raw-transcribe

name	raw-transcribe
description	Transcribe an audio/video file referenced in a raw entry (typically a Drive recording) and fold the transcript into raw/meetings/. Provider and model come from raw.transcribe.* config. Standalone-invokable; also wrapped by /dd:raw:transcribe.

You are the raw-transcribe skill. Convert a recording referenced from a raw entry into a transcript and fold it into the meetings subfolder.

Setup

source "${CLAUDE_PLUGIN_ROOT}/lib/config.sh"; dd_config_load
source "${CLAUDE_PLUGIN_ROOT}/lib/frontmatter.sh"

Procedure

1. Resolve target: input is a path to a raw entry whose frontmatter attachments lists at least one media reference (drive:<file-id> or local path). No attachments → exit: "no attachments referenced in ; nothing to transcribe."

2. Read transcription config:

   provider="$(cfg_get raw.transcribe.provider)"
   model="$(cfg_get raw.transcribe.model)"
   

   provider unset → exit: "raw.transcribe.provider not configured. See docs/extension-model.md § raw.transcribe."

3. Resolve attachment: drive:<file-id> → fetch via Drive MCP (or streaming URL if provider supports). Local path → read directly. Required MCP absent → fail with install guidance.

4. Run transcription: invoke configured provider with media. Capture transcript text plus structural markers (speaker turns, timestamps) where the provider supports them. Surface progress to the user — long operation.

5. Fold into entry: Edit raw entry. Append ## Transcript to body containing the transcript (preserve speaker turns / timestamps). Update frontmatter: set proposed_category: meetings (if not set), add transcribed_at: <ISO 8601> and transcribe.{provider, model} for provenance.

6. Move to meetings/: git mv <current-path> <repo-root>/$(cfg_get paths.raw_root)/$(cfg_get paths.raw_subdirs.meetings)/<filename>. Optionally normalize filename to YYYY-MM-DD_<topic>.md if user opts in.

7. Report: provider:model, transcript length, new path.

Guardrails

• Never call a transcription provider without raw.transcribe.provider explicitly set.
• Preserve provenance — transcript metadata (provider, model, transcribed_at) in frontmatter.
• Multi-attachment entries: transcribe each, concatenate transcripts under separated headers.
• Long-running — surface progress, never run silently.