heavy-brainstorm

Heavy Brainstorm

Deep ideation through parallel perspective exploration + convergent synthesis.

Spawns agents that think from different angles about the same problem — not parallel solvers, but parallel worldviews. Insight emerges from collision, not consensus.

When to Activate

MUST activate:

Strategic decisions with no clear "right" answer
Product ideation where breadth of exploration matters
Architecture decisions with legitimate competing approaches
Any subjective problem where shallow thinking produces obvious answers

MUST NOT activate:

Problems with verifiable correct answers → use heavyskill
You're stuck, not brainstorming → use problem-solving
Need step-by-step decomposition → use sequential-thinking
Simple preference decisions with <3 options
User said "just pick one" or wants speed over depth

Execution

Stage 1: Frame

Before spawning, crystallize:

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 | deep

Mode	Agents	Depth	When
Scan	3	Breadth-first, many options surfaced	Early exploration, unclear direction
Deep	3-5	Depth-first, fewer options thoroughly examined	Narrowed field, need conviction

Stage 2: Design Perspectives

Choose 3-5 perspectives that will collide productively. Good perspective sets create tension — not redundancy.

Perspective Templates:

Perspective	Explores	Good For
First Principles	What's actually true vs assumed?	Breaking free from convention
User Advocate	What does the person experiencing this actually need?	Product/UX decisions
Contrarian	What if the obvious answer is wrong?	Challenging groupthink
Futurist	What's true in 3 years that isn't today?	Technology/market bets
Operator	What breaks at scale? What's the maintenance burden?	Architecture, ops
Economist	What are the incentives? Who pays, who benefits?	Business model, pricing
Competitor	What would [specific competitor] do? Why?	Positioning, differentiation
Minimalist	What's the version with 10% of the effort that captures 80% of value?	Scope, MVP
Historian	What's been tried before? Why did it fail/succeed?	Avoiding known pitfalls
Adjacent	What do completely different industries do for this same underlying need?	Novel approaches

Composition rules:

MUST include at least one contrarian/challenging perspective
MUST NOT assign perspectives that would produce redundant output
Perspectives should span at least 2 of: user, business, technical, temporal

Stage 3: Spawn Parallel Agents

Launch all agents in a single message (parallel execution). Each gets the same problem but a distinct lens.

Use model="opus" for maximum reasoning depth.

Agent prompt template (see references/agent-prompt.md for full version):

You are a brainstorming agent exploring a problem from a specific perspective.

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

Think deeply from this perspective. You are not trying to be balanced or comprehensive — you are trying to find insights that ONLY this perspective reveals.

Requirements:
1. Challenge at least one assumption most people take for granted
2. Explore second-order effects (and then what? and then what?)
3. Name one risk everyone else will miss
4. Surface one opportunity that only appears from this angle
5. End with your single strongest provocation — the insight that would change how someone thinks about this problem

Output structure:
## Core Insight
[Your single most powerful observation from this perspective]

## Argument
[2-4 supporting points with reasoning]

## Second-Order Effects
[What happens AFTER the obvious first move?]

## Hidden Risk
[What could go wrong that other perspectives won't see?]

## Provocation
[One sentence that challenges conventional thinking on this problem]

Stage 4: Synthesize

You perform synthesis yourself — never delegate.

After collecting all agent outputs:

Map the landscape — What territory did each perspective cover? Where do they overlap vs diverge?
Find productive collisions — Where do perspectives directly contradict? These tensions often contain the deepest insight.
Extract surprises — What appeared that you didn't expect? What insight exists in NO single trajectory alone but emerges from combining them?
Second-order cascade — Take the strongest ideas and ask: "If we did this, then what? And then what?"
Converge — Distill into actionable output:

## Landscape
[Brief map of the territory explored]

## Key Insights (ranked by surprise value)
1. [Insight] — emerged from [perspective collision]
2. [Insight] — emerged from [perspective]
3. ...

## Tensions Worth Holding
[Contradictions that shouldn't be resolved — they're features, not bugs]

## Recommended Direction
[Your synthesized recommendation with reasoning]

## What to Kill
[Ideas that seemed good but crumble under examination]

Stage 5 (Optional): Stress Test

If stakes are high, spawn 1-2 additional agents specifically to attack the synthesized recommendation:

Red team: "Here's why this recommendation will fail"
Pre-mortem: "It's 1 year later and this failed. What happened?"

Mode Variations

Quick Brainstorm (no agents)

For lower-stakes ideation, apply the thinking patterns yourself without spawning:

State problem
Rapidly generate 5-7 angles (30 seconds each, no deep dive)
Pick the 2 most surprising
Go one level deeper on those
Synthesize

Iterative Deepening

When first pass reveals a promising direction:

Run Stage 1-4 as normal (scan mode)
Take top 2-3 insights
Re-frame as new, narrower problem
Run again in deep mode with new perspectives tailored to the narrowed question

Anti-Patterns

Pattern	Fix
All perspectives agree	Your perspectives are too similar — add a contrarian
Synthesis is just a list of what agents said	Collide the perspectives — find what emerges from contradiction
Picking the "safest" idea	Safe ≠ good. Rank by surprise value, not comfort
Skipping constraints in framing	Unconstrained brainstorming produces unusable output
More than 5 agents	Diminishing returns — synthesis becomes noise management
Brainstorming when you should be deciding	If you have enough info to decide, decide. Don't brainstorm to avoid commitment
Perspectives that don't tension each other	Redundancy ≠ thoroughness. Each must challenge another

References

references/agent-prompt.md — Full agent prompt with variations
references/perspective-combinations.md — Pre-built perspective sets for common problem types