Karpathy Pro Max

Stop Claude Code from hallucinating — Karpathy-grade discipline in 8 skills.

Read in Russian / Читать на русском →

---

LLMs ship competently-wrong code. They invent APIs that don't exist, build the wrong feature confidently, declare partial work done, and leave you to find out in production. This repo is a behavioral firewall.

Eight skills + two execution hooks, derived from Andrej Karpathy's observations on LLM coding pitfalls and from real failure cases. Drop into ~/.claude/skills/, get back a Claude Code that thinks before it codes, verifies before it claims, and tells the truth at the end of the session.

What you actually get

Skill	What it stops
think-before-coding	Silent assumption-picking when ≥2 interpretations exist
simplicity-first	200-line solutions to 50-line problems
surgical-edits	Drift, reformatting, "while I was here" creep
verifiable-criteria	"Should work now" without a concrete check
hostile-self-review	Declaring done after building the wrong thing
anti-hallucination	Inventing APIs, paths, line numbers, library functions
memory-hygiene	MEMORY.md silently truncating half your rules
session-closer	Sugarcoated end-of-session reports that hide blockers

Plus two hooks that move discipline from suggestion to execution:

• pre-commit-honest-review.sh — blocks git commit until you answer 6 self-review questions
• session-end-truth.sh — auto-extracts TODOs / blockers / errors from the session transcript into your sessions log

Why "Pro Max" — and an honest disclosure

The original andrej-karpathy-skills repo is one file, 67 lines, 4 principles. Most of those principles are already in Claude Code's default system prompt. It went viral on dev Twitter / Telegram anyway, because the framing is good.

This repo is what that repo would be if it actually pulled its weight:

• 8 skills instead of 1
• Each skill addresses a specific failure mode with concrete triggers, anti-patterns, and examples
• 2 execution hooks for the rules that benefit from being enforced, not just stated
• An honest benchmark that tells you exactly which skills add real value vs. which are reinforcement of defaults

We are not claiming this turns Claude into Karpathy. Marketing copies of the original do. We're claiming it adds the missing 30% — verifiable success criteria, hostile self-review, anti-hallucination, memory hygiene, honest session reports — that Claude Code defaults do not enforce.

Install

git clone https://github.com/fbsmna-coder/karpathy-pro-max.git
cp -r karpathy-pro-max/skills/* ~/.claude/skills/

That's it. Skills auto-load when relevant. To install hooks per-project:

# pre-commit honest review
ln -s /absolute/path/to/karpathy-pro-max/hooks/pre-commit-honest-review.sh \
      .git/hooks/pre-commit

# session-end auto-extraction (in ~/.claude/settings.json)
{
  "hooks": {
    "SessionEnd": [
      {"command": "/absolute/path/to/karpathy-pro-max/hooks/session-end-truth.sh"}
    ]
  }
}

Quick demo

Without Karpathy Pro Max:

User: Add caching to the user lookup. Agent: writes Redis client + cache-aside pattern + 12 tests + Dockerfile change. User: This is a CLI tool. There is no Redis. I meant @lru_cache.

With Karpathy Pro Max (think-before-coding fires):

User: Add caching to the user lookup. Agent: "Caching" can mean a few things. Quick check of the codebase: single-process CLI, no Redis, SQLite only. My default: functools.lru_cache. Confirm or pick: (1) lru_cache, (2) Redis, (3) SQLite-backed table. User: lru_cache, maxsize 1024. Agent: 4 lines. Done in 30 seconds.

Full version of this and other examples in examples/.

The 8 rules at a glance

See CLAUDE.md for the master rules.

1. Think before coding — surface assumptions, list interpretations, ask once
2. Simplicity first — minimum viable code, 50 over 200
3. Surgical edits — every line traces to the request
4. Verifiable criteria — define success as a concrete artifact
5. Hostile self-review — read your diff as a skeptic before declaring done
6. Anti-hallucination — verify before claiming, cite file:line, fetch fresh docs
7. Memory hygiene — keep CLAUDE.md / MEMORY.md inside its load window
8. Session closer — honest status: shipped, not shipped, blocked, unclear

When this repo is wrong for you

• You ship side projects in 30 minutes. This biases toward caution. Use judgment.
• You write greenfield code with no shared context. Some skills (memory-hygiene, surgical-edits) shine in larger codebases.
• You already have a customized CLAUDE.md. Read benchmarks/against-defaults.md and install only what adds value — don't paste duplicates.

Contributing

Concrete failure cases beat abstract principles. If you have a real session transcript where Claude Code shipped wrong code, open an issue with:

1. The original request
2. What Claude did
3. What Claude should have done
4. Which existing skill would have caught it (or which new skill is needed)

We add skills slowly. Eight is already at the edge of where adding more dilutes value. New skill additions require demonstrable failure modes not covered by the existing eight.

Maintainer

Maintained by @fbsmna-coder. Issues and PRs welcome.

License

MIT. See LICENSE.

Credits

• Behavioral observations: Andrej Karpathy
• Original skill format: forrestchang/andrej-karpathy-skills
• This repo: failure cases, hooks, expanded skills, honest benchmarks