Skip to main content
Lab Grimoire
TW EN
☕ Coffee
← Spell Book

andrej-karpathy-skills

ai-tools

Spell Rating
🔮🔮🔮🔮○
Pricing
open-source
Difficulty
beginner

Best For

Any developer using AI coding agents (Claude Code, Cursor, Codex), regardless of experience level or project type. If you have ever had an AI silently make wrong assumptions, write 1000 lines when 100 would suffice, or casually refactor code it should not have touched, this config file was designed for you.

How I Actually Use It

Installed globally as a Claude Code Plugin with one command. No special invocation needed. The four principles automatically influence AI behavior in every session. In practice, I observed the AI more frequently listing its assumptions and asking about uncertain parts, with noticeably fewer unrequested refactoring attempts. For ambiguous requests, it now tends to say "I'm assuming X, let me know if that's wrong" instead of silently choosing a direction.

What It Does Well

The four principles are orthogonal, each solving an independent problem. "Think Before Coding" handles the input side, requiring the AI to state assumptions and ask when uncertain. "Simplicity First" handles output scale, prohibiting unasked features and unnecessary abstraction layers. "Surgical Changes" handles modification boundaries, where every changed line must trace directly to the user's request. "Goal-Driven Execution" handles the verification loop, converting vague instructions into verifiable steps with checkpoints.

EXAMPLES.md provides extensive before/after code comparisons. The deepest insight: LLM over-engineering is usually not about lacking technical ability but about "applying correct knowledge at the wrong time." It knows the strategy pattern, so it uses the strategy pattern for every simple calculation. The knowledge is right, the timing is wrong.

Failure Modes and When Not to Use

Fundamentally a prompt-level soft constraint. In long conversations or highly complex multi-step tasks, the AI may gradually "forget" these guidelines. No lint, hook, or test enforces compliance. May create unnecessary friction for trivial tasks, such as the AI listing assumptions before fixing an obvious bug. The author acknowledges: "These guidelines bias toward caution over speed. For trivial tasks, use judgment."

Pricing, Learning Curve, and Risk

MIT License, completely free. Near-zero installation barrier via Plugin marketplace or a single curl command. Risk is minimal: pure Markdown config, no remote calls, no data collection. The only consideration is priority conflicts with existing CLAUDE.md content. Merge by appending rather than replacing.

Verdict

Behind 158K stars is a real need. This 65-line config file will not turn AI into a perfect engineer, but it will make it commit fewer structural errors. As a foundational behavior layer for AI coding agents, it is worth installing globally. The honest assessment: "significant improvement" rather than "complete solution."

Sources