conversation-distill
At the natural end of a meaningful conversation, show a one-line soft reminder asking the user if they want to distill — do NOT auto-start. Only ask when the conversation has distillation value (decisions, insights, lessons, open questions, action items). If user says yes, run the full 5-step classify→confirm→write flow. Trigger reminder when: (1) user says closing phrases like 'that's all', 'thanks', 'done', '好的就这样', '没了', '谢谢'; (2) 3+ turns with no new topics AND conversation had substantive content. Do NOT remind for: casual chat, pure Q&A, pure coding/debugging with no decisions, when user already wrote to KnowMine this session.
Conversation Distill (KnowMine Edition)
The biggest waste of a conversation isn't that nothing was saved — it's that valuable insights are buried in the process and never revisited.
This skill closes every meaningful conversation with one explicit action: classify → confirm → write to KnowMine.
When to Use
The core problem: real-time capture ≠ session-level distillation.
Real-time capture handles individual highlights. This skill is the closing ritual — a full scan of the session to see what was produced, map relationships, and catch what slipped through.
Trigger when:
- User says a closing phrase: "that's all", "got it", "thanks", "done for now", "wrap up", "收尾", "好的就这样"
- 3+ consecutive turns with no new topics
- User switches to an unrelated topic; previous topic had substantive output not yet saved
- User explicitly invokes: "distill", "save this session", "wrap up", "沉淀一下",
/knowmine:conversation-distill
Do NOT trigger for:
- Quick single-turn queries
- Casual conversation or emotional support
- Pure coding/debugging/execution tasks with no knowledge output
- When user is already actively calling KnowMine write tools this session
- When user says "don't save" or "skip it"
Soft Ask Pattern（先问再做）
检测到结束信号后，先问一句，不要直接启动 5 步流程。
触发判断（两个条件同时满足才提醒）
条件 A — 结束信号（满足其一）：
- 用户说：「谢谢」「好的就这样」「搞定了」「完成了」「没了」「就这些」「thanks」「done」「that's all」「good」
- 连续 3 轮以上无新话题，对话自然收尾
条件 B — 对话有实质内容（满足其一）：
- 出现了决策或方案选择
- 讨论了架构、设计、策略
- 产生了值得复用的经验教训或踩坑记录
- 有未记录的 TODO 或开放问题
两个条件缺一不可。纯闲聊说「谢谢」不提醒；有实质内容但对话还在进行中不提醒。
提醒前：做一次轻量预扫描
出提醒之前，先扫描对话，找出「最有代表性的一条」作为预告。预告要具体——不是「有内容」，而是「有决策 / 有踩坑 / 有 TODO」加一句简短描述。
提醒话术（带具体预告）
中文对话：
💾 这次对话有 {N} 条值得沉淀的内容（比如{最代表性的一条，10字内}...）。要收尾整理吗？（说「要」开始，「不用」跳过）
英文对话：
💾 Found {N} things worth saving from this conversation ({one-line preview, e.g. "the decision about X"}...). Quick distill? (say "yes" to start, "skip" to pass)
示例：
- ✅
💾 这次对话有 3 条值得沉淀的内容（比如你那个关于 MCP 权限分级的设计决策...）。要收尾整理吗？ - ❌
💾 这次对话有些值得沉淀的内容，要收尾整理一下吗？（太泛，不体现理解）
用户响应规则
| 用户回复 | Claude 行为 |
|---|---|
| 「要」「好」「是」「收尾」「沉淀」「distill」「yes」「go」「start」 | 立即启动完整 5 步流程 |
| 「不用」「算了」「跳过」「skip」「no」「nope」「pass」 | 一句「好的，跳过。」然后停止，本次对话不再重复提醒 |
| 用户无回应，直接继续新话题 | 不视为永久跳过——下次出现结束信号时，可以再提醒一次 |
Five-Step Flow
Step 1: Full Scan — 6-Category Classification
Scan the entire conversation. Classify everything with distillation value. Skip any empty category — don't force it.
| Category | KnowMine Tool | Tagging Rule |
|---|---|---|
| 💡 Insights / Conclusions | add_knowledge | type: insight; bilingual tags |
| 🎯 Decisions | add_knowledge | type: note; title prefix [Decision Log] |
| 📊 Facts / Data | add_knowledge | type: reference; stable: ✅, time-sensitive: 🕒 + date |
| 🪞 Observations about the user | observe_user_trait or save_memory | Use save_memory for preferences/habits; observe_user_trait for inferred traits |
| ✅ Action items / TODOs | add_knowledge | type: note; tag todo:open; NO new folder — use tag |
| ❓ Open questions | add_knowledge | type: note; tag open-question |
Step 2: Relationship Mapping
Find connections. Default to granular over aggregated:
- Same decision, different angles → separate entries, cross-reference in body
- A is prerequisite for B → mention A's title/ID in B's body
- Insight came from a fact → note the source
Do not merge into one hub document. Granular entries have higher vector search precision.
Step 3: User Confirmation (Mandatory)
Present the list:
This conversation produced N items worth saving to KnowMine:
💡 Insights (2)
1. {title} → add_knowledge [insight]
2. {title} → add_knowledge [insight]
🎯 Decisions (1)
3. [Decision Log] {title} → add_knowledge [note]
✅ Action items (2)
4. {title} → add_knowledge [note] #todo:open
5. {title} → add_knowledge [note] #todo:open
Tell me:
- Numbers to remove
- Numbers to edit (number + new version)
- Numbers to merge
- Say "write" when ready
Iron rule: do not call any KnowMine tool until the user explicitly says "write" or equivalent.
Step 4: Batch Write to KnowMine
After confirmation, write entries sequentially. Report back the KnowMine ID for each success. List failures separately — user decides: retry / rewrite / skip.
Tool mapping:
- Insights, decisions, facts, action items, open questions →
add_knowledge - User preferences, habits, self-observations →
save_memory(type:preference) orobserve_user_trait
Title format: [Decision Log] {title} for decisions; plain title for everything else.
Tags: always include at least one English tag. Add native language tag if the content is in another language.
Step 5: Surface Leftovers
Output anything not worth saving as plain Markdown:
## Leftovers (not saved)
- [rough idea or reminder]
- [something to try next time]
Key Principles
- Granular over hub — one insight per entry beats one long summary
- Confirm before write — mandatory, no exceptions
- Tags over folders for TODOs —
todo:opentag, not a new folder - Bilingual tags — EN + native language improves cross-language recall
- Time-sensitivity — flag stale-prone data with 🕒 + date
Self-Check Before Step 3
- Any empty categories removed?
- Every decision has
[Decision Log]prefix? - Time-sensitive data marked 🕒 + date?
- Action items tagged
todo:open, not put in a new folder? - Any "fake hub" entries that should be split?
Standalone Version
This KnowMine edition is part of the knowmine-claude-plugin.
A tool-agnostic version (works with any notes tool or plain Markdown output) is available separately:
→ github.com/YIING99/conversation-distill
→ npx clawhub@latest install conversation-distill