Meeting Minutes

What you're buying: a download package (skill instructions + Python scripts) that runs on your own computer inside your own AI agent — it is not a cloud service.
You need: an AI agent that supports skills (e.g. Claude Code or Codex) — its subscription or API costs are yours; Python 3.9+, ffmpeg, faster-whisper and python-docx.
Third-party services: none. Transcription runs locally; the free Whisper model (~500 MB) downloads once from Hugging Face. Caption files and pasted transcripts need no internet at all.

One meeting recording in, minutes you can send out. Transcription happens on your computer — the recording is never uploaded.

What you get

• 📄 Word minutes (.docx): info table (date, attendees, length), summary, discussion points, decisions, action-item table, open questions, next meeting
• 📝 Markdown version: paste straight into Notion, Slack or email
• ✅ Action-items CSV: opens cleanly in Excel; import into Jira, Asana, Trello or Google Sheets
• ⏱️ Timestamps: every decision and action shows where it was said; if the recording is a YouTube or web link, timestamps are clickable

Example

A 3-minute product meeting with 3 people (Chinese) was transcribed in about 30 seconds and turned into a 2-page A4 Word document: 3 summary points, 3 topics, 4 decisions, 7 action items and 2 open questions.

• A task nobody was assigned was marked Unassigned in red — not guessed.
• Speech recognition heard "Dec 1" as "Feb 1"; the decision was flagged To confirm with the reason.
• The anti-fabrication check caught all 4 planted errors: a made-up owner, a due date with no source, a quote not in the transcript and a timestamp past the end of the recording.

The Word minutes it produced (page 1 and 2):

Try it — copy a prompt

• "Make minutes from weekly_sync.m4a. Attendees: Anna, Raj, Mei. Terms: Jira, SLA, Q3 roadmap. Standard format, English."
• "Here's our Teams caption file — write stand-up minutes and give me the action items as CSV."

Why buy it

One payment, minutes for every meeting — no monthly subscription, no per-minute transcription fees, and your recordings stay on your machine.

Who it's for

Team leads, project managers, consultants, assistants, researchers doing interviews and students recording lectures.

Inputs

• Audio or video: mp3, m4a, wav, mp4, mov, webm… (local Whisper, 99 languages; Chinese can be output in Traditional)
• Teams, Zoom or Google Meet captions (.vtt): imported instantly, speaker names kept
• Existing .srt subtitles or a pasted transcript

Five formats

Standard meeting, stand-up (yesterday / today / blockers), client meeting (requirements, commitments, timeline, risks), interview (key answers, quotes, observations), lecture (key concepts, examples, homework).

Minutes that don't make things up

• Owners: only people who were named or volunteered ("I'll do it"); otherwise Unassigned in red
• Due dates: only dates actually said, with the exact words quoted from the transcript
• Built-in check: owners not in the transcript or attendee list, due dates without a source, missing quotes and out-of-range timestamps are blocked until fixed
• Numbers and dates are cross-checked across the whole transcript; likely mis-hearings are marked To confirm

Better accuracy

Tell your agent the attendees' names and jargon (product names, terms) before starting — recognition gets noticeably better. Repeated mis-heard words can be fixed across the whole transcript with a replacement list.

Long meetings

Recordings over 25 minutes are split into 20-minute parts; the agent summarises each part and merges them, so two-hour meetings work too.

Requirements

• Python 3.9+ and ffmpeg (the skill checks your setup and tells you what to install)
• pip install faster-whisper python-docx (Traditional Chinese: also opencc-python-reimplemented)
• First transcription downloads the Whisper model (small ≈ 500 MB); works offline afterwards
• A normal CPU is enough: the small model transcribes 3 minutes of audio in about 30 s

Limitations (please read)

• These are AI-written minutes — have attendees check them before sending.
• It does not tell speakers apart by voice. Names come from Teams/Zoom captions or from introductions and call-outs in the conversation; uncertain attributions are left out, not guessed.
• Poor audio, cross-talk and strong accents increase recognition errors.
• Timestamps for local files are for reference only; they are clickable only when the recording is a URL.
• Not a legal verbatim record and not a substitute for official board or court minutes.
• Only process recordings you have the right to use, and make sure participants know they were recorded.

Privacy

Recordings, transcripts and minutes are processed only on your computer and never sent to the author, Capafy or any third party. The exception is your own AI agent, which reads the transcript to write the minutes — its provider's privacy policy applies. The skill collects nothing.

Files

• SKILL.md — agent workflow
• scripts/transcribe.py — local transcription / caption import, term hints, replacement list, long-recording split
• scripts/build_minutes.py — Word / Markdown / CSV output and anti-fabrication check
• references/templates.md — five minutes formats
• references/minutes-schema.md — data format
• references/writing-rules.md — writing rules
• references/example_minutes.json, example_teams.vtt — examples