Willy, Instagram), not an official Anthropic term, not in the agent-skills docs
- The 'arbitrate vs inform' line can blur: a skill returning a boolean decision may look like it arbitrates but remains a skill (no execution autonomy)
Best for
- Before creating a new SKILL.md or a new sub-agent, 30 seconds of qualification
- Auditing a .claude/ repo that mixes skills and agents with no clear logic
- Training a product team on Claude Code artifact governance in prod