Ai Meeting Helper
会议纪要生成器 - 自动将会议录音转为结构化纪要
MIT-0 · Free to use, modify, and redistribute. No attribution required.
⭐ 0 · 37 · 0 current installs · 0 all-time installs
MIT-0
Security Scan
OpenClaw
Suspicious
medium confidencePurpose & Capability
The code, SKILL.md, and skill.json all implement a meeting-transcription-and-summary tool that uses OpenAI Whisper and an LLM — requesting an OPENAI_API_KEY is coherent with that purpose. However, the top-level registry metadata provided to you (which claimed no required env vars / no primary credential) contradicts the included skill.json and SKILL.md that require OPENAI_API_KEY. The mismatch is unexpected and should be resolved before trusting the package metadata.
Instruction Scope
Runtime instructions and the Python code stay within the declared purpose: they read local audio files, call OpenAI APIs for transcription and summarization, format results, optionally backup original audio to a local .ai_meeting_backup/ directory, and write outputs. The instructions do not reference unrelated system paths, other credentials, or external endpoints beyond the OpenAI client.
Install Mechanism
Installation is a local install.sh that runs pip3 install openai python-dotenv (no virtualenv) and creates a .env template and backup/log directories in the skill folder. This uses a public PyPI package (openai) — not high risk by itself — but the script installs packages system-wide (no venv) which may be undesirable. No downloads from arbitrary URLs or obfuscated installers were found.
Credentials
The code requires a single credential (OPENAI_API_KEY), which is proportionate to the stated functionality. However, the provided registry metadata (at the top of this report) claims no required env vars/primary credential while skill.json and SKILL.md declare OPENAI_API_KEY required — that inconsistency is suspicious and could indicate sloppy packaging or metadata manipulation.
Persistence & Privilege
The skill does not request 'always: true' or elevated agent privileges. It creates and writes files only within its own skill directory (.env, .ai_meeting_backup/, .ai_meeting_logs/) and does not modify other skills or global agent config. Uninstall.sh offers to remove these files.
What to consider before installing
What to consider before installing:
- Metadata mismatch: the skill.json and SKILL.md require OPENAI_API_KEY, but the top-level registry metadata claimed no required env vars. Treat that inconsistency as a red flag — confirm which metadata is authoritative before trusting automated installers.
- Review code yourself (you have the files): the Python script uses only the OpenAI client and local file I/O; no hard-coded external endpoints or obfuscated code were found. If you don't review, install only in an isolated environment.
- Install in a virtual environment or sandbox rather than system-wide (the install.sh runs pip3 install without a venv). This avoids contaminating global Python packages.
- Backups: the tool will copy original audio into .ai_meeting_backup/. Those files can contain sensitive conversations — ensure you control that directory, or disable backup if not needed, and securely delete backups when appropriate.
- API key scope and monitoring: provide an OpenAI key with appropriate usage controls; monitor usage and billing. If possible use an account with limited permissions or monitoring alerts.
- Trust and provenance: owner is unknown and there's no homepage. If you need high assurance, request more provenance or prefer a skill from a known source.
If you decide to proceed: run install.sh only after inspecting it, execute the tool inside a dedicated virtualenv or container, and consider auditing network traffic or logs during first runs. If you need help verifying specific lines of code or testing in a sandbox, provide the environment and I can help step through it.Like a lobster shell, security has layers — review code before you run it.
Current versionv1.0.0
Download ziplatestmeetingsummarizationtranscription
License
MIT-0
Free to use, modify, and redistribute. No attribution required.
SKILL.md
AI Meeting Helper - 会议纪要生成器
自动将会议录音转换为结构化会议纪要,包括行动项、决策点和待办事项。
功能
- 🎙️ 语音转文字:使用 OpenAI Whisper API 将会议录音转为文本
- 📋 自动总结:使用 LLM 分析对话内容,生成结构化纪要
- ✅ 提取要点:自动识别行动项、决策点、待办事项
- 📤 多格式输出:支持 Markdown、纯文本、JSON 格式
- 🔄 批量处理:一次处理多个会议录音文件
- 👀 预览模式:不实际生成文件,只显示预览
- ↩️ 撤销功能:自动备份原始文件,可撤销操作
快速开始
1. 安装技能
clawhub install ai-meeting-helper
cd ~/.openclaw/workspace/skills/ai-meeting-helper
./install.sh
2. 配置 OpenAI API
export OPENAI_API_KEY="your-api-key"
3. 使用示例
# 单个文件处理
ai-meeting-helper process meeting_recording.mp3 --output meeting_notes.md
# 批量处理文件夹
ai-meeting-helper batch ./meetings/ --output ./notes/ --format markdown
# 预览模式(不生成文件)
ai-meeting-helper process meeting.mp3 --preview
# 启用撤销功能
ai-meeting-helper process meeting.mp3 --output notes.md --backup
参数
| 参数 | 类型 | 必需 | 描述 |
|---|---|---|---|
input | 路径 | 是 | 输入音频文件或目录 |
--output | 路径 | 否 | 输出纪要文件路径(默认:当前目录) |
--format | 选项 | 否 | 输出格式:markdown/text/json(默认:markdown) |
--model | 字符串 | 否 | Whisper 模型:tiny/base/small/medium/large(默认:base) |
--llm-model | 字符串 | 否 | 总结使用的 LLM 模型:gpt-4o/gpt-4o-mini(默认:gpt-4o-mini) |
--preview | 布尔 | 否 | 预览模式,不生成文件 |
--backup | 布尔 | 否 | 启用备份(原始文件保存到 .ai_meeting_backup/) |
--help | 布尔 | 否 | 显示帮助信息 |
输出格式
Markdown(默认)
# 会议纪要
**日期**:2026-03-15
**时长**:45分钟
**参会人数**:5人
## 讨论要点
1. 讨论了Q1产品发布计划
2. 确定了市场营销策略
3. 分配了开发任务
## 行动项
- [ ] @张三:完成用户文档(3月20日前)
- [ ] @李四:设计海报素材(3月18日前)
## 决策点
- 采用A方案作为最终设计
- 预算增加10%
## 待办事项
- 下周一下午2点开会复盘
JSON
{
"date": "2026-03-15",
"duration": "45分钟",
"participants": 5,
"summary": "讨论了Q1产品发布计划...",
"action_items": [
{"assignee": "张三", "task": "完成用户文档", "due": "2026-03-20"}
],
"decisions": ["采用A方案", "预算增加10%"],
"todo": ["下周一下午2点开会复盘"]
}
技术栈
- 语音识别:openai-whisper-api(需 OPENAI_API_KEY)
- 文本处理:Python 标准库 + 正则
- LLM 总结:OpenAI GPT-4o-mini(默认)或 GPT-4o
- 文件处理:glob, json, pathlib
依赖
pip install openai python-dotenv
注意事项
- 需要有效的 OpenAI API key
- 音频文件支持:MP3, WAV, M4A, FLAC, OGG
- 较大文件处理时间较长(建议 < 30 分钟音频)
- 建议使用高质量录音以提高识别准确率
与其他技能配合
- audio-note-taker:类似功能,但 ai-meeting-helper 专为会议场景优化,输出更结构化
- wechat-formatter:生成的纪要可直接排版发布到公众号
- social-publisher:纪要内容可一键分发到多平台
License
MIT
Files
5 totalSelect a file
Select a file to preview.
Comments
Loading comments…
