Subagent Prompting Guide
# Subagent Prompting Guide
Subagents differ from skills: they run in isolation, receive delegated work from an orchestrator, and return results. Their output is consumed by the orchestrator, not a human. This constraint shapes everything below.
The deletion test: imagine removing a line from your subagent prompt. If behavior doesn't change, the line isn't earning its place.
---
## Core — Apply to Every Subagent
### 1. Outcome-First with Phase Gates
Define what good looks like at each phase boundary. Trust the model to navigate within.
A subagent prompt should answer three questions:
- What does "done" look like? (outcome)
- What must be true before claiming done? (phase gates)
- What should never happen? (constraints)
Do not prescribe the path between gates unless sequence is the point. GPT-5.5 finds efficient solutions when given destination and constraints — process steps narrow its search space.when to use it
Community prompt sourced from the open-source GitHub repo rolandtolnay/llm-toolkit (MIT). A "Subagent Prompting Guide" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.
tags
productivitycommunitydeveloper
source
rolandtolnay/llm-toolkit · MIT
more in Productivity
Productivity✓ tested
Summarize a doc into decisions & actions
chief of staff who extracts what to DO, not just what was said
Productivity✓ tested
Draft a reply to a hard email
calm, direct communicator who de-escalates without caving
Productivity✓ tested
Turn a brain-dump into a weekly plan
planning coach who protects your focus, not just your calendar