Why a 65‑line Markdown file outshines Anthropic’s docs: 4 rules to stop AI coding mistakes

A single 65‑line Markdown file, the andrej-karpathy-skills project, has amassed 176 000 GitHub stars, surpassing Anthropic’s anthropics/skills repository, illustrating a shift in AI coding from "can it write?" to "can it avoid mistakes when unsupervised?".

Karpathy’s observation—LLM agents often make wrong assumptions, over‑abstract code, and modify unrelated sections—highlights the prevalent "乱写" failure mode. Developers who applied the CLAUDE.md rules reported a jump in AI coding task pass rate from 65 % to 94 %.

The file encodes four concrete principles:

Ask before acting: When a request is ambiguous, the model must first present trade‑offs and seek clarification, preventing it from silently filling in product assumptions.

Conciseness first: Limit generated code to the minimal necessary lines (e.g., 50 lines instead of 1 000), curbing the model’s tendency toward bloated architectures.

Surgical edits: Every line change must be traceable to an original requirement, reducing unintended side‑effects and enforcing a strict edit budget.

Goal‑driven execution: Transform vague bug‑fix commands into explicit tests—write a reproducing test, then make it pass—so success is objectively verifiable.

These rules address three systemic error classes: mis‑interpreting requirements, over‑engineering, and uncontrolled modification. By embedding the constraints directly in a short, auditable file, the model’s behavior becomes predictable and auditable.

Why 65 lines instead of 200? Experiments show that extending the rule set dilutes core constraints; the model’s attention window may miss peripheral rules, leading to lower code quality. The concise format aligns with the model’s effective context length, ensuring critical constraints are always read.

VoidLight00’s project demonstrates automated rule optimization: an eval.json test suite quantifies rule impact, and a loop of "modify rule → re‑run evaluation" retains changes only when they improve scores, turning manual prompt tuning into a regression‑tested workflow.

From a broader LLM engineering perspective, three layers emerge:

Prompt Engineering: Crafting instructions the model can understand.

Workflow Engineering: Structuring multi‑step processes (e.g., ReAct, reflection) to guide execution.

Agent Governance: Defining hard boundaries—rules files, sandboxes, approvals, audits—that constrain the model’s actions. CLAUDE.md exemplifies this third layer.

Thus, the real scarcity in AI‑assisted software development is not model intelligence but the design of robust, enforceable constraints that keep the model honest when no human is watching.