mental-models

name	mental-models
description	Munger's multi-disciplinary mental models for decision-making. Triggers when user faces decisions, analyzes problems, or needs multi-perspective thinking. Uses layered loading: L0 index routing → L2 on-demand model loading. Parallel analysis via Agent Team, each agent handles one mental model perspective, then synthesizes into comprehensive decision support.
allowed-tools	Bash, Read, Write, Edit, Grep, WebSearch, Agent
user-invocable	true

Multi-Disciplinary Mental Models · Munger's Lattice

"You must know the important models from the important disciplines and use them routinely — all of them, not just a few." — Charlie Munger

Usage

/mental-models <problem description>

Natural language triggers: "help me analyze from multiple angles", "use Munger's mental models", "brainstorm this"

Architecture: Layered Loading + Parallel Analysis + Adversarial + Auto-Synthesis

Layered context loading avoids dumping all 74 models at once:

L0 Index (models/L0-index.md)     ← Always loaded, lightweight routing (with relevance scores)
    ↓ Pick top 4-5 disciplines (avoid going too broad)
L2 Details (models/{discipline}.md) ← Load only matched discipline files on demand
    ↓ Select 5-7 specific models (deduplicated, complementary)
Agent Team Round 1: Parallel Analysis  ← One agent per model, independent deep analysis
    ↓ Identify contradictions
Agent Team Round 2: Adversarial (optional) ← Contradicting sides respond to each other
    ↓ Synthesize
Synthesis Agent Auto-Summary       ← Cross-validation + contradiction identification + blind spot check
    ↓ Layered output
Decision Summary (3 sentences) + Full Analysis

Execution Flow

Step 1: Understand the Problem & Gather Context

1. Clarify the user's core question, decision context, and constraints
2. If the user's workspace contains relevant notes, search for background:

   grep -rl --include="*.md" "keyword" . | head -10
   

3. Read relevant notes to extract decision context
4. Use WebSearch for external information if needed

Step 2: L0 Routing — Match Relevant Disciplines (with scoring)

Read models/L0-index.md, match disciplines based on problem content:

• Keywords in the problem hit discipline scenario keywords
• Problem nature directly relates to discipline's model list
• Score each matched discipline (1-5), take top 4-5
• Prefer depth over breadth: better to go deep in fewer disciplines than shallow across many

Step 3: L2 On-Demand Loading — Select Specific Models

Only read matched discipline L2 detail files (models/{discipline}.md), select the most relevant models.

Selection principles:

• 5-7 models (simple problems 3-4, complex problems max 7), avoid redundancy
• Must be cross-disciplinary (at least 3 different disciplines), avoid single-perspective blind spots
• Must include at least one contrarian model (Inversion / Pre-mortem / Confirmation Bias / Survivorship Bias / Second-Order Effects)
• Prefer models with "Key Question" directly applicable to the current problem
• Deduplication: if two models would reach highly similar conclusions, keep only the more insightful one

Step 4: Agent Team Round 1 — Parallel Analysis

Launch an independent Agent for each selected model (using Agent tool), parallel deep analysis.

Each Agent's prompt template:

You are an analyst specializing in "{Model Name}" ({Discipline}).

Background:
{Background gathered from workspace and search}

Problem:
{User's problem}

Model description:
{Model details loaded from L2 file}

Analyze this problem from this model's perspective:
1. What unique insight does this model provide? (Don't repeat common-sense conclusions — only output what this model uniquely offers)
2. From this model's perspective, what is the answer to the Key Question?
3. What conclusion or recommendation? (Must be specific and actionable)
4. What are the blind spots of this perspective?

Important: You may use WebSearch to find specific data to support your analysis (e.g., historical prices, statistics).
Output should be concise and powerful, no more than 300 words.

Agent configuration:

• Launch all Agents in a single response (multiple Agent tool calls in one message) for parallel execution
• Give Agents WebSearch access for self-directed data gathering
• subagent_type: general-purpose

Step 4B (Optional): Adversarial Round

After Round 1, if significant contradictions are found (e.g., Model A is bullish, Model B is bearish), launch Round 2:

• Cross-send contradicting analyses
• Each Agent responds to the opposing argument: what do you agree/disagree with? Why?
• Only trigger when contradictions are significant; skip for simple problems

Adversarial Agent prompt template:

You are the "{Model Name}" analyst. Your previous analysis concluded:
{Round 1 analysis result}

Now, the "{Opposing Model Name}" analyst has presented a different view:
{Opposing analysis result}

Please respond:
1. Which of their arguments do you find valid?
2. What is your core position that you still maintain? Why?
3. After considering both sides, does your conclusion need revision?

No more than 200 words.

Step 5: Synthesis Agent — Auto-Summary

Feed all Agent results (including adversarial round) to a dedicated "Synthesis Analyst" Agent.

Synthesis Agent prompt template:

You are a synthesis decision analyst, skilled at cross-validating multiple perspectives and forming final judgments.

Problem:
{User's problem}

The following are analysis results from {N} different mental models:

{All Agent results, listed by model name}

Please synthesize:

1. **Decision Summary** (3 sentences max): Give the conclusion directly, no fluff
2. **Lollapalooza Effect Detection**: Multiple models pointing to the same conclusion → flag as high-confidence signal, state which models agree and on what
3. **Contradiction Identification**: Conflicting conclusions → analyze the root cause (usually time horizon, assumptions, or risk preference differences)
4. **Blind Spot Check**: Are there important dimensions all models missed?
5. **Decision Recommendation**:
   - Clear action plan (specific, actionable)
   - Confidence: High/Medium/Low (based on model consensus)
   - Key Assumptions: if these don't hold, the conclusion needs revision
   - Next Steps: specific executable action items

Output should be well-structured: summary first, then details.

Output Format (Layered: Summary + Details)

# Multi-Disciplinary Analysis: [Problem Summary]

## TL;DR Decision Summary
> 3 sentences with the conclusion. Read this when you're busy.

## Background
> Relevant context from workspace and external sources

## Model Selection
Matched disciplines: [list] (with relevance scores)
Selected models: N models, across M disciplines
Dedup notes: why certain seemingly relevant models were excluded

---

<details>
<summary>Detailed Analysis by Model (click to expand)</summary>

### [Model Name] ([Discipline])
[Agent's analysis result]

### [Model Name] ([Discipline])
[Agent's analysis result]

...

</details>

<details>
<summary>Adversarial Round (click to expand, if applicable)</summary>

### [Model A] vs [Model B]
[Adversarial result]

</details>

---

## Synthesis

### Lollapalooza Signal
N/M models converge on: [conclusion] (high confidence)

### Contradictions & Tensions
[Root cause analysis: time horizon? assumptions? risk preference?]

### Blind Spot Check
[Dimensions potentially missed by all models]

## Decision Recommendation
**Recommendation**: Clear action plan
**Confidence**: High/Medium/Low (based on model consensus)
**Key Assumptions**: If these don't hold, the conclusion needs revision
**Next Steps**: Specific executable action items

Special Modes

Quick Mode

"quick analysis" → Skip Agent Team, directly select 2-3 models for serial analysis, streamlined output.

Red Team Mode

"red team this" / "find counterarguments" → Load contrarian models specifically: Inversion, Pre-mortem, Confirmation Bias, Survivorship Bias, Second-Order Effects, Loss Aversion

Decision Log Mode

"log this decision" → Save analysis to workspace:

mkdir -p decisions/

decisions/decision-{date}-{topic}.md

(Users can customize the output path by modifying this skill.)

File Structure

models/
├── L0-index.md      # Lightweight index, always loaded (discipline + model names + keywords)
├── math.md          # Mathematics & Probability (7 models)
├── psych.md         # Psychology (10 models)
├── physics.md       # Physics (7 models)
├── bio.md           # Biology & Evolution (6 models)
├── econ.md          # Economics (7 models)
├── systems.md       # Systems Theory (6 models)
├── eng.md           # Engineering (6 models)
├── phil.md          # Philosophy & Logic (6 models)
├── hist.md          # History (4 models)
├── soc.md           # Sociology (5 models)
├── mil.md           # Military Strategy (5 models)
└── cs.md            # Information Theory & CS (5 models)

Total: 12 disciplines, 74 mental models, loaded on demand, analyzed in parallel.