Happy Scribe

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 subtitling pipeline that replaces Happy Scribe.

**Transcription is free now; subtitle craft is not.** The value you are building is in the segmentation and timing rules, which most homemade versions ignore and which decide whether subtitles are readable.

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

THE DATA MODEL
- media, words (start_ms, end_ms, text, confidence), cues (start_ms, end_ms, lines_json, speaker)
- styles: id, name, font, size, position, background, safe_area
- reviews: id, cue_id, reviewer, note, resolved_at
- Words are the source of truth; cues are derived and rebuildable. **Store both, because a retiming must not lose the corrections**

SEGMENTATION, WHICH IS THE CRAFT
- Two lines maximum, around 42 characters each. Longer and the eye cannot take it in
- Minimum duration around one second, maximum around seven
- **Reading speed capped at about 17 characters per second.** This one number is the difference between subtitles people can read and subtitles that flash past
- Break lines at clause boundaries, never between an article and its noun, never leaving a preposition stranded
- A gap of a couple of frames between consecutive cues so they do not visually merge
- Never let a cue cross a hard cut in the video if you can avoid it — detect cuts with ffmpeg and snap to them

TIMING
- Start a cue slightly before the first word and end slightly after the last, from the word timings
- Snap to the frame rate; a cue starting mid-frame is a rounding error the player shows
- A shift-all control for the classic problem of audio offset by a constant

REVIEW
- Play with the cue overlaid on the actual video, at the actual size. **A subtitle judged in a text editor is judged wrong** — position, contrast and safe area only exist on screen
- Comments per cue, resolvable
- A validation pass listing every rule violation: too fast, too long, overlapping, crossing a cut. Fix the list, not the file

EXPORT
- SRT, VTT and TTML. Burnt-in through ffmpeg with the style applied, into the safe area
- Contrast: a background box or a strong outline, never plain white text over unknown video

TRANSLATION
- Translate cue by cue, keeping the timings, then **re-run the reading-speed check** — translated text is often much longer and the timing that worked no longer does. This is the step everybody skips and it is why translated subtitles are unreadable

WHAT MATTERS MOST
The reading-speed cap and re-checking after translation. Those two rules are most of what a professional subtitler is doing, and they cost nothing to encode.