Otter.ai

Promptfree, for everyone, and the only version there is

Open the prompt in Claude CodeOpen the prompt in CodexOpen the prompt in Cursor

Copy

Build me local meeting transcription with speaker labels and a summary I can trust.

STACK
- whisper.cpp for transcription, pyannote or an equivalent for diarisation
- Node for the pipeline, SQLite for the index, everything on my own machine

INPUT
- A recording file, or a system-audio capture on macOS through a virtual device
- Never a meeting bot: joining calls on your behalf is a permissions and consent problem, and the README should say so

TRANSCRIPTION
- Chunk long audio with overlap and stitch, removing the duplicated overlap by aligning on word timings
- Store word-level timestamps, not just segments

SPEAKERS
- Diarise first, then transcribe per speaker turn — the other order loses turn boundaries
- Label speakers by a voice sample enrolled once per person; unknown speakers stay numbered rather than guessed

THE SUMMARY, HANDLED HONESTLY
- Summarise with a local or API model, but every claim in the summary must cite a timestamp range
- Anything the model asserts without a citable range is dropped before display
- The transcript, not the summary, is the record; the README says so

SEARCH
- SQLite FTS5 over transcripts, results linking to the timestamp
- Playback jumps to the word, not the segment

CONSENT
- A prompt to state that recording is happening, and a per-recording flag confirming everyone was told

Include a README on the model sizes and their real-time factor, and the legal position on recording where you are.