• Document conversion: Use

  markdown-tools

  skill to convert .docx/.pdf to Markdown first (preserves tables/images)
• Transcript cleanup: Use

  transcript-fixer

  skill to fix ASR/STT errors if transcript quality is poor
• Context file: Prepare

  context.md

  with team directory for accurate speaker identification

1. Read the transcript provided by user
2. Load project-specific context file if provided by user (optional)
3. Intelligent file naming: Auto-generate filename from content (see below)
4. Speaker identification: If transcript has "Speaker 1/2/3", identify speakers before generation
5. Multi-turn generation: Use multiple passes or subagents with isolated context, merge using UNION
6. Self-review using references/completeness_review_checklist.md
7. Present draft to user for human line-by-line review
8. Cross-AI comparison (optional): Human may provide output from other AI tools (e.g., Gemini, ChatGPT) - merge to reduce bias
9. Iterate on feedback until human approves final version

YYYY-MM-DD-<topic>-<type>.md

2026-01-25

api-design

product-roadmap

review

sync

planning

retro

kickoff

• 2026-01-25-order-api-design-review.md

• 2026-01-20-q1-sprint-planning.md

• 2026-01-18-onboarding-flow-sync.md

• Meeting topic and attendees
• Key decisions with supporting quotes
• Action items with owners
• Deferred items / open questions

context.md

1. Load team directory section
2. Match feature patterns to known team members
3. Cross-reference roles with speaking patterns

undefined

📊 Transcript Quality Assessment:
- Content: 41,837 chars (+2)
- Filler ratio: 3.6% (+2)
- Main speaker: 99% (+1)
- Technical depth: High (+1)
→ Quality Score: 10/10 (High)
→ Recommended: Full structured minutes with diagrams

"Transcript quality is low (碎片对话/噪音较多). Generate full minutes or summary only?"

// Launch ALL 3 subagents in PARALLEL (single message, multiple Task tool calls)
Task(subagent_type="general-purpose", prompt="Generate complete meeting minutes from transcript...", run_in_background=false) → version1.md
Task(subagent_type="general-purpose", prompt="Generate complete meeting minutes from transcript...", run_in_background=false) → version2.md
Task(subagent_type="general-purpose", prompt="Generate complete meeting minutes from transcript...", run_in_background=false) → version3.md

// After all complete:
Main Agent: Read all versions → UNION merge, consolidate duplicates → draft_minutes.md

1. Full path to transcript file
2. Full path to output file (version1.md, version2.md, version3.md in transcript-specific subdirectory)
3. Context files to load (project-specific context if provided, meeting_minutes_template.md)
4. Reference images/documents if provided by user
5. Output language requirement (match user's language preference, preserve technical terms in English)
6. Quote formatting requirement (see Quote Formatting Requirements section below)

• Each pass independently analyzes the SAME content
• Different context states catch different details (no single pass catches everything)
• Pass 1 might catch decision X but miss action item Y
• Pass 2 might catch action item Y but miss decision X
• UNION merge captures both X and Y

• Each pass/agent starts fresh without prior assumptions
• No cross-contamination between passes
• Different "perspectives" emerge naturally from context isolation

<output_dir>/intermediate/

<output_dir>/intermediate/<transcript-name>/version1.md
<output_dir>/intermediate/<transcript-name>/version2.md
<output_dir>/intermediate/<transcript-name>/version3.md

project-docs/meeting-minutes/2026-01-14-api-design.md

• Intermediate files:

  project-docs/meeting-minutes/intermediate/2026-01-14-api-design/version1.md

• This prevents conflicts when multiple transcripts are processed in the same session
• The

  intermediate/

  folder should be added to

  .gitignore

  (temporary working files)

// Create transcript-specific subdirectory first
mkdir: <output_dir>/intermediate/<transcript-name>/

// Launch all 3 subagents IN PARALLEL (must be single message with 3 Task tool calls)
Task 1 → Write to: <output_dir>/intermediate/<transcript-name>/version1.md (complete minutes)
Task 2 → Write to: <output_dir>/intermediate/<transcript-name>/version2.md (complete minutes)
Task 3 → Write to: <output_dir>/intermediate/<transcript-name>/version3.md (complete minutes)

Merge Phase:
  Read: <output_dir>/intermediate/<transcript-name>/version1.md
  Read: <output_dir>/intermediate/<transcript-name>/version2.md
  Read: <output_dir>/intermediate/<transcript-name>/version3.md
  → UNION merge, consolidate duplicates, INCLUDE ALL DIAGRAMS → Write to: draft_minutes.md

Final Review:
  Read: draft_minutes.md
  Read: original_transcript.md
  → Compare & add omissions → Write to: final_minutes.md

• Conversation context stays clean (avoids token overflow)
• Intermediate results persist (can be re-read if needed)
• Each pass starts with fresh context window
• Merge phase reads only what it needs
• Human can inspect intermediate files for review - Critical for understanding what each pass captured
• Supports very long transcripts that exceed context limits
• Enables post-hoc debugging - If final output is missing something, human can trace which pass missed it

• <output_dir>/intermediate/<transcript-name>/version1.md

  ,

  version2.md

  ,

  version3.md

  - Each subagent output
• These files help human reviewers understand the merge process
• Do NOT delete intermediate files after merge
• Human may want to compare intermediate versions to understand coverage gaps
• Add

  intermediate/

  to

  .gitignore

  - These are temporary working files, not final deliverables
• Transcript-specific subdirectory prevents conflicts when processing multiple transcripts

• Chinese output with English technical terms preserved
• Evidence-based decisions - Every significant decision needs a supporting quote
• Structured sections - Executive Summary, Key Decisions, Discussion, Action Items, Parking Lot
• Proper quote formatting - See Quote Formatting Requirements section below
• Mermaid diagrams (STRONGLY ENCOURAGED) - Visual diagrams elevate minutes beyond pure text:
  • ER diagrams for database/schema discussions
  • Sequence diagrams for data flow and API interactions
  • Flowcharts for process/workflow decisions
  • State diagrams for state machine discussions
  • Diagrams make minutes significantly easier for humans to review and understand
• Context-first document structure - Place all reviewed artifacts (UI mockups, API docs, design images) at the TOP of the document (after metadata, before Executive Summary) to establish context before decisions; copy images to

  assets/<meeting-name>/

  folder and embed inline using

  ![description](assets/...)

  syntax; include brief descriptions with the visuals - this creates "next level" human-readable minutes where readers understand what was discussed before reading the discussion
• Speaker attribution - Correctly attribute decisions to speakers

• Never assume - Ask user to confirm if unclear
• Quote controversial decisions verbatim
• Assign action items to specific people, not teams
• Preserve numerical values (ranges, counts, priorities)
• Always use multiple passes - Single turn is guaranteed to lose content
• Normalize equivalent terminology - Treat trivial variations (e.g., "backend architecture" vs "backend", "API endpoint" vs "endpoint") as equivalent; do NOT point out or highlight such differences between speakers
• Single source of truth - Place each piece of information in ONE location only; avoid duplicating tables, lists, or summaries across sections (e.g., API list belongs in Discussion OR Reference, not both)

• Minutes require careful line-by-line human review
• Domain experts catch terminology conflicts AI may miss
• Final version emerges through iterative refinement

• Accept as-is (rare for complex meetings)
• Request deeper review for missing content
• Identify terminology issues (e.g., naming conflicts with existing systems)
• Provide another AI's output for cross-comparison

• Different AI models have different biases - Each AI catches different details
• Cross-validation - Content appearing in both outputs is likely accurate
• Gap detection - Content in one but not the other reveals potential omissions
• Error correction - One AI may catch factual errors the other missed (e.g., wrong date, wrong attendee name)

1. Read the external AI output carefully
2. Identify items present in external output but missing from our draft
3. Verify each item against original transcript before adding (don't blindly copy)
4. Identify items where external AI has errors (wrong facts) - note but don't copy errors
5. UNION merge valid new content into our draft
6. Document any corrections made based on cross-comparison

• Missing decision about API authentication method ✓ (add to our draft)
• Missing naming convention specification ✓ (add to our draft)
• Wrong date (2026-01-13 vs actual 2026-01-14) ✗ (don't copy error)
• Wrong attendee name ✗ (don't copy error)
• Missing database performance concern ✓ (add to parking lot)

1. Re-read transcript section by section
2. Compare each section against current minutes
3. Look for: entities, field names, numerical ranges, state transitions, trade-offs, deferred items
4. Add any omitted content
5. Never claim "nothing missing" without thorough section-by-section review

1. Keep ALL content from existing version
2. Add ALL new content from incoming version
3. Consolidate duplicates (don't repeat same info)
4. Preserve more detailed version when depth differs
5. Maintain logical section numbering

1. Inventory ALL diagrams from each version (v1, v2, v3)
2. Include ALL unique diagrams - don't assume a diagram is redundant
3. If similar diagrams exist, keep the more detailed/complete one
4. Check every section that could contain diagrams: Executive Summary, Discussion, API design, State machines, Data flow, etc.

• Sequence diagrams (data flow, API interactions)
• ER diagrams (database schema, table relationships)
• State diagrams (state machines, status transitions)
• Flowcharts (decision flows, process workflows)
• Component diagrams (system architecture)

• Primary: Match the language of the transcript (or user's preference if specified)
• Preserve in English: Technical terms, entity names, abbreviations (standard practice)
• Quotes: Keep original language from transcript

• Data flow discussions → Sequence diagram
• Database schema discussions → ER diagram
• Process/workflow decisions → Flowchart
• State machine discussions → State diagram
• System architecture → Component diagram

* **Quote:** > "This is wrong" - **Speaker**

* **Quote:**
  > "This is the correct format" - **Speaker**

* **Quote:**
  > "First quote from the discussion" - **Speaker1**
  > "Second quote supporting the same decision" - **Speaker2**

• * **Quote:**

  on its own line (no quote content on this line)
• Blank line NOT needed after

  * **Quote:**

• Quote content indented with 2 spaces, then

  > 

  prefix
• Speaker attribution at end of quote line:

  - **SpeakerName**

• Multiple quotes use same indentation, each on its own line

• Decision: Specific decision made
• Logic:
  • Reasoning point 1
  • Reasoning point 2
• Quote:

  "Exact quote from transcript" - Speaker Name

undefined

1. Terminology conflicts: Humans know existing system naming (e.g., "Note" already means comments in the existing system)
2. Domain context: Humans catch when a term could be confused with another (e.g., "UserProfile" vs "Account")
3. Organizational knowledge: Humans know team conventions and prior decisions
4. Completeness gaps: Humans can request "deep review" review when something feels missing

• ❌ Single-pass generation - One turn through transcript will absolutely lose content
• ❌ Divided sections without overlap - Each pass must cover FULL transcript, not split by sections
• ❌ Narrow-focused passes - Each pass must generate COMPLETE minutes (all sections), not just one section type (wastes tokens, causes bias)
• ❌ Generic summaries without supporting quotes
• ❌ Action items assigned to "team" instead of specific person
• ❌ Missing numerical values (priorities, ranges, state counts)
• ❌ State machines with incomplete states
• ❌ Circular debate transcribed verbatim instead of summarized
• ❌ Removing content during multi-version merge
• ❌ Claiming "nothing missing" without section-by-section review
• ❌ Treating first draft as final without human review
• ❌ Using terminology without checking for conflicts with existing systems
• ❌ Shared context between subagents (causes cross-contamination and missed content)
• ❌ Keeping all intermediate outputs in conversation context (causes token overflow, use file system)
• ❌ Pure text minutes without diagrams - Architecture/schema discussions deserve visual representation
• ❌ Deleting intermediate files after merge - Preserve for human review and debugging
• ❌ Blindly copying external AI output - Always verify against transcript before merging
• ❌ Ignoring cross-AI comparison opportunity - Different AI models catch different details
• ❌ Sequential subagent execution - MUST launch v1, v2, v3 subagents in PARALLEL using multiple Task tool calls in a single message
• ❌ Flat intermediate directory - MUST use transcript-specific subdirectory

  intermediate/<transcript-name>/

  to avoid conflicts
• ❌ Inline quote formatting - Quotes MUST use blockquote format on separate lines, not inline

  > "quote"

• ❌ Omitting diagrams during merge - MUST aggressively include ALL diagrams from ALL versions, even if they seem similar
• ❌ Highlighting trivial terminology variations - Do NOT point out differences like "backend architecture" vs "backend" or "API" vs "endpoint" between speakers; these are equivalent terms and highlighting such differences is disrespectful
• ❌ Duplicate content across sections - Do NOT repeat the same information in multiple sections (e.g., API endpoint table in both "Discussion" and "Reference"); place content in ONE authoritative location and reference it if needed
• ❌ Creating non-existent links - Do NOT create markdown links to files that don't exist in the repo (e.g.,

  [doc.md](reviewed-document)

  ); use plain text for external/local documents not in the repository
• ❌ Losing content during consolidation - When moving or consolidating sections, verify ALL bullet points and details are preserved; never summarize away specific details like "supports batch operations" or "button triggers auto-save"
• ❌ Appending domain details to role titles - Use ONLY the Role column from Team Directory for speaker attribution (e.g., "Backend", "Frontend", "TPM"); do NOT append specializations like "Backend, Infrastructure" or "Backend, Business Logic" - all team members with the same role should have identical attribution