Transcribe

Transcribe audio → markdown (local, whisper.cpp)

Local CPU transcription, language auto-detected (multilingual; tuned for EN/FR), no API/GPU. Output: markdown with frontmatter + full text + timestamps. The wrapper scripts/transcribe.sh self-discovers whisper.cpp ($WHISPER_DIR → ~/tools/whisper.cpp → ~/whisper.cpp → /opt/whisper.cpp). Invoke it via the plugin root: ${CLAUDE_PLUGIN_ROOT}/skills/transcribe/scripts/transcribe.sh.

Instructions

1. Resolve the audio file from the request ($ARGUMENTS). Accept m4a/mp3/wav/ogg/flac; if none given, ask.
2. Ensure the pipeline exists — probe once: bash "${CLAUDE_PLUGIN_ROOT}/skills/transcribe/scripts/transcribe.sh" --help >/dev/null 2>&1; echo $?. Exit 2 (or "whisper.cpp not found") → follow setup.md to build it (warn: one-time ~10-min build + ~1.5 GB download), then continue.
3. Run: bash "${CLAUDE_PLUGIN_ROOT}/skills/transcribe/scripts/transcribe.sh" "<audio-path>" (output defaults to <input>.md). Flags on request: --lang fr|en, --model small|large-v3 (only medium ships by default — others must be downloaded first, see setup.md), --no-timestamps, custom output as 2nd positional arg.
4. Set expectations: a 10–20 min file takes several minutes on CPU — let it run.
5. When done: Read the .md and show a short preview (frontmatter + first lines).

Notes

• WSL-safe: wrapper does all decode/temp in native /tmp; only final .md + source may live on /mnt/c.
• Non-standard install: export WHISPER_DIR=/path/to/whisper.cpp.
• Frontmatter fields, limitations, and upgrade paths (diarization, word-level timestamps, long files): see setup.md § Reference.