Boardmeeting

• 1-3 → small roster, 4-5 → medium, 6-7 → large, 8+ → mega

If the topic is empty or under 5 words, ask the user to elaborate (one follow-up max).

Execution Flow

Phase 1: Gemini First Pass

Output: [BOARD] Convened: N experts on "Topic" Output: [BOARD] Gemini first-pass running...

Run the Gemini board meeting script:

bash ~/scripts/gemini_board_meeting.sh "$TOPIC" $EXPERT_COUNT

Parse the last line of stdout as JSON to get the output file path. Read the Gemini output file.

Shallowness check: If ANY of these are true, mark Gemini output as shallow:

• Script returned ok: false
• Output is under 800 words
• Output is missing both "Recommended Next Steps" AND a vote tally table

Output: [BOARD] Gemini complete (X words). Claude panel reviewing. Or: [BOARD] Gemini returned shallow results. Claude panel running independently. Or: [BOARD] Gemini dispatch failed. Claude panel running independently.

Phase 2: Claude Expert Panel

Spawn ALL agents in ONE message with run_in_background: true.

Select agents from this priority-ordered roster (top N based on size):

Compressed context packet (max 200 tokens per agent):

Role: [ROLE] | Topic: "[TOPIC]" | Check: [2-3 criteria from role] | Output: ADOPT/ADOPT_CTRL/DEFER/REJECT + 2-3 actions, <500w | Rules: challenge Gemini don't confirm, read files don't speculate, no file writes, be concrete to this codebase
[IF GEMINI]: Gemini advisory (challenge it): [key_sections trimmed] | Focus: what Gemini missed
[IF NO GEMINI]: No Gemini first-pass. Fully independent analysis.

Model routing:

• Opus agents (roles 1, 2, 4): Use model: "opus" in Agent tool
• Sonnet agents (all others): Use model: "sonnet" in Agent tool

Live status — compressed pipe-delimited format as each agent completes:

[BOARD] 1/N | Chief Architect | ADOPT | scalable, fits existing arch | next: synthesis
[BOARD] 2/N | Devil's Advocate | REJECT | hidden coupling risk | next: synthesis

Phase 3: Synthesis

After ALL agents return, produce the final verdict.

Determine verdict:

• Count votes: ADOPT, ADOPT WITH CONTROLS, DEFER, REJECT
• Majority wins. Ties broken by: ADOPT WITH CONTROLS > DEFER > ADOPT > REJECT
• If ADOPT + ADOPT WITH CONTROLS combined > 50%, verdict is ADOPT WITH CONTROLS

Determine confidence:

• HIGH: >= 70% agreement on one category
• MEDIUM: 50-69% agreement
• LOW: no category has majority (< 50%)

Output format (verdict-first):

# BOARD VERDICT: [Topic]
> Date: YYYY-MM-DD | Board: N experts | Confidence: HIGH/MEDIUM/LOW

## VERDICT: [OUTCOME] (X ADOPT / Y DEFER / Z REJECT)

[2-3 sentence executive summary of the decision and key reasoning]

## Key Dissent
[Strongest opposing view: who, what vote, and their core argument]

## Recommended Actions
1. [P1] [Highest-priority action]
2. [P2] [Second priority]
3. [P3] [Third priority]

Transcript saved: docs/board-meetings/YYYY-MM-DD-<slug>.md

Phase 4: Persist

1. Save full transcript to docs/board-meetings/YYYY-MM-DD-<slug>.md with all expert assessments
2. Append one-liner to wiki/_inbox.md using idempotent write:
   • Generate a deterministic ID: md5(entry_text + date).substring(0, 12) where entry_text is the one-liner without any id suffix
   • Read the last 50 lines of wiki/_inbox.md (not the whole file)
   • If any line already contains (id: <computed_id>) — skip the write and log [BOARD] Dedup: skipped duplicate inbox entry
   • Otherwise append: [BOARD] YYYY-MM-DD: [Topic] — [VERDICT] ([N] experts, [confidence]) (id: <id>)

Important Rules

• Gemini is top-tier (Google AI Ultra subscription, gemini-2.5-pro). Never describe as "free tier" or inferior.
• Never skip Devil's Advocate for medium+ boards — it is the primary bias correction mechanism.
• All Claude agents launch in ONE message — never sequential.
• Claude agents are READ-ONLY — no file writes, edits, or code changes during the meeting.
• The main session orchestrates — no meta-orchestrator agent.
• Anchoring bias mitigation: Every agent prompt explicitly says "challenge Gemini, don't confirm."
• Never exceed Gemini dispatch cap (200/day). If ok:false, proceed Claude-only.
• Never surface size choice interactively — parse from args or default to medium.

Cost Reference

Gemini cost is covered by Google AI Ultra subscription ($249.99/mo, unlimited). Claude cost scales with prompt size — estimates assume ~3,000 token prompts.

Integration Points

• Initiative Tracker: If topic matches an active initiative in MEMORY.md, note the link in synthesis
• Wiki: One-liner appended to wiki/_inbox.md for future compilation
• Command Deck: Transcripts in docs/board-meetings/ surfaceable in decisions tab
• Adversarial Evaluator: For ADOPT WITH CONTROLS verdicts with complex controls, recommend /adversarial on the implementation plan
• Advisor Strategy: For ADOPT verdicts on implementation tasks, recommend /advisor for execution phase