建立會議記錄 Skill：從逐字稿生成結構化會議紀錄

2026-06-18
ai-agentmeeting-notesskill

建立會議記錄 Skill：從逐字稿生成結構化會議紀錄

這份文件可以讓同事在自己的電腦上建立一個獨立 skill：輸入一份會議逐字稿，輸出和我們目前 meeting note 接近的繁體中文會議紀錄格式。

它不依賴 OpenClaw、Google Drive、Slack、speaker ID pipeline 或公司內部資料庫。這份 skill 只處理一件事：根據逐字稿整理會議記錄。

適合使用的情境

建立方式

在你的 skills 目錄中新增資料夾：

meeting-transcript-notes/
└── SKILL.md

把下面整段內容存成 meeting-transcript-notes/SKILL.md。

---
name: meeting-transcript-notes
description: Use this skill whenever the user provides a meeting transcript and wants a structured meeting note, 繁體中文會議記錄, meeting summary, action items, decisions, or notes organized by topics. This skill turns raw transcripts into concise, attributable, decision-oriented meeting records.
---

# Meeting Transcript Notes

You turn meeting transcripts into structured Traditional Chinese meeting records.

The transcript is the only source of truth. Do not invent context, decisions, names, deadlines, owners, metrics, or background that does not appear in the transcript.

## Trigger

Use this skill when the user provides:

- a meeting transcript
- a recording transcript exported from Google Meet, Zoom, PLAUD, Whisper, VibeVoice, Otter, Fireflies, or similar tools
- raw speaker-labeled dialogue
- a request such as "整理會議記錄", "做 meeting note", "summarize this transcript", "extract action items", or "產生會議紀錄"

## Required Input

Ask for missing information only when it is necessary. If metadata is missing, use sensible placeholders and continue.

Preferred input:

- Meeting title
- Meeting date
- Attendees
- Duration
- Transcript with speaker labels and timestamps

Acceptable transcript forms:

- `[00:03:12] Marcus: ...`
- `Marcus: ...`
- `Speaker 1: ...`
- raw text without clear speaker labels

If speaker names are unclear, keep labels such as `Speaker 1` instead of guessing real names.

## Language Rules

- Write in Traditional Chinese.
- Preserve product names, company names, technical terms, tools, and English phrases in English.
- Add spaces around English terms inside Chinese sentences, for example `Linear`, `Sprint Planning`, `context layer`.
- Do not translate proper nouns unless the transcript itself uses a translated name.
- Use concise business writing. Avoid dramatic wording.

## Ground Truth Rules

- Treat the transcript as the canonical source.
- Do not use external knowledge unless the user explicitly provides it.
- Do not infer that a person owns an action item unless the transcript clearly assigns or implies it.
- Do not turn vague discussion into a decision.
- If something is ambiguous, mark it as unresolved instead of filling the gap.
- If speaker labels are unreliable, mention that speaker attribution may need confirmation.

## Output Format

Return the meeting note directly in Markdown. Do not wrap the final answer in a code block unless the user asks for a file.

Use this structure:

---
title: "{meeting_title} — {YYYY-MM-DD}"
domain: meetings
lifecycle: record
status: completed
tags: []
created: {YYYY-MM-DD}
updated: {YYYY-MM-DD}
date: {YYYY-MM-DD}
author: "@MeetingBot"
acl: company
source_type: transcript
meeting_series: ""
attendees: [{attendees}]
duration_minutes: {duration_minutes}
source_recording: ""
---

## Overview

**參與者：** {attendees}
**會議目的：** {一句話說明這場會議想處理什麼}

### Summary
{100-200 字，說明這場會議的主軸、重要討論、主要結論。不要寫成逐點流水帳。}

---

## 議題

### 1. {議題標題}
- **狀態：** ✅ 已決議 / ⏳ 未決議
- **結論：** {一句話描述決議內容；若未決，說明卡住原因}
- **討論重點：**
  - {Speaker} 認為⋯⋯
  - {Speaker} 補充⋯⋯
  - {Speaker} 提到的關鍵數字、時程、限制或風險是⋯⋯

### 2. {議題標題}
- **狀態：** ✅ 已決議 / ⏳ 未決議
- **結論：**
- **討論重點：**
  - ...

---

## Action Items

- [ ] @{Name}：{任務描述}（deadline: YYYY-MM-DD）
- [ ] @{Name}：{任務描述}

If there are no clear action items, write:

沒有明確 action item。

## Extraction Rules

### Summary

The summary should answer:

- 這場會議在談什麼？
- 主要脈絡是什麼？
- 最終有哪些明確結果？

Do not overload the summary with all details. Put detailed discussion under topics.

### Topics

Split topics by meaning, not by timestamp.

Each topic should include:

- a short title under 30 Chinese characters
- a status: resolved or unresolved
- a one-sentence conclusion
- discussion bullets that attribute views to speakers where possible
- key numbers, dates, deadlines, metrics, percentages, costs, or constraints mentioned in the transcript
- disagreements or tradeoffs if they affected the conclusion

Do not merge unrelated discussions into one topic just because they happened close together.

### Decisions

A decision is only a decision if the transcript shows agreement, explicit selection, or a clear next direction.

Examples:

- "先不做" can be a decision if the group clearly agrees.
- "觀察兩週後再決定" can be a decision if the next checkpoint is clear.
- "有人提出一個想法" is not a decision.

When a decision has conditions, include the conditions.

### Action Items

Only include tasks that are clearly assigned or strongly implied.

Each action item should include:

- owner
- concrete action
- deadline, if mentioned

Do not include generic intentions such as "大家再想想" unless a specific person is responsible.

### Open Questions

If the meeting ends with unresolved issues, capture them inside the relevant topic. Use `⏳ 未決議` and explain what must be clarified next.

### Sensitive Meetings

Default `acl` is `company`.

Change `acl` when needed:

- `restricted` for HR, hiring, compensation, performance review, legal, finance-sensitive, or private strategy conversations
- `private` for personal notes or one-on-one coaching that should not be shared broadly

If the transcript appears sensitive, mention that the note should not be broadly shared until the owner confirms access level.

## Quality Self-Check

Before finalizing, check:

- Did I only use information from the transcript?
- Did I avoid inventing names, owners, deadlines, or decisions?
- Is the summary concise and non-repetitive?
- Are topics split by actual discussion themes?
- Does every topic include status, conclusion, and attributed discussion points?
- Are important numbers, dates, deadlines, metrics, tools, and risks preserved?
- Are action items limited to real commitments?
- Are unresolved items clearly marked?
- Is the output in Traditional Chinese?

## When Transcript Quality Is Poor

If the transcript is messy:

- Preserve speaker labels as given.
- Fix obvious repeated filler words and transcription noise.
- Do not silently correct names unless the transcript provides enough evidence.
- Add a short note at the top: `注意：逐字稿品質較差，speaker attribution 可能需要人工確認。`

## User-Facing Follow-Up

After producing the meeting note, optionally add a short final section:

## 需要人工確認

- {需要確認的人名、speaker label、決策或 action item}

Only include this section when there is real ambiguity.

建議的使用方式

建立好 skill 後，可以直接貼逐字稿並這樣要求：

請用 meeting-transcript-notes skill，把下面逐字稿整理成會議記錄。

會議名稱：Product + Ops Weekly
日期：2026-06-18
參與者：Marcus, Amy, David
時長：45 分鐘

逐字稿：
[00:00:01] Marcus: ...

如果你沒有完整 metadata，也可以只貼逐字稿：

請把這份 transcript 整理成繁體中文會議記錄，保留 action items 和未決議題。

與我們內部版本的差異

這份 shareable skill 保留的是「會議記錄的格式與整理邏輯」，不包含我們內部 pipeline 的自動化部分。

不包含：

保留：

小提醒

如果你想讓輸出更像自己的團隊，可以另外加一份 references/team-terms.md，放常見產品名稱、人名、專案名、容易轉錯的詞。使用時先把那份詞彙表一起貼給 AI，會比讓模型憑印象猜準得多。