Write the frontmatter for a Claude Code subagent at `.claude/agents/{{name}}.md`.

```
---
name: {{name}}
description: {{when_to_dispatch}}
tools: {{tools}}
model: {{model}}
---
```

Then a system prompt body that:
- States the agent's single responsibility: {{responsibility}}
- Tells it to return its final answer as raw data (the parent reads it, not a human).
- Lists hard constraints (read-only? no network? output schema?).

Keep the scope narrow — one agent, one job.