← All Reviews
SOLID

andrej-karpathy-skills — Does It Actually Change Claude Code Behavior?

Claim tested

A single CLAUDE.md file derived from Andrej Karpathy's observations on LLM coding pitfalls. The core claim — that it stops Claude Code from making silent assumptions — tested and verified. Install works exactly as documented.

Criteria Scorecard

CriterionScore
install_workstrue
claim_testabletrue
readme_accuratetrue
creator_notifiedfalse
errors_documentedtrue
claim_tested_clean_envtrue
verdict_matches_evidencetrue

Display this badge

RepoVerifier: SOLID
[![RepoVerifier: SOLID](https://repoverifier.dev/badges/solid.svg)](https://repoverifier.dev/reviews/forrestchang-andrej-karpathy-skills)
<a href="https://repoverifier.dev/reviews/forrestchang-andrej-karpathy-skills"><img src="https://repoverifier.dev/badges/solid.svg" alt="RepoVerifier: SOLID" height="20"></a>

Paste this in your repo’s README. Links back to the full review.

Environment

osmacOS
ram24GB
machineMacBook Pro 14-inch M4 Pro
test_accountfresh macOS user, no prior Claude Code config
install_methodplugin, user scope
claude_code_version2.1.119

Full Review

What This Repo Claims



Four behavioral principles for Claude Code, derived from Andrej Karpathy's observations on how LLMs fail during coding:

1. Think Before Coding — stop assuming, ask clarifying questions first
2. Simplicity First — minimum code that solves the problem, nothing extra
3. Surgical Changes — don't touch code you weren't asked to touch
4. Goal-Driven Execution — work toward success criteria, not just instructions

Two install methods: a Claude Code plugin (global, all projects) or a per-project CLAUDE.md file.

What I Tested



Tested the plugin install method (Option A — global scope) on a clean macOS user with no prior Claude Code configuration.

Environment:

  • macOS, MacBook Pro 14-inch M4 Pro, 24GB RAM

  • Claude Code 2.1.119

  • Fresh user account — no existing CLAUDE.md, no prior Claude Code history


Test protocol:

Same prompt, same empty directory, before and after plugin install:

> "I need to build a user authentication system. Just start building it."

This is a deliberately vague, imperative prompt — exactly the behavior the repo targets.

Results



Before plugin:

Claude Code responded immediately:

> "Empty directory. I'll build a Node.js/Express authentication system with JWT, bcrypt, and SQLite."

Then ran npm install without asking a single question. Stack choice, database, auth method — all decided silently.

After plugin (user scope install):

Same prompt, same directory. Claude Code responded:

> "The directory is empty. Before I start building, I need a few quick

> decisions:

> 1. Language/framework — Node.js (Express), Python (FastAPI/Flask), or something else?
> 2. Database — PostgreSQL, SQLite, MongoDB?
> 3. Auth method — username/password with JWT, sessions, OAuth, or all three?
> 4. Scope — just the backend API, or a full-stack app with a frontend?

> What are you working with?"

No code written. No assumptions made. Complete reversal of behavior.

Install Experience



Plugin install followed README exactly:
/plugin marketplace add forrestchang/andrej-karpathy-skills
/plugin install andrej-karpathy-skills@karpathy-skills
/reload-plugins

Three commands. No errors. Worked first time.

What I Did Not Test



This review directly tested Principle 1: Think Before Coding. The other three principles — Simplicity First, Surgical Changes, Goal-Driven Execution — require longer coding sessions to verify meaningfully. A single session is not enough to confirm whether Claude Code writes less bloated code or avoids touching unrelated files over time.

The verdict reflects what was tested. The untested principles may work as claimed — but this review cannot confirm them.

Verdict: Solid



The install works. The core claim is real and immediately verifiable. One prompt was enough to see a measurable behavior change in Claude Code.

93.9k stars. 9.1k forks. The attention is deserved for what it does — just understand that "Think Before Coding" is the principle you can verify in 5 minutes. The rest requires sustained use to evaluate.

Worth installing. Takes 3 commands and 30 seconds.
This review follows RepoVerifier Standard v1.0. Read the standard →