- name
- feishu-doc-summarizer
- description
- Summarize Feishu/Lark cloud documents the user can read. Use when the user sends a Feishu doc link (docx or wiki) and expects an auto summary without extra instructions, or explicitly asks to summarize a Feishu document. Workflow: resolve wiki to docx, read doc content via feishu_doc, summarize using the user’s fixed summary schema from MEMORY.md, and send the summary back to the current chat.
Feishu Doc Summarizer
Overview
Given a Feishu/Lark document link (docx or wiki), read the document content and reply in-chat with a structured summary following the user’s fixed schema stored in MEMORY.md.
Workflow
0) Input detection
- Trigger when the user message contains a Feishu/Lark docx link like:
- https://...larkoffice.com/docx/DOC_TOKEN
- Also trigger when the user message contains a Feishu/Lark wiki link like:
- https://...larkoffice.com/wiki/WIKI_TOKEN
The user may provide no additional instructions; default to summarizing.
1) Resolve link → doc token (if needed)
- If it is a docx link: extract
doc_tokendirectly. - If it is a wiki link:
- Use feishu_wiki(action=get, token=wiki_token) to resolve the underlying object. - If the object type is docx, extract its doc token. - If not docx, tell the user what type it is and what you can support.
2) Read document content
- Use
feishu_doc(action=read, doc_token=doc_token)to retrieve the full document content. - If permissions fail or content is empty, ask the user to confirm they granted read access.
3) Summarize with the fixed schema (from memory)
- Before drafting, retrieve the summary schema from memory (search MEMORY.md for “Feishu 云文档摘要偏好” / “固定模板”).
- Produce the summary strictly following the section order and rules in memory.
- Citations: In the “引用原文片段” section, only quote text that appears in the document; do not fabricate.
Long documents
If the document is very long: 1) Chunk by headings/paragraph groups, summarize each chunk briefly. 2) Merge chunk summaries into the final schema. 3) Keep “引用原文片段” short and representative.
4) Reply back to the user
- Send the final formatted summary back to the current conversation.
- Include the original link in “文档元信息/链接”.
Output format (must follow memory)
Follow the user’s schema in MEMORY.md exactly; keep missing items as “无/未知” rather than removing sections.