Audio

Audio support is intentionally a thin bridge. source-to-skill does not bundle a speech model. It can call an installed Whisper-compatible CLI, write a transcript file, and then use the existing transcript cleanup and readiness gate.

Command

source-to-skill transcribe-audio recording.m4a \
  --out out/recording.vtt \
  --model base \
  --language en

Then clean and score the transcript:

source-to-skill clean-transcript out/recording.vtt \
  --out out/recording-clean.md \
  --title "Recording Transcript"

source-to-skill analyze out/recording-clean.md

For long recordings, split the cleaned transcript before building:

source-to-skill split-source out/recording-clean.md --out out/recording-topics
source-to-skill build-segment out/recording-topics 2 --level seed --out out

If the source is strong enough:

source-to-skill build out/recording-clean.md --level auto --out out

Product Boundary

The default posture for recordings should be conservative. A voice note or meeting often contains useful local context, but not enough reusable judgment for a standalone skill.

Prefer:

• Note for ordinary meeting context
• Skill Seed for one or two reusable rules
• Mini Skill only when the recording has a clear method, examples, and repeatable decision rules

For long recordings, make that decision per topic segment rather than for the whole transcript.

Requirements

Install a Whisper-compatible command line tool separately. The default command name is whisper, but another executable can be passed:

source-to-skill transcribe-audio recording.m4a \
  --engine whisper \
  --out out/recording.vtt

The command writes vtt, srt, or txt transcripts.