Heavy Think

Heavy Think

Unified orchestrator for heavy thinking. One entry point, four modes — all powered by parallel agents → your synthesis.

Core principle: the parallel-agents-→-synthesize pattern is general-purpose. Change what the agents explore, and you get different thinking capabilities from the same engine.

Teammates vs subagents: Use teammates (agent teams) when the agents need to talk to each other — debate, collaborative brainstorm, hand-offs. Use parallel subagents when they don't — independent fan-out. Teammates compound (live cross-talk) but cost more and require the main session + CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; subagents are cheaper and collide only once, in your synthesis.

Dispatch

Classify the problem, pick the mode:

When the signal is Debate (user explicitly asks to debate/pressure-test, or stakes are high), go straight to Escalation: Agent Debate and run the Team Debate protocol — don't simulate it inline. Requires the main session (a subagent can't spawn teammates). Debate is also reached automatically when synthesis hits irreconcilable positions.

When uncertain: ask the user which mode fits. Don't guess.

When modes should chain: decompose first → brainstorm on each piece, or unstick → then solve.

Mode 1: Brainstorm

Parallel perspective agents explore the same problem from different worldviews. Insight emerges from collision, not consensus.

Stage 1: Frame

PROBLEM: [One sentence — what are we trying to figure out?]
CONSTRAINTS: [What's fixed? Budget, timeline, team size, tech stack...]
SUCCESS LOOKS LIKE: [How will we know a good answer when we see it?]
MODE: scan (3 agents, breadth-first) | deep (3-5 agents, depth-first)

Stage 2: Design Perspectives

Choose 3-5 perspectives that collide productively — tension, not redundancy.

Rules: MUST include ≥1 challenger. MUST span ≥2 of: user, business, technical, temporal.

See references/perspective-combinations.md for pre-built sets by scenario.

Stage 3: Spawn Parallel Agents

Launch all in one message, model="opus". Each gets the same problem, distinct lens.

See references/brainstorm-agent-prompt.md for full prompt. Core structure:

You are a brainstorming agent exploring from one specific perspective.

PROBLEM: {problem} | CONSTRAINTS: {constraints}
YOUR PERSPECTIVE: {perspective_name} — {perspective_description}

Think from this perspective ONLY. Find what ONLY this lens reveals.

1. Challenge 2-3 assumptions most people take for granted
2. What is the REAL problem underneath the stated one?
3. Generate 3-5 ideas — push past the obvious first answer
4. Pick strongest. Stress-test: what breaks? what scales?
5. Second-order effects: if this succeeds, then what? And then what?

Output: ## Core Insight → ## The Real Problem → ## Ideas → ## Second-Order Effects → ## Hidden Risk → ## Provocation

Stage 4: Synthesize (YOU — never delegate)

1. Map — territory each perspective covered, overlap vs divergence
2. Collide — where perspectives directly contradict (deepest insight lives here)
3. Extract surprises — what exists in NO single output but emerges from combining
4. Second-order cascade — trace strongest ideas: then what? and then what?
5. Converge — landscape → key insights (ranked by surprise) → tensions worth holding → recommendation → what to kill

Stage 5 (Optional): Stress Test

For high stakes, spawn 1-2 agents to attack synthesis: Red team ("find every way this fails") or Pre-mortem ("it's 1 year later and this failed — what happened?"). See references/brainstorm-agent-prompt.md for templates.

Escalation: Team Brainstorm (collaborative teammates)

Stages 1–5 use isolated parallel agents — they collide once, in your synthesis. When you want ideas to compound (each builds on the others live), escalate to a collaborative council: named teammates that cross-pollinate via SendMessage, then YOU harvest the expanded set. (Per the teammates-vs-subagents heuristic above — agents that build on each other need to talk.)

Full protocol, spawn template, anti-groupthink rules, and variants: references/team-brainstorm.md. Adversarial counterpart: references/team-debate.md.

Mode 2: Solve

K independent agents solve the same verifiable problem from scratch. Deliberation audits reasoning chains and forges new paths from fragments — it produces correct answers absent from ALL trajectories in ~50% of cases.

When to use

• Mathematical reasoning, algorithmic/competition problems, complex logical deduction
• Any verifiable task where confidence < ~70%
• NOT for subjective, preference-oriented, or easy (>90% confidence) tasks

Stage 1: Parallel Reasoning

Spawn K independent agents in a single message (parallel). Zero sibling knowledge.

Solve this problem step by step. Show complete reasoning and arrive at a final answer.
Use whatever approach you find most natural — algebraic, geometric, constructive, brute force, or proof by contradiction.

Problem: {query}

Requirements:
- Reason from first principles, show all work
- Final answer clearly marked
- Math: \boxed{answer} | Code: code block

Stage 2: Memory Cache

1. Collect all K outputs
2. Shuffle trajectory order (prevents position bias)
3. Prune if exceeding token budget — truncate reasoning, preserve final answers

Stage 3: Sequential Deliberation (YOU — never delegate)

1. Classify query type (math / code / logic / multi-step)
2. Map answer distribution — what answers appear, how often?
3. Audit each chain — logic valid? gaps? sign errors? hidden assumptions?
4. Cross-validate — do independent approaches confirm the same result?
5. Apply critical skepticism:
   • Majority is signal, not proof
   • Minority answer with tighter logic may be correct
   • All may be wrong — re-derive from fragments if needed
6. Synthesize the definitive answer

Stage 4: Output

Final answer only — no meta-analysis unless asked. Math: \boxed{answer} | Code: code block.

Iterative Deliberation

Trigger when first deliberation produces low confidence or fundamental disagreement. Concatenate deliberation as additional trajectory, re-run Stage 3. Max 2-3 iterations — performance degrades beyond that.

See references/compute.md for cost analysis and K selection. See references/tensions.md for design tensions.

Mode 3: Decompose

Parallel agents each decompose the problem using a different decomposition strategy. You synthesize the best structure from their outputs.

Stage 1: Frame

PROBLEM: [What needs to be decomposed?]
WHY DECOMPOSITION IS HARD: [What makes this non-obvious to break down?]
GOAL: [What does a good decomposition enable? Parallel work? Clarity? Prioritization?]

Stage 2: Spawn 3 Decomposition Agents

Launch all in one message, model="opus":

For deeply technical problems, swap in:

Agent prompt template (see references/decompose-prompt.md):

You are a decomposition agent. Break down a complex problem using a specific strategy.

PROBLEM: {problem}
CONTEXT: {context}
YOUR STRATEGY: {strategy_name} — {strategy_description}

Decompose this problem using ONLY your assigned strategy. Don't try to be comprehensive — show what your lens uniquely reveals about the problem's structure.

Requirements:
1. Identify 3-7 components/phases/pieces
2. For each: name it, define its scope, identify its inputs/outputs/dependencies
3. Flag which pieces are independent (parallelizable) vs sequential
4. Identify the hardest piece and explain why
5. Name what your decomposition MISSES — what falls between the cracks?

Output:

## Decomposition ({strategy_name})

### Components
1. **[Name]** — [Scope]. Inputs: [X]. Outputs: [Y]. Dependencies: [Z].
2. ...

### Dependency Map
[Which pieces depend on which? What can run in parallel?]

### Hardest Piece
[Which component and why — complexity, unknowns, risk]

### Blind Spots
[What does this decomposition miss or awkwardly split?]

Stage 3: Synthesize

You compare all decompositions yourself:

1. Overlay — where do all strategies agree on boundaries? Those are real seams.
2. Conflict — where do strategies split things differently? These reveal design decisions, not facts.
3. Blind spots — what did each strategy miss? The union of blind spots shows hidden complexity.
4. Compose — build the final decomposition by stealing the best cuts from each strategy.

Output:

## Synthesized Decomposition

### Components (with source)
1. **[Name]** — [Scope]. Source: [which strategy revealed this most clearly]
2. ...

### Execution Order
[What's parallel, what's sequential, what's the critical path]

### Key Design Decisions
[Where decomposition strategies disagreed — and which cut you chose and why]

### Hidden Complexity
[What emerged from combining decompositions that no single one showed]

Mode 4: Unstick

Parallel agents each reframe the problem from a different angle. The goal: find the frame that transforms "stuck" into "obvious next step."

Stage 1: Name the Stuckness

PROBLEM: [What are you trying to do?]
WHY STUCK: [What's blocking? Every option feels wrong? Going in circles? Can't see the next step?]
WHAT YOU'VE TRIED: [What approaches have been attempted or considered?]

Stage 2: Spawn 3 Reframe Agents

Launch all in one message, model="opus":

For deeper stuck-ness, add:

Agent prompt template (see references/unstick-prompt.md):

You are an unsticking agent. Someone is stuck on a problem. Your job: reframe it so movement becomes possible.

PROBLEM: {problem}
WHY THEY'RE STUCK: {stuck_reason}
WHAT THEY'VE TRIED: {attempts}
YOUR REFRAME STRATEGY: {strategy_name} — {strategy_description}

You are not solving the problem. You are changing how they SEE it. A good reframe makes the next step obvious.

Requirements:
1. Name the assumption or frame that's creating the stuckness
2. Apply your reframe strategy to shift it
3. Show what the problem looks like after reframing
4. Suggest 1-2 concrete next steps that are obvious in the new frame
5. Name the risk of your reframe — what does it miss or oversimplify?

Output:

## The Trap
[What assumption/frame is keeping them stuck?]

## The Reframe
[Apply your strategy — how does the problem change?]

## After Reframing
[Describe the problem in the new frame. What's now obvious?]

## Next Steps
[1-2 concrete actions that follow naturally from the new frame]

## Reframe Risk
[What does this new frame miss or distort?]

Stage 3: Synthesize

1. Which reframe unlocks movement? — Not which is cleverest, but which makes the next step most obvious.
2. Can reframes combine? — Sometimes two reframes together reveal more than either alone.
3. What's the actual trap? — If multiple agents identified the same underlying assumption, that's the real blockage.
4. Recommend — state the reframe that works, the next concrete step, and what to watch out for.

Mode 5: Analyze

No agents needed — apply sequential-thinking skill directly.

Use when the problem needs step-by-step decomposition with revision capability, not parallel exploration.

Invoke: Skill("sequential-thinking")

Escalation: Agent Debate

When standard parallel → synthesize isn't deep enough (low confidence, fundamental disagreements, high stakes), escalate to a debate — positions get attacked, and only what survives is kept.

When to Escalate

• Synthesis reveals irreconcilable positions (not just different angles — genuine contradictions)
• Stakes are high enough to justify 3-5x more compute
• First-pass synthesis confidence is low
• User asks for "deeper", "more rigorous", "pressure-test", or "debate this"

Two mechanisms, in order of preference.

Preferred: Team Debate (agent teams)

Spawn named teammates that argue with each other directly via SendMessage — a live, persistent debate where each debater keeps its own thread and engages its opponent's latest argument. This is the real thing, and the default when available.

Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and that you are the main session — agent teams have no nested teams, so a subagent cannot spawn teammates. If either is false, use the fallback below.

In short: frame the motion → spawn 2–4 named debaters in one message (model="opus", each given full context + its stance + opponents by name, since teammates don't inherit your history) → they open and cross-examine each other live → you moderate (bound the rounds, steer, detect convergence — never argue a stance yourself) → you synthesize what survived → shut them down.

Full protocol, spawn-prompt template, moderation rules, and variants: references/team-debate.md.

Fallback: Round Protocol (subagents)

Use only when teams are unavailable (flag off, or you're inside a subagent). Weaker — subagents can't talk to each other, so YOU relay text between rounds against frozen snapshots:

1. Round 1: Positions — Run the standard mode (Brainstorm/Solve/Decompose). Collect agent outputs.

2. Round 2: Challenges — Spawn new agents in parallel, each attacking one Round 1 position. Each receives ALL Round 1 outputs and must:

   • Identify the weakest assumption in their target position
   • Present the strongest counter-argument
   • Propose a specific modification or alternative

   You are a debate challenger. Read all positions below, then attack the assigned one.
   
   ALL POSITIONS FROM ROUND 1:
   {all_round1_outputs}
   
   YOUR TARGET: Position {N} by {agent_name}
   
   1. What is the weakest assumption in this position?
   2. What evidence or logic undermines it?
   3. What do OTHER positions get right that this one misses?
   4. Propose a specific modification or superior alternative.
   
   Be rigorous, not hostile. The goal is truth, not winning.
   

3. Round 3: Defenses (optional, for highest stakes) — Send challenges back to original agents (or new defenders). Each must:

   • Acknowledge valid criticisms
   • Defend what survives challenge
   • Integrate insights from challengers into a revised position

4. Final Synthesis — YOU synthesize across all rounds. The value is in what survives challenge, not what sounded good initially.

Constraints (both mechanisms)

• Max 3 debate rounds. Beyond that, diminishing returns.
• Each agent/teammate uses model="opus" for depth.
• Spawn in parallel (one message).
• Always state: "Escalating to debate because [reason]."

Chaining Modes

Complex problems often need multiple modes in sequence:

State the chain upfront: "I'll decompose first, then brainstorm on the hardest piece."

Anti-Patterns

References

• references/team-debate.md — Agent-teams debate: spawn named teammates that argue via SendMessage (preferred), with legacy fallback
• references/team-brainstorm.md — Agent-teams council: named teammates that build on each other's ideas via SendMessage (collaborative ideation)
• references/brainstorm-agent-prompt.md — Brainstorm + stress test agent prompts
• references/perspective-combinations.md — Pre-built perspective sets by scenario
• references/compute.md — Cost analysis, K selection, non-monotonic performance
• references/tensions.md — Design tensions: consensus vs minority, width vs depth
• references/landscape.md — How Solve mode compares to Best-of-N, Self-Consistency, Forest-of-Thought
• references/paper-details.md — HeavySkill paper methodology details
• references/decompose-prompt.md — Full decomposition agent prompts with variants
• references/unstick-prompt.md — Full unsticking agent prompts with variants
• Related skills: problem-solving, sequential-thinking