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

Routing overlap? If this intent better matches a sibling skill, redirect per references/routing-tiebreaks.md (prefer internal priority debate over external research).

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

Trigger	Example
Product/strategy decisions	"Should we build X?" "A or B first?"
Scope decisions	"How far do we go?" "What's in Phase 1?"
Tradeoff analysis	"Performance vs features vs maintenance cost"
Risk assessment	"What's the blast radius of this change?"
CEO Agent hits strategic decision	CEO autonomously invokes autopilot:think-tank

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

Role	voltagent subagent_type	Focus
Architect	voltagent-qa-sec:architect-reviewer	Technical feasibility, coupling, performance, tech debt
Product Director	voltagent-biz:product-manager	Feature ROI, priority, scope, success metrics
UX Advocate	voltagent-biz:ux-researcher	User experience, interaction flow, edge cases
QA Devil	voltagent-qa-sec:qa-expert	Test coverage, regression risk, failure modes
Ops/SRE	voltagent-infra:sre-engineer	Deploy complexity, monitoring, rollback, resource usage
Customer Advocate	voltagent-biz:customer-success-manager	User needs, pain points, retention impact

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:

Topic Type	Suggested Roles
Technical approach	Architect + QA + Ops
Feature scope	Product + UX + Customer
Performance vs features	Architect + Product + Ops
New user-facing feature	UX + QA + Customer

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

Situation	Action
Agent timeout	Produce partial brief from available roles, mark missing ones
All roles agree	Normal — mark as "strong consensus", but verify no angle was missed
All roles oppose	Report to user, suggest not proceeding or redefining the topic
Topic too broad (multiple independent decisions)	Split into sub-topics, run one round per sub-topic

See Also

Skill	Relationship
autopilot:survey	Survey finds external data, think-tank does internal debate. Can chain.
autopilot:think-tank-dialectic	Escalation target for LOW consensus + irreversible decisions. Think-tank maps perspectives; dialectic resolves genuine stalemates via Hegelian synthesis. Use dialectic only after think-tank shows real tension, not as a default.
autopilot:ceo-agent	CEO autonomously invokes think-tank for strategic decisions
Parallel dispatcher (per .claude/dispatch-config.md, default: native Task tool)	think-tank fans roles out in parallel via the first available entry — superpowers:dispatching-parallel-agents if configured and installed, otherwise multiple Task tool calls in a single response