Close out the current coding session in a way that is easy to resume later. Focus on what was actually finished, what is still open, what was learned, and what should happen next.
Language Detection
Detect the user's language from their request:
- Chinese or mixed Chinese input → output in Chinese (default)
- English input → output in English
- Keep identifiers, paths, branch names, and commit hashes in their original form
Core Rules
- Scope is the current session unless the user explicitly asks for a broader range.
- Prioritize outcomes, decisions, learnings, and follow-up actions over process chatter.
- Do not output a chronological running account.
- Do not claim work is completed unless there is evidence in the session, git status, or validation output.
- If the user asks to write files, commit, or update docs, summarize first and then confirm the action.
Workflow
Step 1: Confirm scope
Default to the current session. If the user already gave a clear wrap-up request, do not ask extra questions.
Clarify only when needed:
- Current session only or also include earlier sessions
- Whether to include commit advice
- Whether to produce a handoff note or just a chat summary
Large or multi-topic sessions:
- If the session covered 3+ unrelated topics, group by topic rather than by time. Ask the user which topics matter most if grouping is ambiguous.
- If the conversation context is very long (compacted multiple times), rely more on git evidence and less on conversation recall. State which portions are based on git vs memory.
Step 2: Inspect the working tree
Collect the minimum evidence needed for an accurate summary:
git status --short
git diff --stat
git log --oneline -n 10
If the repository state is not available, say so explicitly and continue with session context.
Step 3: Summarize completed work
Summarize by workstream instead of by timestamp.
Structure:
- Completed this session
- Key decisions
- Files and modules touched
- Verification performed
If there were no meaningful code changes, say that clearly instead of inventing results.
Step 4: Extract learnings and open items
Capture what matters for the next session. Focus on signals that would change how someone approaches the work if they resumed it cold.
Worth extracting:
- A debugging path that was non-obvious (e.g., "the real cause was X, not the apparent Y")
- A design decision and the reasoning behind it, especially if alternatives were considered
- An API/library behavior that surprised you or contradicted docs
- A workaround for a tool limitation or environment issue
- An unfinished task with enough context to resume without re-investigation
- A known risk that has no mitigation yet
Not worth extracting:
- Routine operations that went as expected
- Tool invocations or file reads that were just part of the process
- Decisions that are already obvious from the code diff
Structure:
- Learnings: 2-5 bullet points, each one sentence
- Open items: task + what's blocking or what's needed next
- Risks: risk + current mitigation status (none / partial / mitigated)
Keep this section concise and operational.
Step 5: Offer next actions and commit guidance
When relevant, finish with:
- Suggested next steps (prioritized)
- Whether a commit is appropriate now
- A possible commit message direction
- Whether additional verification is still needed
Do not auto-commit. Recommend commit timing only when there is enough evidence.
Step 6: Generate handoff note (optional)
If the user asks for a handoff note, or if the session involves substantial unfinished work, generate a structured handoff using the template in $SKILL_DIR/references/handoff-format.md.
The handoff note captures context that would be lost between sessions — decisions made, approaches tried and rejected, environment state, and clear next steps.
Step 7: Save / Export (optional)
Save policy:
- If the user explicitly asks to save, write to a file.
- If session persistence tools are available (OMC
notepad_write_working, Claude Code/export-summary), suggest using them. - Default filename:
YYYY-MM-DD-session-wrap.md - If the user does not ask to save, return the summary in chat only.
Output Template
## 本次会话总结
### 已完成
- ...
### 关键决策
- ...
### 涉及文件 / 模块
- ...
### 验证情况
- 已验证：...
- 未验证：...
### 经验与风险
- ...
### 下一步建议
- ...
When Not to Use
- User wants all of today's sessions summarized → use
project-daily-summary - User wants a pure commit-based daily report → use
commit-daily-summary - Current session has almost no real work content, just simple Q&A
Quality Checklist
Before responding, verify:
- Summary is scoped to the current session unless user asked otherwise
- Completed work is grouped by workstream, not by timeline
- Validation is reported honestly
- Risks and unfinished items are explicit
- No fake completion, fake commit readiness, or fake verification claims
- Output language matches the user's language