glebis/meeting-processor

Meeting Processor

Intelligent meeting transcript processor that auto-detects meeting type and applies type-specific extraction with optional interactive clarification.

When to Use

• After syncing Fathom or Granola transcripts (/fathom --today, /granola export)
• When asked to process, analyze, or summarize a meeting transcript
• When a new meeting transcript appears in the vault root matching YYYYMMDD-*.md
• For coaching sessions, delegate to coaching-session-summarizer skill instead

Prerequisites

pip install openai pyyaml

Requires CEREBRAS_API_KEY environment variable (uses Cerebras API with llama-3.3-70b).

Supported Meeting Types

| Type | Description | Key Extractions | |------|-------------|-----------------| | leadgen | Sales/business development calls | Commitments, pain points, budget, timeline, decision makers, deal stage, sentiment | | partnership | Collaboration/partnership exploration | Opportunity overview, value proposition, strategic alignment, technical needs, fit assessment | | coaching | Coaching/mentoring sessions | Insights, decisions, action items, themes, emotional arc, techniques, session quality | | internal | Internal team meetings | Coming soon |

Usage

Interactive Mode (default)

Run the processor, which auto-detects meeting type and asks clarifying questions:

python3 ~/.claude/skills/meeting-processor/scripts/process.py <transcript-file> --mode interactive

Interactive flow:

1. Script analyzes transcript and detects meeting type
2. Extracts structured data via LLM
3. Identifies missing/ambiguous fields
4. Returns questions as JSON (exit code 2 signals interaction needed)
5. Parse the JSON between __INTERACTIVE_QUESTIONS__ markers
6. Use AskUserQuestion to collect answers for each question
7. Save answers to a temp JSON file and re-run with process_with_answers.py

Handling interactive questions:

When the script exits with code 2, parse the output for questions JSON. Each question has:

• question: The question text
• header: Short label (used as answer key)
• options: Array of {label, description} for AskUserQuestion

After collecting answers, create two temp files:

• questions.json — the original questions context (includes partial_data, meeting_type, transcript_file)
• answers.json — map of {header_lowercase: selected_label}

Then run:

python3 ~/.claude/skills/meeting-processor/scripts/process_with_answers.py questions.json answers.json

Batch Mode

Extract only high-confidence information without user interaction:

python3 ~/.claude/skills/meeting-processor/scripts/process.py <transcript-file> --mode batch

Force Meeting Type

Skip auto-detection:

python3 ~/.claude/skills/meeting-processor/scripts/process.py <transcript-file> --type leadgen
python3 ~/.claude/skills/meeting-processor/scripts/process.py <transcript-file> --type partnership

Output

Analysis is appended to the transcript file as a ## Meeting Analysis section. Frontmatter is updated with meeting_type, processed_date, and processing_mode.

Leadgen Output Structure

• Commitments & Actions — with deadlines and owners
• Follow-up — next meeting date if scheduled
• Client Context — pain points, budget, timeline, decision makers
• Deal Assessment — stage (cold/warm/hot), probability (1-5), blocker, sentiment

Partnership Output Structure

• Opportunity — description and value proposition for both sides
• Commitments & Actions — with deadlines and owners
• Follow-up — next meeting date if scheduled
• Partnership Context — strategic alignment, technical needs, resources, challenges
• Opportunity Assessment — fit (strong/medium/weak), readiness, success factors, sentiment

Step 2: Auto-Link Prep Notes

After the meeting analysis is complete (Step 1), automatically link any matching meeting-prep notes to the session note. This replaces the need to manually run /meeting-prep link.

How It Works

1. Derive the meetings directory from the processed session note's parent directory (do not hardcode paths).

2. Extract session metadata from the processed note:

   • date from frontmatter (YYYYMMDD format)
   • participants from frontmatter (list of names)
   • If no participants field, extract names from the transcript header or attendee list

3. Search for matching prep notes:

   find <MEETINGS_DIR> -name "YYYYMMDD-prep-*" -type f 2>/dev/null
   

   Where YYYYMMDD is the session date.

4. Validate the match: For each candidate prep note, read its frontmatter and confirm:

   • The date field matches the session date
   • The participant field matches one of the session's participants (fuzzy: check both full name and first name, case-insensitive)
   • The session_note field is empty ("") — skip already-linked prep notes

5. Update both files when a match is found:

   In the prep note:

   • Set session_note: "[[session-note-filename]]" (without .md extension)
   • Set status: done

   In the session note:

   • If a ## See also section exists, add - [[YYYYMMDD-prep-participant-slug]] to it
   • Otherwise, append a new section at the end:

     ## Prep Note
     - [[YYYYMMDD-prep-participant-slug]]
     

   • Never create duplicate links — check if the link already exists before adding

6. Report in the processing output which prep notes were linked, skipped, or not found.

Rules

• Derive MEETINGS_DIR from the session note path, not from hardcoded values
• If the meeting-prep config.yaml is available, read prep_notes.prefix (default: prep) and prep_notes.type_tag (default: meeting-prep)
• This step is non-blocking: if it fails or finds no prep notes, processing still succeeds