| name | subagent-dispatch |
| description | Construct a precise, structured delegation prompt before every delegated-agent or worker launch. Use whenever you are about to delegate work to a background agent, run a task in parallel, or spawn a worker. Injects scope constraints, a structured output contract, obstacle reporting, and allowed-action boundaries directly into the input prompt to compensate for generic or unconfigured delegated-worker descriptions. Activate on: 'delegate this', 'run this in parallel', 'use the research agent', 'background task', 'spawn a worker', 'launch a subagent', 'hand this off'. |
Subagent Dispatch
When: immediately before launching any delegated agent / background worker.
Constraint: you usually cannot edit the worker’s system prompt — put intelligence in the delegation input.
Decision gate
Does intermediate work matter to the main thread?
| Answer | Action |
|---|
| No — need result, not journey | Delegate |
| Yes — each step depends on prior discovery | Keep in main thread |
Delegate: exploration that would clutter context; independent review; persona/tone tasks; clean-room TDD implementation after tests written in main.
Do not delegate: tightly coupled multi-step diagnosis; cases where you need full raw tool output in main (except clean-room TDD loops).
Four pillars (all inside the prompt)
- Precise scope — exact deliverable + explicit out-of-scope; ≤ ~500 lines pasted context.
- Structured output — numbered sections +
TASK_COMPLETE done signal.
- Obstacle reporting — force listing of workarounds / env quirks (or NONE).
- Allowed actions — READ / EDIT / RUN / FORBIDDEN lists (soft contract; pair with environment-level permissions when available).
Full pillar text, domain findings templates, and allowed-action examples:
references/pillars-and-templates.md
Copy-paste full prompt skeleton + anti-patterns + receive protocol:
references/prompt-template-and-anti-patterns.md
Minimal output contract (always include)
## 1. Objective Recap
## 2. Findings
## 3. Obstacles Encountered (or NONE)
## 4. Confidence & Caveats
## 5. Done Signal
TASK_COMPLETE
Preflight (before launch)
After return: scan for TASK_COMPLETE, read obstacles, treat missing sections as incomplete.
