Capture Thoughts

Capture Thoughts

Capture ideas in the user's existing Notion Thoughts area with minimal interruption. Preserve the user's meaning and voice, confirm the database tag with the user, create a useful title, and return the created page link.

Choose the interaction mode

• Use quick capture by default. Treat a clear request to record an accompanying thought as authorization to prepare the entry, but do not write until its Tags value is explicit.
• Use conversational capture when the user asks to think aloud, explore, refine, or talk first. Help develop the idea across turns and wait until the user says to save, record, or put it into Thoughts.
• If the user explicitly says not to save yet, continue the conversation without writing.
• Beyond the required tag selection, ask one concise question only when the intended note is missing or the transcription is too ambiguous to preserve reliably.

Confirm the tag

• Require an explicit Tags choice before every create operation. If the user already named one exact existing tag for this entry, treat it as confirmed; otherwise ask: “这条用哪个 TAG？卡片盒 / 永久笔记 / 文献笔记 / 网络摘抄 / My thoughts / AI学习笔记”.
• In quick capture, hold the prepared draft while asking for the tag, then write after the answer.
• In conversational capture, wait until the user is ready to save, then ask for the tag if it is still unspecified.
• Never infer the tag from whether the input was spoken, typed, pasted, or produced through an AI conversation. A thought developed with AI may belong to any existing tag.
• Use AI学习笔记 only when the user chooses it. Its intended meaning is the user's learning notes or reflections about vibe coding and current AI knowledge; it does not mean every thought discussed with AI.
• If the user gives a label outside the current options, show the existing options and ask them to choose. Do not create a new option.

Create an entry

1. Read references/thoughts-database.md before every write for the canonical destination and property mapping.
2. Fetch the canonical data source before writing to verify its current schema. If it is unavailable or materially changed, follow the recovery procedure in the reference instead of guessing.
3. Read the Notion enhanced Markdown MCP resource before sending page content when the resource is available. Follow the current connector specification.
4. Prepare the entry:
   • Generate Name as a specific, concise title in the same language as the thought. Prefer the central claim or question over generic titles such as “New thought” or a date.
   • Set Tags to the exact existing option explicitly confirmed by the user. Never supply a default.
   • Set Also see only when the user provides related concepts, titles, or links. Do not invent connections. Omit it when empty.
   • Put the substantive note in the page body. Remove obvious speech fillers, repeated fragments, and clear transcription artifacts while preserving tone, uncertainty, examples, and meaning.
   • Keep a short thought as one or two paragraphs. For a long thought, use a few descriptive headings or bullets only when they improve readability.
   • Do not repeat the generated page title as a heading in the body.
5. Create exactly one page under the canonical data source with the mapped properties and body. Do not apply the database's literature-note template.
6. Report success concisely with the final title and a clickable Notion page link. Mention inferred fields only when that helps the user notice a potentially important interpretation.

Handle follow-up changes

• Interpret “再加一句”, “改成…”, or similar immediate follow-ups as edits to the entry just created in the current conversation.
• Fetch that page and the data source schema before updating it.
• Use the smallest suitable update: update properties for title/tag changes, targeted content replacement for corrections, or append content for additions.
• If the user starts a clearly separate thought, create a new entry rather than overwriting the previous one.
• Never create a second entry merely because a connector response was delayed or uncertain; verify the first operation before retrying.

Guardrails

• Do not change the database schema, views, templates, option names, or parent page.
• Do not silently assign any tag, including My thoughts.
• Do not add claims, conclusions, emotional certainty, citations, or links the user did not supply.
• Do not expose internal data-source IDs in the user-facing confirmation.

Examples

• “帮我记一下：我发现拖延很多时候不是懒，而是任务定义得不够清楚。” → prepare the entry, ask which existing tag to use, then create it after the answer.
• “把这条存成 My thoughts：我发现拖延很多时候不是懒，而是任务定义得不够清楚。” → create without asking the tag again because it is already explicit.
• “我想跟你聊聊为什么我最近不想做研究，最后帮我存进 Thoughts。” → discuss first; when the user signals completion, ask for the tag if it has not been specified, then create one synthesized entry.
• “我们刚才通过 AI 对话产生了这个想法。” → do not infer AI学习笔记; ask the user to choose any existing tag.
• “先别存，我还没想清楚。” → continue talking and do not write.