think-tank

name: think-tank description: > Multi-role debate (architect, ops, QA, product, UX, finance) for strategic decisions. Use when: "I want perspectives from different roles", "let's debate this", "tradeoff analysis", "which should we prioritize", "blast radius of X", "scope decision", "hear from architect/ops/product", "what's the impact?", "要辯論一下", "多角度分析", "幫我評估利弊". Not for: pure tech selection (→ survey), open brainstorming, or full delegation (→ ceo-agent).

Think Tank — Multi-Role Decision Brief

Dispatch Chains (auto-injected)

!cat .claude/dispatch-config.md 2>/dev/null || true

6 specialized roles analyze the same topic in parallel, each producing independent perspectives, then cross-collide. Value lies in divergence, not consensus.

When to Use

Not for: pure technical selection (X library vs Y) → use autopilot:survey.

Escalation path: If R1 reveals LOW consensus (roles fundamentally disagree) AND the decision is irreversible or high-stakes, the Decision Brief will recommend escalating to autopilot:think-tank-dialectic for Hegelian cross-examination. Do not escalate for medium decisions — dialectic is expensive and only justified when genuine tension exists on an irreversible choice.

Relationship to Other Skills

survey                → external research (what does the industry do?)
think-tank            → internal multi-perspective debate (what should WE do?)  ← this skill
    ↓ (if LOW consensus + irreversible)
think-tank-dialectic  → Hegelian resolution (how do we reconcile real tension?)
    ↓
ceo-agent             → autonomous execution after decision

Chain: survey output → feed into think-tank → Decision Brief → (optional escalate to dialectic for irreversible LOW-consensus calls) → CEO executes.

Role Configuration

6 Standard Roles

voltagent is optional — degrade gracefully (autopilot is standalone-capable). The voltagent-* subagent_type values are preferred, not required. If no voltagent-* agents are installed (autopilot ships only reviewer/debugger/planner), dispatch each role via general-purpose with the inline role prompt (this panel MUST run with zero voltagent agents present).

Quick Mode (3 Roles)

Small decisions don't need 6 roles. Pick the 3 most relevant:

Execution Flow

Step 1: Define the Topic

Extract from user input:

• Topic: one sentence describing the decision
• Context: current state, known constraints
• Survey results: if a survey was run previously, summarize conclusions

If the topic is too vague, ask one clarifying round first.

Step 2: Prepare Domain Context

Read codebase context relevant to the topic (architecture docs, domain skills, related source code). Compile into a shared context block for all roles. Each role needs sufficient domain knowledge to produce valuable perspectives.

Step 3: Parallel Dispatch

Model routing: Read .claude/model-routing-config.md if it exists; otherwise use defaults from references/model-routing.md. Think-tank roles map to think-tank-role → default: model: "sonnet", mode: "plan".

Spawn all roles simultaneously (6 or 3), each agent gets:

• Role prompt (see references/role-prompts.md)
• Shared domain context
• Topic + constraints

Agent({
  subagent_type: "<voltagent-type>",
  model: "sonnet",           // from model-routing config (think-tank-role)
  mode: "plan",              // analysis only — no implementation
  prompt: "<role-prompt> + <domain-context> + <topic>",
  run_in_background: true,
  name: "tt-<role>"
})

Dispatch ALL roles simultaneously — do not wait for one to complete before dispatching the next.

Step 4: Cross-Collision Synthesis

After all roles report back, synthesize into a Decision Brief:

1. Find consensus — conclusions all roles agree on
2. Find divergence — conflict points between roles (this is the most valuable part)
3. Find collision insights — new perspectives emerging from role interactions (A's view + B's view → insight neither had alone)
4. Summarize recommendations — each role's stance (approve/conditional/reject) + one-sentence rationale

Step 5: Produce Decision Brief

Use the fixed format (see references/brief-template.md), including:

• Consensus
• Divergence map (ASCII diagram showing tensions between roles)
• Key divergence table (pro vs con + tension level)
• Role recommendation summary table
• Top 2-3 collision insights
• CEO recommendation (if in CEO mode)

Output Requirements

• Decision Brief must present complete results in a single response
• Divergence map uses ASCII diagram (not mermaid)
• Each role's output: max 300 words
• Brief should be readable in under 2 minutes

Error Handling

See Also