Skill: session-handoff

When to Use

Use this skill when:

• Pausing mid-task and planning to resume later
• Handing off to a different AI session or developer
• A task is too large to finish in one session
• You want a clean resume point before a context window reset

This is more detailed than wrap-up — it focuses on producing a handoff artifact
that another session can use to continue without asking any questions.

---

Instructions

Step 1 — Capture current state

Gather the raw facts:

git status                          # what's staged / unstaged
git log --oneline -5                # recent commits
git stash list                      # any stashes
cat .claude/state/agent-state.json  # machine state
tail -30 .claude/state/checkpoint.jsonl  # recent steps

Step 2 — Write the handoff document

Update .claude/state/NEXT_ACTION.md with the full handoff:

# Session Handoff

**Date:** YYYY-MM-DD HH:MM
**Branch:** <current branch>
**Objective:** <one sentence — what is being built?>

---

## What Was Accomplished This Session

- <completed item 1>
- <completed item 2>
- <completed item 3>

## Current State

**Working tree:** <clean | N files modified | N staged>
**Tests:** <passing | failing — see blockers>
**Last commit:** `<hash> <message>`

## Exactly Where We Left Off

<Describe the precise stopping point. Include file name and line number if mid-edit.>

## Next Actions (in order)

1. **First:** `<exact command or action>`
2. **Then:** `<next action>`
3. **Finally:** `<last action to complete the task>`

## Blockers

- <blocker 1 — what is it and why is it blocking?>
- None (if clear)

## Key Decisions Made

- **Decision:** <what was decided>  **Reason:** <why>
- <repeat for each significant decision>

## Files Changed This Session

| File | Change |
|------|--------|
| `path/to/file.py` | Added X, modified Y |

## Do Not Touch

- `<file>` — in progress, half-written
- `<module>` — requires risky-module-review before any changes

## Context for Next Session

<2–3 sentences of context that would take 10 minutes to re-derive.
Include any gotchas, unexpected behaviour, or things that will matter.>

Step 3 — Update machine-readable state

# Update agent-state.json

Set in .claude/state/agent-state.json:

• status: "paused"
• next_step: next action (from Step 2 above)
• last_updated: current ISO8601 timestamp
• pending_risks: any known risks to carry forward

Step 4 — Confirm the handoff is self-contained

Ask: "Could another session read only NEXT_ACTION.md and continue without asking questions?"

If no: add the missing context.

---

Acceptance Checks

• NEXT_ACTION.md written with all sections complete
• Next actions are numbered and specific (not vague)
• Blockers explicitly listed (or "None")
• Files changed this session listed
• agent-state.json updated with status: "paused" and next_step
• Handoff is self-contained — next session needs no archaeology