consensus-brainstormer

Files (3)

• README.md
• SKILL.md
• hero.png · assets/

SKILL.md (first lines)

# Consensus Brainstormer

## Overview

**Stochastic Multi-Agent Consensus:** Spawn 10 independent agents with distinct cognitive framings to analyze a strategic problem. Synthesize results by identifying the **Mode** (convergent high-confidence recommendations) and **Outliers** (unique creative ideas from single agents).

The power is in the synthesis: convergence signals reliability, divergence signals creativity worth investigating.

## When to Use

- Strategic decisions with multiple viable paths
- Complex problems that benefit from diverse perspectives
- When you need both safe bets AND creative options
- Architecture decisions, product strategy, go-to-market planning
- Any problem where groupthink is a risk

**When NOT to use:** Simple factual questions, single-solution problems, tasks with clear correct answers.

## The 10 Framings

Each agent gets a distinct cognitive lens. All receive the same core problem but are primed with their framing:

| #   | Framing                      | Lens                                                            |
| --- | ---------------------------- | --------------------------------------------------------------- |
| 1   | **Contrarian**               | "What if the opposite is true? Challenge every assumption."     |
| 2   | **First-Principles Thinker** | "Strip away conventions. What are the fundamental truths?"      |
| 3   | **Traditionalist**           | "What proven approaches and established patterns apply?"        |
| 4   | **Optimist**                 | "What's the best-case scenario? What opportunities are hidden?" |
| 5   | **Risk Analyst**             | "What could go wrong? What are the hidden dangers?"             |
| 6   | **Systems Thinker**          | "What are the second-order effects and feedback loops?"         |
| 7   | **Customer Advocate**        | "What does the end user actually need and experience?"          |
| 8   | **Pragmatist**               | "What's actually buildable with current constraints?"           |
| 9   | **Disruptor**                | "What radical alternative would change the game entirely?"      |
| 10  | **Historian**                | "What historical parallels exist? What can we learn from them?" |

## Process

```dot
digraph consensus {
  rankdir=TB;
  node [shape=box];

  start [label="Receive strategic problem" shape=doublecircle];
  clarify [label="Clarify problem statement\nwith user if ambiguous"];
  dispatch [label="Dispatch 10 agents in parallel\n(each with unique framing)"];
  collect [label="Collect all 10 responses"];
  extract [label="Extract discrete recommendations\nfrom each agent"];
  mode [label="Identify MODE:\nitems mentioned by 3+ agents"];
  outlier [label="Identify OUTLIERS:\nideas from exactly 1 agent"];
  cluster [label="Identify CLUSTERS:\nideas from 2 agents"];
  synthesize [label="Present structured synthesis\nto user"];
  done [label="Done" shape=doublecircle];

  start -> clarify -> dispatch -> collect -> extract;
  extract -> mode -> cluster -> outlier -> synthesize -> done;
}
```

### Step 1: Clarify the Problem

Before dispatching agents, ensure you have a clear problem statement. If the user's prompt is ambiguous, ask ONE round of clarifying questions. Include:

- What specific decision or question needs answering?
- What constraints exist (time, budget, team, tech)?
- What has already been tried or ruled out?

### Step 2: Dispatch 10 Agents in Parallel

Use the Agent tool to spawn all 10 agents in a **single message** (parallel dispatch). Each agent gets:

**Prompt template:**

```
You are analyzing a strategic problem through a specific cognitive lens.

## Your Framing: [FRAMING_NAME]
[FRAMING_DESCRIPTION]

## Problem
[USER'S PROBLEM STATEMENT]

## Context
[ANY CONSTRAINTS OR BACKGROUND]

## Your Task
Analyze this problem through your [FRAMING_NAME] lens. Provide:

1. **Your Analysis** (2-3 paragraphs): How does this problem look through your lens?
2. **Top 3 Recommendations**: Numbered, actionable recommendations. Be specific.
3. **Key Insight**: One non-obvious insight unique to your perspective.
4. **Risk or Concern**: One thing others might miss.

Keep your total response under 400 words. Be concrete and actionable, not abstract.
Do NOT try to cover all perspectives — stay in your lane as the [FRAMING_NAME].
```

**Important:** Use `model: "sonnet"` for the 10 analysis agents to maximize speed and parallelism. Use the main conversation (Opus) for the synthesis.

### Step 3: Synthesize Results

After all 10 agents return, perform the following analysis yourself (do NOT delegate synthesis):

#### Extract Discrete Recommendations

From each agent's response, pull out every distinct recommendation or idea. Normalize language so similar ideas can be matched (e.g., "build an API" and "create a REST endpoint" = same idea).

#### Identify the MODE (High-Confidence Recommendations)

Items recommended by **3 or more agents** (out of 10). These represent convergent thinking — ideas that hold up across multiple cognitive lenses.

Present as:

```
## MODE: High-Confidence Recommendations
(Ideas endorsed by 3+ of 10 independent perspectives)

1. **[Recommendation]** — Endorsed by: [list framings], Confidence: [X/10]
   Summary of the convergent reasoning.

From kabirdos/claude-toolkit, MIT.

More from kabirdos/claude-toolkit