
bye
by jimmy-academia
SKILL.md
name: bye description: Clean exit with session context capture. Writes detailed resume log to docs/sessions/, updates CLAUDE.md with pointer. Use when ending a session. user-invocable: true allowed-tools: Read, Write, Edit, Bash, Grep, Glob, TodoWrite, Task, AskUserQuestion
/bye - Clean Exit with Session Persistence
Ensures session context is captured for seamless resume in next session.
Core Principle
The next session's Claude has NO memory of this session. Your job is to create a "briefing document" that lets future-Claude immediately understand and continue the work.
Exit Protocol
1. Auto-Summarize Session Context
This is the most critical step. Analyze the conversation and extract:
- Topic: What was the main focus? (1-2 sentences)
- Decisions Made: Any conclusions reached
- Current State: Where did we leave off?
- Next Steps: What should happen next session?
- Open Questions: Unresolved issues or choices
- Key Files: Files that were discussed or modified
Create a session log at docs/sessions/session_YYYY-MM-DD_<topic-slug>.md:
Use compact format (max 20 lines):
# [YYYY-MM-DD] <Topic> <✅|🔄|⛔>
**What**: One-sentence summary of what was worked on.
**Impact**: Key outcome or result (optional, if significant).
**Files**:
- `path/to/file.py` - Brief note
- `path/to/other.py` - Brief note
**Next**:
- Next step 1
- Next step 2
**Status**: <Uncommitted: X files | Blocked: reason | Ready: next action>
**Context**: Any critical background (1-2 sentences max, optional)
Format guidelines:
- Use emojis: ✅ (completed), 🔄 (in-progress), ⛔ (blocked)
- Keep total length under 20 lines
- Be specific but concise
- Focus on actionable information
2. Session Log Location
Session logs are stored in docs/sessions/. Do NOT update CLAUDE.md with a pointer - the user runs multiple concurrent sessions, so a single pointer would cause conflicts.
At session startup, Claude automatically reads the latest session logs from docs/sessions/ to understand recent context.
3. Git Status Check
git status --short
Display if there are uncommitted changes. Do NOT offer to commit - just inform.
4. Todo Check & Clear
Check current todo list:
- List any incomplete todos
- Capture these in the session log's "Next Steps"
- Clear the todo list after archiving:
TodoWrite(todos=[]) - This ensures fresh start for next session (todos persisted in session log only)
5. Background Process Check
ps aux | grep -E '(python|node|npm|pytest)' | grep -v grep | head -5
Alert if processes are running.
6. Documentation Reminder
If significant changes were made this session, remind about doc maintenance:
📝 You made changes this session. Consider:
/sync - Check doc consistency (.claude/ ↔ docs/)
/audit - Verify docs match code (deep check)
7. Final Message with Two Paths
✓ Session context saved to: docs/sessions/session_YYYY-MM-DD_<topic>.md
─────────────────────────────────────────
What's next?
Continue working? → /compact (clears context, keeps session)
Done for now? → /exit (ends session)
─────────────────────────────────────────
📊 Project progress: /roadmap
📋 Session history: /hello
Two Exit Paths
/bye
│
├─► Save session context
│
└─► Two options:
│
├─► /compact
│ └─► Clear context, stay in session
│ (good for switching to unrelated task)
│
└─► /exit
└─► End session entirely
Important Guidelines
Session Summarization
DO:
- Be specific about technical decisions (e.g., "Decided to use Option A: single batch mode")
- Include code snippets or file paths if relevant
- Capture the "why" behind decisions
- Note any blockers or dependencies
DON'T:
- Be vague (e.g., "worked on some stuff")
- Skip technical details
- Assume future-Claude remembers anything
- Write a novel - be concise but complete
Log File Naming
Use descriptive topic slugs:
session_2025-01-18_gt-generation-batch-mode.mdsession_2025-01-18_fix-auth-bug.mdsession_2025-01-18_refactor-query-system.md
Multiple Sessions Same Day
If multiple sessions on same day, append time or number:
session_2025-01-18_gt-generation.mdsession_2025-01-18_gt-generation-2.md
Completed Work
If work is completed (not just paused):
- Set status to "completed"
- Remove "Active Work" section from CLAUDE.md
- Or update it to point to next priority
Example Session Log
# [2025-01-18] Ground Truth Generation Script 🔄
**What**: Designed batch mode approach for GT generation. Chose Option A (single OpenAI batch) over concurrent batches for simplicity.
**Files**:
- `src/addm/tasks/cli/extract.py` - Has `--mode batch` support
- `src/addm/tasks/cli/compute_gt.py` - Needs PolicyIR support (currently uses old formulas)
- `src/addm/llm_batch.py` - BatchClient infrastructure
**Next**:
- Decide: modify compute_gt.py or create new unified script
- Implement batch submission for judgment extraction
- Test with G1_allergy_V2 policy
**Status**: Uncommitted: 3 files
**Context**: Project now uses PolicyIR (V0-V3) instead of old formula modules (a/b/c/d). Batch API gives 50% cost discount, 24hr window.
Decision Tree
User invokes /bye
│
├─► 1. Analyze conversation
│ └─► Extract: topic, decisions, state, next steps, questions
│
├─► 2. Write session log
│ └─► docs/sessions/session_YYYY-MM-DD_<topic>.md
│
├─► 3. (No CLAUDE.md update - concurrent sessions)
│ └─► Session logs auto-discovered at startup
│
├─► 4. Check git status (inform only)
│
├─► 5. Check todos (captured in log)
│
├─► 6. Check processes
│
├─► 7. Show doc maintenance reminder (if changes made)
│
└─► 8. Display two options:
├─► /compact (continue working, fresh context)
└─► /exit (done for now)
Related
/hello- Session startup, reads session logs, produces overview/orient- General project orientation (reads CLAUDE.md)/doc- Quick documentation capture/roadmap- Project progress tracking/sync- Quick doc consistency check/audit- Deep doc verification
Score
Total Score
Based on repository quality metrics
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 100以上
3ヶ月以内に更新がある
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
Reviews
Reviews coming soon