For detailed rules on converting spoken language to written form, see references/polishing-guide.md. For the lecture notes output template, see references/lecture-notes-template.md—required reading before organizing. For design decisions, see DECISIONS.md.
Overview
Organize the recording transcripts of legal AI courses (spoken verbatim transcripts produced by tools such as Tingwu/FunASR/Jianying) intostructured written lecture notes in Markdown: remove spoken-language redundancy, correct transcription errors, and reorganize into modular lecture notes for publishing, archiving, or study use.
Boundary with lecture-review : This skill only performscontent organization(spoken → written); it does not evaluate the instructor's performance or analyze verbal tics and rhythm;lecture-review only performs review of teaching performance and does not rewrite content. The two skills have the same input but completely different outputs. When the user asks to "organize into lecture notes/written draft," use this skill; when the user asks to "review teaching performance," use lecture-review。
Dependencies
No external dependencies, works out of the box: this skill is a pure text organization workflow, read through and rewritten directly by the agent, and does not depend on any scripts or Python packages. For legal provision verification, an already-connected legal retrieval tool in the environment (such as yuandian MCP) may optionally be used; when that environment is unavailable, handle it by marking "to be verified" (see Core Principle 4).
Core Principles (Non-Negotiable)
- Faithful to the original meaning, only organize, do not create. Remove spoken-language redundancy, correct transcription errors, and reorganize structure, butdo not add new viewpoints, do not rewrite conclusions, and do not supplement content the speaker did not say. The organized draft is the speaker's draft, not the agent's draft.
- The standard for converting spoken language to written form: It is permissible to complete missing subjects, merge slips of the tongue, adjust word order, and make connectives more written in style; it is not permissible to "package" the speaker's plain-language viewpoints into professional expressions he never said (for detailed rules, see polishing-guide.md).
- Citations must be traceable. Quotable lines, cases, and data in the lecture notes should have a sense of provenance—retain the approximate position of the paragraph in the original transcript (timestamp or sequence number) so the user can check against the original text.
- Mark when uncertain. When legal provision numbers, case names, tool names/versions, or numerical data in the transcript cannot be confirmed, mark them as "⚠ To be verified" and do not correct them on your own. When a legal retrieval tool is available, verify first and then mark.
- Explicitly mark missing information. For meta-information such as course topic, speaker, and date that is not in the transcript, mark "not mentioned/to be supplemented" in the lecture notes information header, and do not speculate.
- Reminder about sensitive information. When the transcript contains real parties or unpublished case information, after organizing, remind the user to review privacy and compliance.
Workflow
任务进度：
- [ ] Step 1 预检输入
- [ ] Step 2 通读全文，识别课程结构
- [ ] Step 3 拟定大纲（长稿先确认）
- [ ] Step 4 口语转书面整理
- [ ] Step 5 专业内容核对
- [ ] Step 6 按模板生成讲义
- [ ] Step 7 校验与交付
Step 1: Pre-check the input
- Use Read to read the transcript and confirm completeness (whether the beginning and end are truncated, whether multiple speakers are mixed in).
- Record available meta-information: presence of timestamps, number of speakers, word count of the draft. For drafts without timestamps, use "original sequence number (near paragraph N)" for traceback positioning instead.
- When the transcript lacks meta-information (course name/lecturer/date), ask the user once; if the user does not answer, mark "To be supplemented" in the information header and continue.
Step 2: Read the full text and identify the course structure
Read through completely (read all pages if very long), and mark:
- Course skeleton: opening introduction → knowledge modules → case explanations → operation demonstrations → Q&A → closing
- Content types: concept explanation / legal analysis / tool demonstration (including prompts) / case / personal experience and opinions / interactive Q&A
- Colloquial problem areas: obvious slips of the tongue, transcription typos (homophones), repeated paragraphs
Step 3: Draft an outline (confirm first for long manuscripts)
- Manuscript ≤ 5000 characters: organize directly, and just write the outline into the handout table of contents.
- Manuscript > 5000 characters: first give the user a one-page outline (module division + key points of each module), and expand after confirmation. If the user explicitly says "organize directly", skip confirmation.
Step 4: Convert colloquial speech to written form
Follow references/polishing-guide.md , and remember the boundaries of the three types of operations:
- Delete: filler words (um/ah/this and that), information-free repetition, off-topic chatter (can be moved into the appendix "Digressions" rather than simply discarded)
- Change: add subjects, rebreak and reorganize sentences, written connectives, correct homophone typos (limited to obvious transcription errors)
- Keep: the speaker's memorable quotes (keep the original wording and add a quote block), first-person experience, case details, valuable on-site Q&A
Step 5: Verify professional content
- Legal provision citations: when the number and the content of the provision do not match, mark "⚠ To be verified"; if the environment has legal research tools (such as yuandian MCP), search and verify first, and when the verification result is inconsistent with the speaker's original words,keep the speaker's original words and add a note"The speaker said X orally, but upon search it is Y".
- AI tool names/functions/versions: when obviously outdated or wrongly written, likewise "keep the original words + editor's note".
- Numbers (dates, amounts, proportions): corrections that can be self-evidenced from context may be made; those that cannot be self-evidenced should be marked as to be verified.
Step 6: Generate the handout according to the template
Follow references/lecture-notes-template.md to output the Markdown handout and save it to the location specified by the user; if unspecified, use the same directory as the transcript, named {课程名}-讲义.md(if the course name is unknown, use course-notes-{日期}.md）。
Step 7: Validate and deliver
Before delivery, self-check item by item:
- No new viewpoints: spot-check the concluding sentence of each module to ensure it can be traced back to the original manuscript
- No colloquial remnants: randomly read 3 paragraphs and ensure there is no accumulation of "um/ah/that is to say"
- All "⚠ To be verified" and "Not mentioned" annotations are centrally findable (there is a list in the handout appendix)
- The quoted memorable lines have not changed the original wording
- The meta-information header is complete (missing items are marked as to be supplemented rather than left blank)
When delivering, report to the user: an overview of the handout structure, the extent of organization (a rough description of the deletion/modification ratio), the number of items on the to-be-verified list, and sensitive information reminders (if applicable).
Outputs
| Artifact | Description | Default |
|---|---|---|
| Handout Markdown | Structured written handout (see references for the template) | Required output |
| Digressions/behind-the-scenes appendix | Off-topic but valuable content is moved into the handout appendix | Merged into the handout by default |
| Clean memorable quotes collection | Output separately only when the user requests *-quotes.md |
Do not produce |