AI.MD — Convert CLAUDE.md to AI-native structured-label format

[sstklen]

What it does

AI.MD converts human-written CLAUDE.md files into AI-native structured-label format.
Same rules, zero deletions — just restructured for how LLMs actually parse.

Results

Tested with a real production CLAUDE.md across 4 models:

Metric	Before (prose)	After (structured)	Change
File size	13,474 bytes	6,332 bytes	-53%
Line count	224 lines	142 lines	-37%
Codex compliance	6/8	8/8	✅

Claude Code recommends keeping CLAUDE.md under 200 lines. The original was 224 (over limit). After conversion: 142 (well within).

10 techniques used

1. One concept per line (split attention → full attention)
2. Explicit trigger/action/exception labels (zero inference)
3. Semantic anchors (labels as hashtags for direct matching)
4. State-machine gates (if → then → else → priority)
5. XML section isolation (<gates> <rules> <rhythm>)
6. Strip all "why" — keep only "what"
7. Bilingual labels (English for AI parsing, native language for output)
8. Cross-references instead of repetition
9. Not-triggered lists (explicit exclusions prevent over-firing)
10. Conflict detection with priority resolution

Install

mkdir -p ~/.claude/skills/ai-md && curl -o ~/.claude/skills/ai-md/SKILL.md https://raw.githubusercontent.com/sstklen/ai-md/main/SKILL.md

Then say "AI.MD" or "distill my CLAUDE.md" to Claude Code.

Links

• GitHub: https://github.com/sstklen/ai-md
• Includes before/after examples in examples/ directory
• Full methodology documented in SKILL.md (6-phase conversion process)

[aniruddhaadak80]

This is interesting because it is testing whether instruction structure is actually more machine-readable, not just shorter. The part I would want to see next is an ablation on real coding tasks, since compression alone can look impressive while hiding subtle losses in priority, exceptions, or edge-case behavior.nnIf you keep pushing it, side by side examples where the structured version improves one concrete agent behavior would make the case much stronger than byte reduction by itself.

[Nyrok]

The benchmark is compelling. 53% size reduction with full rule preservation proves that prose is inefficient as a medium for machine-readable instructions. The model was reconstructing structure from free text every time.

Same issue at the prompt layer. When you write role, constraints, and objective as a single prose blob, the model has to infer those boundaries. Explicit semantic blocks let it parse cleanly without guesswork. The compliance improvement you see with structured labels is that same effect.

I've been building flompt for exactly this, a visual prompt builder that decomposes prompts into 12 semantic blocks and compiles to Claude-optimized XML. Open-source: github.com/Nyrok/flompt