Claude Code skill · open source · MIT
transcribe
Turn a long recording into the document someone would otherwise spend a day writing.
Give it a two-hour meeting, a workshop, an interview or a folder of voice notes. It gives you back one Word file: a short TL;DR, the decisions with who owns them and by when, a glossary of the terms people used, and the whole conversation written out clean, so you can check every line of the summary.
/plugin marketplace add dihannahdi/transcribe
/plugin install transcribe@dihannahdiThen type /transcribe with your files. The first run checks your setup and tells you exactly what’s missing.
What you get
One document, in the order people actually read it. This is the opening page of the sample.
Lumen Launch Planning
TL;DR, Summary and Full Transcript
| Recording | Product planning call, two speakers |
|---|---|
| Length | 1 min 37 s, two files joined |
TL;DR
- Launch stays on 12 December. Budget is $42,000; marketing gets about half.
- Cut the second onboarding screen. 6 of 10 test users dropped off there.
- Priya builds a prototype by Friday, with Marco from engineering for two days.
- Pricing page is wrong: it says $9 per seat; the agreed price is $12. Daniel fixes it tonight.
- TL;DR8–12 lines, each starting with the decision or the fact.
- SummaryBy topic, with a timeline, follow-ups (who, what, by when) and what still needs confirming.
- GlossaryEvery term, acronym and name that came up, explained.
- Full transcriptThe whole conversation, cleaned, in nine-minute sections with exact start times.
- How it was madeWhich model, what was edited, and what the markers mean.
Short on time? Choose Summary only. Need every word as spoken? Choose Verbatim. You pick at the start.
How it works
You answer a few questions once. Then it runs on its own and comes back with the file.
- 1
Listen
Your files are joined in order and cut into nine-minute pieces at real pauses, so no sentence is split in half. Gemini transcribes five pieces at a time, word for word, in whatever languages were spoken.
- 2
Check
Before running the whole recording, it transcribes one moment you can vouch for, like your own introduction, and compares. Afterwards an audit flags every piece that came back empty, cut short or with a gap, and those are redone.
- 3
Clean
Claude rewrites each piece as readable prose: fillers and false starts go, every fact, number and name stays. Anything it can’t be sure of goes on one short list for you to answer.
- 4
Write
Claude reads the whole transcript and writes the TL;DR, the summary and the glossary, builds the Word file, and opens it to check the pages before handing it to you.
Why this one
Works on what you already recorded
No bot joins your call. A phone recording, a stack of WhatsApp voice notes, a seminar video, three files that belong together: any of these works.
Decisions, with the evidence attached
Speech-to-text gives you every word. Meeting apps give you bullets you can’t verify. This gives you a summary to act on, with the full transcript right behind it.
It tells you what it checked
Models sometimes invent a passage or skip a minute. This one tests itself before trusting a run, and when it hands over the file it says what was verified and what is still a guess.
Built for how people really talk
Indonesian, English and Javanese in one sentence are transcribed as spoken. Nicknames, product names and who was absent come from one short glossary, so nobody gets credited with a line they never said.
No subscription
It runs on your own Gemini API key, and you pay only for what you transcribe. Before a long batch starts, it estimates the cost.
Good to know
- Cost
- About 25–40k tokens per nine minutes of audio, measured. A 2 h 20 min session came to 466k.
- Privacy
- Audio goes to the Gemini API, text to Claude, nowhere else. On Gemini’s free tier Google may use what you send to improve its products; a key with billing turned on keeps it out of that.
- Names
- A speaker gets a name only with evidence: they introduce themselves, or someone calls them by it. Otherwise they stay “Participant”, never a guess.
- Needs
- Claude Code, Python 3.10+, ffmpeg and a free Gemini API key. The first run checks all of it.
It started as a private tool for turning a two-hour Indonesian sharing session, recorded in a noisy café, into minutes for the next year’s committee. Every check in it is there because something went wrong once: a model that made up a passage, a file merge that turned three minutes into nearly eleven hours, a pause detector that heard no pauses in a café.
Get it on GitHub