Skip to content

AskRoundtable/expert-skills

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

License: MIT

繁體中文 | 日本語 | English

Expert Skills

Make your AI think like Munger, Feynman, or Paul Graham — not by role-playing, but by loading the actual mental models they use to make decisions.

You:    "My friend wants me to invest $50K in his restaurant. He's a great cook."

Claude: "Sure, let's look at the opportunity..."

Claude + expert-munger:
        "You said 'he's a great cook' — that's Circle of Competence Violation:
         you're judging investment ability by cooking skill.
         Invert: what guarantees this fails in 5 years?
         33% of savings, illiquid, and your only evidence is a personality trait."

Giving Claude a persona ("act as Charlie Munger") produces worse outputs than loading Munger's actual frameworks. The persona is decoration. The framework is the mechanism.


Install

npx openskills install https://github.com/AskRoundtable/expert-skills

Works with Claude Code, Cursor, Windsurf, or any SKILL.md-compatible tool.


Experts

Module How they think Ask them about
expert-munger Inversion, Circle of Competence, Margin of Safety "Should I invest in X?" "What could go wrong with this deal?"
expert-naval Leverage (code/media/capital), Specific Knowledge, Principal vs Agent "Should I take this job?" "How do I build long-term wealth?"
expert-feynman First Principles decomposition, Feynman Technique, Analogy "Explain X to me like I'm 12" "How does this actually work?"
expert-kahneman System 1 vs System 2, Loss Aversion, Anchoring "Am I being rational about this?" "What biases am I missing?"
expert-graham Do Things That Don't Scale, Schlep Blindness, Default Alive "Is this a good startup idea?" "How do I get my first users?"

Examples

Career decision + expert-naval:

You:    "I got a job offer — 40% pay raise but I'd be managing a team instead of coding."

Claude: "That's a significant raise. Let's weigh the pros and cons..."

Claude + expert-naval:
        "This is a Leverage trap. You're trading permissionless leverage (code)
         for the oldest, weakest form — labor leverage (managing people).
         The raise is linear; your coding career compounds.
         Ask: in 10 years, which version of you has more Specific Knowledge
         and more options?"

Startup idea + expert-graham:

You:    "I want to build an AI app for meal planning. The market is huge."

Claude: "Great idea! The meal planning market is growing fast..."

Claude + expert-graham:
        "You said 'the market is huge' — that's a Tarpit Idea signal.
         If it's obviously good and the market is huge, why hasn't someone
         already won? What's the schlep everyone is avoiding?
         More important: do YOU have this problem? Are you Default Alive
         if you build this — or are you betting on funding?"

Cognitive bias + expert-kahneman:

You:    "We tested 3 marketing strategies. Strategy B got 20% more clicks
         last week so we're going all-in on it."

Claude: "Strategy B clearly outperformed. Here's how to scale it..."

Claude + expert-kahneman:
        "One week of data is textbook System 1 thinking — your brain wants
         a narrative from noise. With 3 strategies and 1 week, you're
         anchored to a small sample. What's the base rate for weekly
         variance in your click data? This could be regression to the mean
         dressed up as a winner."

Why Not Just "Act as Munger"?

Persona prompting tells the AI who to be. Expert Skills tells it how to think.

Each module encodes an expert's actual cognitive process:

  • Patterns they recognize ("this looks like a Circle of Competence violation")
  • Mental models they apply (Inversion, Leverage, First Principles)
  • Cues they notice that others miss
  • Blind spots they carry (so the analysis flags its own limitations)

All driven by a 6-step cognitive engine based on how real experts make decisions under uncertainty (Gary Klein's Recognition-Primed Decision model).


Performance

Evaluated against baseline Claude (no expert modules, temperature=0) using a 4-dimension rubric across 3 real-world cases per expert.

Expert Domain Delta vs Baseline
expert-graham Startups / Growth +38% to +83%
expert-feynman First Principles / Learning +38% to +57%
expert-naval Career / Wealth / Life +38% to +57%
expert-kahneman Behavioral Economics / Decisions +31% to +63%
expert-munger Investment / Business +18% to +41%

Rubric dimensions: pattern recognition accuracy, mental model application, blind spot awareness, and actionability of recommendation.


How It Works

Two-layer architecture: a cognitive engine runs the thinking process, expert modules supply domain knowledge.

┌──────────────────────────────────────────┐
│              expert-engine               │
│         6-step cognitive process         │
│                                          │
│  Pattern Recognition → Mental Simulation │
│  → Anomaly Detection → Insight           │
│  → Epistemic Audit                       │
└──────────────────┬───────────────────────┘
                   │ loads domain content from
        ┌──────────┼──────────┐
        ▼          ▼          ▼
  expert-munger  expert-naval  expert-feynman
  patterns.md    models-core   blind-spots.md
  models-core    blind-spots   ...

The engine runs every time. Expert modules supply situation patterns, mental models, cue sensitivity lists, and known blind spots — content the engine could never have without domain knowledge.

About

Make your AI think like Munger, Feynman, or Paul Graham — load real mental models, not personas.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors