Fireflies.ai

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me the meeting recorder and searchable archive I actually need instead of Fireflies.

**The joining bot is the hard part, not the transcription.** Transcription is free and local now; getting a reliable participant into three different conferencing platforms is a permanent maintenance job. Build the archive, and record locally.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode, FTS5 over the transcripts
- whisper.cpp for transcription, locally
- pyannote or a simpler energy-based split for speakers
- A local model through Ollama for summaries, or your own key
- Caddy in front

THE DATA MODEL
- meetings: id, title, started_at, ended_at, source, calendar_event_id, participants_json, audio_path, status
- segments: id, meeting_id, speaker, start_ms, end_ms, text — the unit of everything, because a search result must jump to a moment
- summaries: id, meeting_id, kind, body, model, generated_at — kinds are recap, decisions, actions
- actions: id, meeting_id, segment_id, text, owner, due, status
- consent: id, meeting_id, participant, method, at

GETTING THE AUDIO WITHOUT A BOT
- Record locally: system audio plus microphone, started by a hotkey or by a calendar trigger. **This sidesteps the entire joining problem and produces better audio than a bot in the room**
- Save a WAV to disk immediately and keep writing to it. A crash must cost the last seconds, not the meeting
- Chunk into fifteen-minute pieces so transcription can start before the meeting ends

CONSENT, WHICH IS NOT OPTIONAL
- **In many places recording without telling everybody is illegal, and in the rest it is a way to lose a client.** Announce it in the meeting, record that you did, and make the archive visible to the participants
- A per-meeting opt-out that stops the recording, not one that hides it afterwards
- A retention policy that is enforced by a sweeper, and deletion on request

TRANSCRIPTION AND SPEAKERS
- whisper.cpp with a medium model; the large one is better and slow, and a queue with a concurrency cap keeps the machine usable
- Diarise into speakers and label them once per person from a short voice sample; without labels the transcript is unusable
- Timestamps per segment, so clicking a line plays the audio there

SUMMARIES THAT ARE WORTH READING
- Three fixed sections: what was decided, what was agreed as an action and by whom, and what was left open. **Not a paragraph of prose** — nobody reads the paragraph
- Feed the model the transcript with speaker labels, and quote the segment id behind every extracted action so it can be checked
- Never let a summary be the only record. It is a convenience over the transcript, not a replacement

THE ARCHIVE, WHICH IS THE ACTUAL VALUE
- Full-text search across every meeting, results as segments with the surrounding lines and a play button
- Filter by participant, date and calendar event
- **This is why the product is worth having**: 'what did we agree with them in March' answered in five seconds

WHAT MATTERS MOST
Local recording and searchable segments. Skipping the joining bot removes the maintenance burden; the segment index is where the value lives.