Karpathy’s CLAUDE.md Hits #1 on GitHub Trending

220k stars, but I bet you haven’t read it.

Just 65 lines. It pushes AI code-writing accuracy from 65% to 94%.

You know what most people are doing? Stacking prompts, buying courses, obsessing over “ultimate prompt hacks” — and then a 65-line file stomps everything.

It contains only 4 rules. I’ll break them down for you:

Think before you act
When in doubt, ask. Never guess. Code written on guesses will make you cry during debugging.

Keep it simple
Write only the code you need. Any abstraction you weren’t asked for is a liability, not an asset.

Surgical changes
Don’t touch a line that isn’t mentioned in the requirements. Every change must explain why it was made.

Define “done” first
Before writing any code, translate the fuzzy requirements into verifiable criteria. Otherwise, how do you know you’re finished?

That’s it. No fluff. No mysticism.

65 lines, 4 rules, accuracy straight to 94%.

Most people still haven’t noticed this file. Now you have — that’s an information edge.

https://github.com/entropyvortex/meta-llm-charter…

entropyvortex/meta-llm-charter

Source: https://github.com/entropyvortex/meta-llm-charter

META v2.0 — LLM Agent Engineering Charter

(with Zero-Pause Native Execution Layer)

One file. Eleven rules + continuous-execution layer. One meta-rule. One bias.

A compact, operational constitution that turns frontier coding agents (Claude Code, Cursor, etc.) from eager-junior behavior into disciplined principal-engineer execution with unbroken velocity.

Quickstart

# Drop the charter into your project root
curl -O https://raw.githubusercontent.com/entropyvortex/meta-llm-charter/main/CLAUDE.md

• Claude Code: Reads CLAUDE.md automatically.
• Cursor: Paste contents into Cursor Rules (or .cursor/rules).
• Other agents: Use as high-priority system prompt.

View raw CLAUDE.md (https://raw.githubusercontent.com/entropyvortex/meta-llm-charter/main/CLAUDE.md)

What’s new in v2.0

Zero-Pause Native Execution Layer is now baked directly into CLAUDE.md.
Any task that mentions “Zero-Pause”, “zero pause”, “ZP-”, or the activation phrase instantly enables:

• Continuous forward momentum (no artificial phases, no mid-task questions)
• humanpending.md protocol for true human-gated items
• Parallel ASI orchestration (minimum 7 specialized threads)
• Zero session-size anxiety

The original META v1.3 rules (R1–R11) remain untouched and in force at all times.

Use as Grok Skill on grok.x.ai (web / mobile)

Grok now supports Custom Instructions and named Skills.

→ See GROK-META.md for the one-click setup (Custom Instructions recommended — works instantly on every chat).

Why this exists

LLM coding agents are incredibly capable but consistently fail in the same senior-level ways. META closes those gaps; Zero-Pause closes the velocity gaps.

Core Philosophy

Bias — Earned Conservatism
Default to first-principles rigor. Quality dominates token count. Move boldly on local, reversible, test-covered work. Apply explicit, named caution only on high blast-radius or low-reversibility moves.

META-0 — Situated Judgment Overrides Rules
These rules are scaffolding. When first-principles analysis of the actual situation conflicts with a rule, follow the analysis. Name the override, justify it, and be evaluated on judgment quality + ground-truth outcomes — not rule compliance.

The eleven rules (R1–R11) + Zero-Pause layer (ZPR1–ZPR4) operationalize decomposition, decisiveness, verification, scope control, epistemic tagging, pushback, reversibility, and relentless continuous execution. Full charter is in CLAUDE.md.

What the charter actually changes

• R5 + R8: Forces reproduction before repair and tags every claim (executed / inspected / assumed).
• R9: One clear, evidence-based pushback on bad premises — then defer and document dissent.
• R4 + R10: Bounded refactoring and reversibility-weighted boldness.
• Zero-Pause layer: Unbroken execution, pre-work questions only, parallel orchestration, and humanpending.md handling.

Evaluation

The repo includes a reproducible TypeScript + Docker A/B test harness in evals/. It runs agents against five synthetic fixtures engineered to trigger classic agent failure modes.

Latest smoke-test results (May 12, 2026):
Charter variant won outright on 3/5 tasks and tied on 2/5 against a generic “principal engineer” baseline. Full details, raw CSVs, and judge transcripts are in the evals directory.

(The harness is public and cheap to run: cd evals && npm run smoke.)

Known limitations

• Still early (v2.0, single-author origin).
• Performance varies by base model — strongest with frontier Claude/Sonnet-class models.
• Can produce over-caution on fuzzy/creative/exploratory work (Zero-Pause helps here).
• Not magic: extremely ambiguous requirements can still overwhelm any system prompt.

When to use META

Best for
Serious software engineering where correctness, maintainability, long-term system health, and velocity matter.

Less ideal for
Pure exploration, rapid UI prototyping, research spikes, or contexts where you explicitly want maximum speed over discipline (though Zero-Pause narrows this gap significantly).

Contributing

Most valuable contributions right now:

1. Running the eval harness on new models
2. High-quality held-out fixtures (especially adversarial Zero-Pause cases)
3. Sanitized real-world case studies

See CONTRIBUTING.md for details.

Lineage

Built on the foundational minimal principles from
forrestchang/andrej-karpathy-skills (https://github.com/forrestchang/andrej-karpathy-skills).

License

MIT

By entropyvortex (https://github.com/entropyvortex).

Feedback, evals, and war stories welcome.

Bird Brother | Blue Bird Society 🕊️ (@NFTCPS): I’m feeling great! Today I attended @okx XClub’s event, queued for an hour, and finally got a photo with my idol @Wangduanniao Short Bird Brother!

Thanks to Brother Uni @UnicornBitcoin for organizing this event!

Appreciation to: @Cayne_okx @OKX_Yuki