ai-md

AI.MD v4 — The Complete AI-Native Conversion System

When to Use This Skill

• Use when your CLAUDE.md is long but AI still ignores your rules
• Use when token usage is too high from verbose system instructions
• Use when you want to optimize any LLM system prompt for compliance
• Use when migrating rules between AI tools (Claude, Codex, Gemini, Grok)

What Is AI.MD?

AI.MD is a methodology for converting human-written CLAUDE.md (or any LLM system instructions) into a structured-label format that AI models follow more reliably, using fewer tokens.

The paradox we proved: Adding more rules in natural language DECREASES compliance. Converting the same rules to structured format RESTORES and EXCEEDS it.

Human prose (6 rules, 1 line)  → AI follows 4 of them
Structured labels (6 rules, 6 lines) → AI follows all 6
Same content. Different format. Different results.

---

Why It Works: How LLMs Actually Process Instructions

LLMs don't "read" — they attend. Understanding this changes everything.

Mechanism 1: Attention Splitting

When multiple rules share one line, the model's attention distributes across all tokens equally. Each rule gets a fraction of the attention weight. Some rules get lost.

When each rule has its own line, the model processes it as a distinct unit. Full attention weight on each rule.

# ONE LINE = attention splits 5 ways (some rules drop to near-zero weight)
EVIDENCE: no-fabricate no-guess | 禁用詞:應該是/可能是 → 先拿數據 | Read/Grep→行號 curl→數據 | "好像"/"覺得"→自己先跑test | guess=shame-wall

# FIVE LINES = each rule gets full attention
EVIDENCE:
  core: no-fabricate | no-guess | unsure=say-so
  banned: 應該是/可能是/感覺是/推測 → 先拿數據
  proof: all-claims-need(data/line#/source) | Read/Grep→行號 | curl→數據
  hear-doubt: "好像"/"覺得" → self-test(curl/benchmark) → 禁反問user
  violation: guess → shame-wall

Mechanism 2: Zero-Inference Labels

Natural language forces the model to INFER meaning from context. Labels DECLARE meaning explicitly. No inference needed = no misinterpretation.