Trint

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 transcription and story workflow that replaces Trint.

The transcription is now free and local. **What you are building is the verification and quoting workflow**, which is where journalism actually happens.

STACK
- Node 20+ with Fastify, server-rendered HTML with a little vanilla JS
- SQLite through better-sqlite3, WAL mode, FTS5 across every transcript
- whisper.cpp with word timings, ffmpeg for conversion
- Caddy in front

THE DATA MODEL
- recordings: id, title, source, recorded_at, path, sha256, duration_ms, language, status
- words: id, recording_id, index, start_ms, end_ms, text, confidence
- edits: id, recording_id, word_index, before, after, editor, at — **append-only, so a corrected transcript still shows what the machine heard**
- highlights: id, recording_id, from_index, to_index, note, created_by
- stories: id, title, body_md
- quotes: id, story_id, highlight_id, text, verified_by, verified_at
- speakers: id, recording_id, label, name, verified

WHY THE EDIT HISTORY MATTERS HERE MORE THAN ANYWHERE
- A quote in a published piece must be defensible. **Keeping the machine's original alongside every human correction means you can always show what was actually said** and who changed what
- The source hash proves the audio has not been replaced
- Never allow a destructive overwrite of the word table

THE EDITOR
- Text with the audio playing along, current word highlighted, click to seek
- Keyboard-first: play and pause, rewind three seconds, and a slow-playback toggle, all without leaving the text
- Low-confidence words visibly marked, so attention goes where the machine was unsure
- Speaker labels applied once and propagated, with a merge and split

QUOTING, WHICH IS THE POINT
- Select text, press a key, and it becomes a quote attached to a story, carrying the recording, the timestamp and the speaker
- **A quote in the story links back to the exact second of audio.** An editor checking a piece can hear it in one click rather than searching a two-hour file
- Mark a quote verified, with who verified it and when. A story shows how many of its quotes are unverified

SEARCH ACROSS EVERYTHING
- Full text across every recording ever transcribed, results as moments with play buttons
- Filter by speaker, date and story. **The archive is the asset**: 'when did he say that' answered in seconds, across years

SECURITY, BECAUSE SOURCES
- Encryption at rest, access per recording, every read logged
- A retention and deletion path that reaches the media, and an option to keep a recording out of any shared index entirely
- Nothing sent to a third-party service. This is why local transcription matters here rather than merely being cheaper

WHAT MATTERS MOST
Append-only edits with the original preserved, and quotes that link to the second. Together they make the transcript evidence rather than a convenience.