A summariser cannot hear. Hand it a garbled span and it launders the noise into a clean fact — "some paying customer" becomes "~8,000 users" — and nothing downstream can separate that from a real figure. This skill holds a chain of custody: every claim in the notes traces back to audio the model actually heard well, and a claim whose custody is broken says so.

Scripts are in scripts/. They need a Python with httpx; --provider local also needs faster-whisper.

1. Establish the recording's context

Ask the user, or read it off the surrounding material — the files, the repo, an existing transcript:

Language — the ISO code to pin. Left unpinned, per-window detection flaps on code-switched speech and the decoder emits fluent, confident text in the wrong language.
Speakers — how many, plus names and roles.
Domain vocabulary — product names, people, companies, jargon, currencies.

Done when you can state the language code, the speaker count, and at least five domain terms.

2. Transcribe

python scripts/transcribe.py AUDIO... --out DIR --language <code> --diarize --keyterms "term,term,..."