transcription-checklisted
Install: claude install-skill Shadysmetools/claude-skills
# /transcription-check
Claude cannot hear audio. So: **ffmpeg cuts the clip, Whisper hears it, Claude judges.**
This is a **shadow** of whatever verifier your product already runs (or the first verifier, if you
have none). It does not replace it. Label its KPI **"Transcript consistency (Claude, text-only),
unverified"** until it has been calibrated against human spot-listens. Accuracy first: never present
it as verified accuracy.
## Inputs (configurable)
- `CLIPS_SOURCE`: where today's sample comes from: a JSON/CSV export or a DB query returning
`{ clip_id, media_url_or_path, start_s, end_s, stored_text, prod_verdict? }`.
- `OUT_DIR`: default `./transcription-check/<day>/`.
- `WHISPER_API_KEY` (env) for a hosted Whisper endpoint, or a local `whisper` / `faster-whisper` install.
- `LANGS`: the languages you expect (e.g. `ar,en`).
- Spend cap per run (default $0.05).
## Steps
1. **Prep** (read-only on production):
- Load the clip list for `[day]` (default today, UTC). If your production verifier has not run
yet and you are shadowing it, stop and say so.
- Cut each window: `ffmpeg -ss <start> -to <end> -i <media> -ac 1 -ar 16000 -c:a libopus <OUT_DIR>/clips/<clip_id>.ogg`
- Transcribe each clip twice: language auto-detect, and forced to each expected language.
- Write `<OUT_DIR>/bundle.json` with `stored_text`, `whisper_auto`, `whisper_forced`, `prod_verdict`.
2. **Judge** each clip by comparing `stored_text` with both Whisper runs:
- **fail** if any of