Meeting Audio Grounding

Meeting Audio Grounding

Convert a meeting audio file into structured meeting grounding outputs by:

1. reusing the existing audio_structuring skill to produce meeting_transcript.txt
2. reusing the existing meeting-grounding skill to turn that transcript into meeting grounding outputs

This skill is for meeting audio inputs where the primary information comes from speech. It is intentionally transcript-first.

When to Use

Use this skill when:

• the input is a meeting recording or discussion audio file
• the main information is expected to come from spoken content
• you want to reuse the existing audio transcription and meeting grounding workflow

Do not use this skill when:

• the task is only to produce a transcript without meeting grounding
• the task is to write a polished final report directly from raw audio
• the input is not a meeting/discussion recording

Input

A single meeting audio file.

Typical examples:

• .mp3
• .wav
• .m4a
• .webm
• .aac
• .ogg

All these formats are supported via WhisperX's ffmpeg backend.

Optional input:

• transcription_language: language code such as en or zh

Output Bundle

For each input audio file, create one bundle directory:

data/grounded_notes/<ground_id>/

Inside that bundle, the expected outputs are always:

<bundle_dir>/
├─ extracted.md
├─ extracted_meta.json
├─ grounded.md
├─ audio/
│  └─ meeting_audio<original_ext>
└─ transcript/
   └─ meeting_transcript.txt

If the meeting contains multiple independent topics that should be researched separately downstream, the bundle may also contain:

<bundle_dir>/
├─ topic_manifest.json
└─ child_outputs/
   ├─ topic_01/
   │  └─ grounded.md
   ├─ topic_02/
   │  └─ grounded.md
   └─ ...

Important separation of responsibilities

• scripts/run.sh is responsible for:
  • preparing the bundle directory
  • ensuring the input audio is present under audio/
  • calling the existing audio_structuring skill
  • creating the bundle files:
    • audio/meeting_audio<original_ext>
    • transcript/meeting_transcript.txt
    • extracted.md
    • extracted_meta.json
• The agent is responsible for:
  • reading the transcript bundle
  • applying the existing meeting-grounding skill
  • always writing the meeting-level:
    • grounded.md
  • and, when appropriate, also writing:
    • topic_manifest.json
    • child_outputs/topic_xx/grounded.md

grounded.md must be a real grounding note. It must not remain a placeholder scaffold.

Required Agent Workflow

1. Run the existing scripts/run.sh entrypoint for this skill.
2. Confirm that the bundle exists and that these files are present:
   • extracted.md
   • extracted_meta.json
   • audio/
   • transcript/meeting_transcript.txt
3. Read transcript/meeting_transcript.txt as the primary grounding evidence.
4. Reuse the existing meeting-grounding skill on that transcript.
5. Always save the meeting-level structured note to grounded.md inside the same bundle directory.
6. If the transcript clearly contains multiple independent topics, also save:
   • topic_manifest.json
   • child_outputs/topic_xx/grounded.md
7. Do not stop after confirming that the transcript bundle exists.

Output Format

The final grounded.md must follow the existing meeting-grounding schema exactly.

If topic children are created, each child grounded note must also follow the same meeting-grounding schema.

Do not invent a new schema here.

Instructions

• Do not implement a new ASR pipeline.
• Do not implement a new meeting summarizer.
• Do not directly summarize the raw audio without first running the existing workflow.
• Reuse the existing audio_structuring skill for transcription.
• Reuse the existing meeting-grounding skill for transcript grounding.
• Treat the transcript as the primary evidence.
• Keep the workflow simple and stable.

Failure Handling

• If the input audio file does not exist, fail clearly.
• If meeting_transcript.txt is not produced, fail clearly.
• Do not pretend the task succeeded if only part of the workflow completed.

Example Invocation

/meeting-audio-grounding

Source: gaotiexinqu/OneResearchClaw — distributed by TomeVault.