usredanchenko/eda-marketing

caption-writer

Three separate things: on-screen captions (video/captions/, props.captions, speech.srt); post text (captions/<platform>.md, skill eda-marketing:content-repurposer); editorial text (accents with spoken: false, editorialCta) — never in the SRT.

Caption text is in the brand's language (brand.json → language). Grouping, typography and syllable rules come from the language pack (packages/core/src/lang/<en|ru>.ts).

Before recording (draft)

npm run mos -- captions build <pkg> — estimated timing from the script (brand.json → speech.syllablesPerSec), "ESTIMATED TIMING" badge, final render blocked.

After recording

1. npm run mos -- footage <pkg> (if sources are in input/).
2. npm run mos -- captions transcribe <pkg> <voice file> — local faster-whisper, no downloads. No model → ask permission before installing (docs/VIDEO_PIPELINE.md).
3. ASR is a draft. Check names, numbers, negations, word endings, quiet words, phrase boundaries. Fixes go to transcript/corrected.<file>.json + a line in transcript/CORRECTIONS.md; transcript/raw-asr.* is immutable. Unclear → [inaudible] + ask the user; never guess. Never claim you listened if you did not.
4. If real speech differs from the script, show the difference and decide with the user (keep speech / re-edit / re-record). Captions follow the actual speech; never add words that are not in the audio.
5. npm run mos -- captions map <pkg> → npm run mos -- captions build <pkg> --source transcribed.
6. Optional: npm run mos -- captions calibrate <pkg> raw-asr.<file>.json --write — the author's real pace into brand.json.

Modes and grouping

Default mode: brand.json → defaults.captionMode; per video: props.captions.mode. replace — short groups replace each other; accumulate — words build a small stepped block; highlight — on request. Emphasis: *word* in script speech (for transcribed speech, mark emphasis in corrected). Group by meaning. Never split a negation from its verb ("don't / know"), a preposition from its word ("in / the app"), a number from its unit ("30 / seconds"), or a name. Avoid one dangling word on a line. Keep brand.json → captions.protectedPhrases intact. Russian rules live in the ru language pack.

Check

Studio: start/end of each page, caption-to-accent handoff without duplicates, readability on a phone, no text over the eyes. mos qa checks line overflow with real font metrics.