michael-f-bryan/obsidian-recording-to-meeting-minutes

Obsidian Recording to Meeting Minutes

Overview

Convert a sparse or partial Obsidian meeting note into a complete, evidence-grounded note with:

• summary and key details
• chapter list
• polished diarised transcript

This workflow is coordinator-led and specialist-executed:

1. coordinator resolves note + recordings, merges audio, transcribes once, and builds manifest
2. specialists handle speaker attribution, transcript polishing, optional secondary cross-check, minutes extraction, and final QA
3. coordinator merges outputs and writes the vault note

Use references/coordinator-workflow.md as the canonical execution order.

Prerequisites

• Know the vault absolute path and read <vault>/AGENTS.md before writing.
• Keep frontmatter, prep notes, and original ![[...audio]] embeds unless asked otherwise.
• Have ffmpeg and scribe available.

Install scribe if needed:

uv tool install git+https://github.com/trailofbits/scribe.git

If the environment cannot run scribe, stop and report the blocker. Do not fabricate transcript content.

Data handling

Treat /tmp/<slug>/ working files as sensitive:

• merged audio
• diarised JSON
• chunk excerpts
• specialist outputs

Do not paste full transcripts into unrelated chats or tools.

Non-negotiable architecture

1. Single transcription pass: resolve all audio embeds, concatenate into one scratch file, run scribe once.
2. Manifest-first orchestration: create /tmp/<slug>/manifest.json and bounded chunk files before specialist prompts.
3. Speaker attribution before polishing: produce a global speaker map from the merged diarisation, then apply per-chapter caveats only when needed.
4. Transcript polishing must use /transcript-polisher: polishing sub-agents follow that skill in strict Obsidian mode (faithful cleanup, not summarisation).
5. QA is a separate sub-agent gate: do not rely on a local verification script as the final check.

Workflow

1. Read rules and source note
   • Read <vault>/AGENTS.md.
   • Read the target note and preserve the existing human-written notes + embeds.
2. Resolve recordings and concatenate
   • Parse all ![[...m4a]], ![[...mp3]], ![[...wav]].
   • Concatenate in note order with ffmpeg into /tmp/<slug>/audio/merged.*.
   • See references/merged-audio-scribe-workflow.md for the exact merged-audio pattern.
3. Transcribe once
   • Run scribe --format json -o /tmp/<slug>/scribe/merged.json on the merged file, not each embed separately.
4. Build manifest and chunks
   • Write /tmp/<slug>/manifest.json and chapter/chunk excerpt files.
5. Run specialist phases
   • Speaker attribution (references/speaker-attribution-agent.md + references/speaker-attribution-from-prior-notes.md).
   • Transcript polishing (references/transcript-polishing-agent.md, explicitly invoking /transcript-polisher).
   • Optional secondary transcript comparison (references/secondary-transcript-agent.md).
   • Minutes extraction (references/minutes-extraction-agent.md).
6. Merge and draft final note
   • Merge specialist outputs using references/final-note-template.md.
   • Keep the original audio embeds and the prep notes unless the user asks for a rewrite.
7. Run QA sub-agent gate
   • Spawn the QA specialist with explicit pass/fix/block criteria from references/qa-audit-agent.md.
   • Apply blockers before writing to vault.
8. Write note + manual skim
   • Write the final note, then re-read the first chapter and one later chapter for fidelity and readability.
   • Fix obvious ASR noise, but keep uncertain speaker labels uncertain.

Speaker attribution rules

• Because diarisation comes from one merged transcription pass, treat speaker IDs as globally consistent by default.
• Validate mappings with:
  • attendee list in frontmatter
  • human-written notes already in the note body
  • explicit transcript anchors and role/context evidence
• Preserve uncertainty explicitly when evidence is weak.

Quick reference

| Need | Default | | ---------------------- | ---------------------------------------------------------------------- | | Transcription strategy | Merge all embedded recordings, then one scribe pass | | Chunking strategy | Build chapter/chunk files after transcription; polish per chapter | | Speaker mapping | Global map first (merged diarisation), caveats where needed | | Transcript polishing | Separate sub-agent, must apply /transcript-polisher | | Verification | Separate QA sub-agent gate + manual skim | | Raw transcript storage | Keep scratch artifacts in /tmp/<slug>/, not in vault note by default |

Pitfalls

• Skipping <vault>/AGENTS.md rules.
• Transcribing each embed separately when merged transcription is available.
• Running polishing without a prior speaker-attribution pass.
• Polishing by rewriting/summarising instead of faithful cleanup.
• Converting speculative discussion into decisions without timestamp evidence.
• Treating QA as optional or replacing it with only a script check.