| name | handoff |
| description | Summarize the current conversation into a compact handoff document for another agent or a future session. Use when the user says "handoff", "summarize for another agent", "write a handoff doc", "pass this to another agent", or when a conversation has reached a natural break point and needs to be resumed later. |
The user wants to pass the current conversation's context to another agent (or future self). Write a handoff document that compresses the entire conversation into something actionable.
1. **Summary first** โ start with a 2-3 sentence summary of what was discussed and decided.
2. **Current state** โ what exists now? What files were changed? What's deployed?
3. **Open questions** โ what's unresolved? What needs decisions?
4. **Next steps** โ what should the next agent do first?
5. **Landmines** โ anything the next agent should know to avoid wasting time (failed approaches, quirks, constraints).
6. **Suggest future skills** โ recommend which skills the next agent should invoke based on the remaining work.
7. **Don't duplicate artifacts** โ if you wrote a spec, don't repeat it. Reference the file path.
8. **Resume anchor (if handing off mid-task to learn something)** โ when the handoff exists so the user can go learn a concept and come back (e.g. paused mid-grill, see `grill-with-docs`), record the precise return point: the exact open question, the undecided trade-off, and what to re-read on return. This is what makes "return to the session clean" actually work.
Write the handoff as a markdown document. Use this structure:
# Handoff: [Topic]
## Summary
[2-3 sentences]
## Current State
- [what exists, what files changed, what's working]
## Open Questions
- [unresolved decisions or unknowns]
## Next Steps
1. [first thing to do]
2. [second thing to do]
## Landmines
- [things that didn't work, constraints, quirks]
## Suggested Skills
- [skill-name] โ [why]
## Resume Anchor (only if pausing mid-task to learn/branch off)
- Stopped at: [exact open question or step]
- Undecided: [the trade-off left open]
- On return, re-read: [file paths / learning-records to reload before resuming]
