Skip to content

zhangyanxs/axiom

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Chinese / 中文 | English

Axiom — First-Principles Assumption Auditor

Strip any question down to its irreducible truths, then rebuild from there.

Axiom is an agentic skill that forces rigorous first-principles thinking on any decision or belief. Unlike generic "think deeper" prompts, Axiom introduces a structured 5-phase process with assumption classification (fact vs convention vs belief vs interest-driven), fragility × impact risk scoring, and anti-sycophancy rules that prevent the AI from simply agreeing with you.

Not framework fill-in-the-blank — this is assumption prosecution.

What Makes This Different

Existing tools Axiom
Template fill-in: "problem → assumptions → rebuild" Force-classifies each assumption's nature — different types need different challenge strategies
English-only examples (SpaceX, microservices) Bilingual: Chinese + English scenarios with culturally-specific assumption checklists
Outputs a list of assumptions with no prioritization Risk matrix: Fragility × Impact scoring → "Most Dangerous Assumptions Top 3"
AI tends to agree with user's original conclusion Anti-sycophancy rules: Hard constraints that prevent flattery and force genuine intellectual challenge

The 5-Phase Process

┌─────────────────────────────────────────────────────────────────┐
│  Phase 1: REFRAME                                               │
│  Confirm the question itself is correctly defined               │
├─────────────────────────────────────────────────────────────────┤
│  Phase 2: MINE                                                  │
│  Surface 8-12 hidden assumptions across 3 depth layers          │
├─────────────────────────────────────────────────────────────────┤
│  Phase 3: CLASSIFY                                              │
│  🔵 Physical Fact  🟡 Convention  🔴 Belief  ⚫ Interest-Driven │
├─────────────────────────────────────────────────────────────────┤
│  Phase 4: RANK                                                  │
│  Fragility (1-5) × Impact (1-5) → Top 3 Most Dangerous         │
├─────────────────────────────────────────────────────────────────┤
│  Phase 5: REBUILD                                               │
│  Reconstruct from verified premises → Show cognitive shift      │
└─────────────────────────────────────────────────────────────────┘

Installation

Pick your agent and run a single command. All methods clone the skill into the directory your tool expects.

Claude Code

# Project-level (recommended)
git clone https://github.com/zhangyanxs/axiom.git .claude/skills/axiom

# Or global (available in all projects)
git clone https://github.com/zhangyanxs/axiom.git ~/.claude/skills/axiom

Then in Claude Code, type /axiom or just say "help me think through this from first principles".

Cursor

# Project-level
git clone https://github.com/zhangyanxs/axiom.git .cursor/skills/axiom

Or use Settings → Rules → Add Rule → Remote Rule (GitHub) and paste the repo URL.

Gemini CLI / Antigravity

git clone https://github.com/zhangyanxs/axiom.git ~/.gemini/antigravity/skills/axiom

Then say Use @axiom to analyze my decision.

Codex CLI

git clone https://github.com/zhangyanxs/axiom.git .agents/skills/axiom

Kiro

git clone https://github.com/zhangyanxs/axiom.git ~/.kiro/skills/axiom

OpenCode / Other Agents

# Clone to any path your agent reads skills from
git clone https://github.com/zhangyanxs/axiom.git /your/skills/path/axiom

Manual (Any LLM)

Just copy the contents of SKILL.md into your AI's system prompt or custom instructions. No installation needed.


Trigger Phrases

Once installed, trigger axiom with natural language:

English:

first principles / break it down / question my assumptions / think from scratch / audit my reasoning / axiom

Chinese:

di-yi-xing-yuanli / bang-wo-xiang-qingchu / chai-jie / axiom

Example Output

User: "I think I should quit my job to start a company"

Axiom Phase 4 output (excerpt):

🔴 A4 (Fragility 4 × Impact 5 = 20)
   "My idea has enough market demand"
   → Do you have ANY paying users or letters of intent?

🔴 A2 (Fragility 3 × Impact 5 = 15)
   "I have the ability to build something independently"
   → What have you independently built? At what scale?

🟡 A5 (Fragility 4 × Impact 3 = 12)
   "If I fail, I can always start over"
   → What does "start over" actually mean given your age/debt/family?

File Structure

axiom/
├── SKILL.md                          # Main entry point
├── references/
│   ├── scenarios.md                  # 8 scenario checklists (4 CN + 4 EN)
│   └── assumption-types.md           # Classification handbook
├── examples/
│   ├── walkthrough-zh.md             # Full Chinese example (quitting to start a business)
│   └── walkthrough-en.md             # Full English example (dropping out for a startup)
├── README.md                         # English README
└── README_zh.md                      # Chinese README

Scenarios Covered

# Scenario
1 Career Decisions (job change, startup, promotion)
2 Business & Product Validation (startup, new feature, pivot)
3 Financial & Life Decisions (property, investing, major purchases)
4 Belief & Worldview Audit (values, methodology, life philosophy)

Each scenario includes 10-15 culturally-specific hidden assumptions and tailored probing questions.

Anti-Sycophancy Rules

Axiom includes hard constraints to prevent the AI from being a yes-man:

  • 🚫 No agreeing with the user's original conclusion during decomposition
  • 🚫 No "That's a great question" or similar flattery openers
  • 🚫 No identical reconstruction — if nothing shifted, the analysis wasn't deep enough
  • ✅ Must output at least one assumption the user won't want to hear
  • ✅ Must hold firm like a devil's advocate until the user provides real evidence

Compatibility

Works with any AI coding assistant or chat interface that supports SKILL.md or custom instructions:

  • Claude Code / Claude.ai
  • Gemini CLI / Antigravity
  • Cursor
  • Codex CLI
  • OpenCode
  • GitHub Copilot
  • Any LLM chat (paste SKILL.md as system prompt)

Contributing

Contributions welcome! Areas that would be especially valuable:

  • Additional scenario checklists (e.g., parenting decisions, health choices, academic career)
  • More walkthrough examples in different languages
  • Edge case documentation for the classification system
  • Integration guides for specific AI tools

License

MIT


Inspired by: Elon Musk's first-principles reasoning, Charlie Munger's mental models, and the observation that most "first-principles" tools stop at listing assumptions without actually auditing them.

About

Deconstruct problems via First Principles. Classify hidden assumptions (Fact/Norm/Belief/Interest) and rank the "Most Dangerous" using a Vulnerability×Impact matrix. Rebuild conclusions from verified premises. Best for major decisions, business validation & belief questioning. Triggers: axiom, first principles, deconstruct, underlying logic.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors