Andrej Karpathy's LLM Coding Skills

Single CLAUDE.md guardrails derived from Karpathy's public observations on LLM coding failures; repo multica-ai/andrej-karpathy-skills hit ~185K GitHub stars after viral 2026 threads, including a Spanish-language April thread framing discipline as the coding-agent bottleneck.

What it is

Community-encoded principles in one file (also Claude Code plugin andrej-karpathy-skills@karpathy-skills). Addresses silent wrong assumptions, overcomplicated abstractions, orthogonal edits, and weak success criteria. Source: github.com/multica-ai/andrej-karpathy-skills, 2026-06-01

As of 2026-07-01, the repo is on main at 2c606141936f1eeef17fa3043a72095b4765b9c2, with no GitHub release tags and no license field exposed by the GitHub API; GitHub reported 186,021 stars and 19,088 forks. Treat it as a high-signal instruction source, not a package with semantic version authority. Source: GitHub API; git ls-remote, 2026-07-01

Four principles

Principle Addresses
Think Before Coding Wrong assumptions, hidden confusion, missing tradeoffs
Simplicity First Bloat, speculative features, 200 lines when 50 suffice
Surgical Changes Drive-by refactors, touching unrelated code
Goal-Driven Execution Tests-first loops, verifiable checkpoints

Source: github.com/multica-ai/andrej-karpathy-skills, 2026-06-01

Karpathy insight quoted upstream: give success criteria and let the model loop, not imperative micromanagement. Source: github.com/multica-ai/andrej-karpathy-skills README, 2026-06-01

Spanish thread context

Nico Bilinkis thread (527 likes, 2,196 bookmarks): one CLAUDE.md, no CLI, no dependencies, paste at repo root. Stressed discipline over model capacity as the 2026 coding-agent bottleneck. Compared hype around harnesses and MCPs to a file that subtracted instead of added. The thread has no local visual media; the reviewed artifacts were the resolved repo README and the self-thread. Source: X/@nicobilinkis, 2026-04-14

Repo stars at enrichment: 163,925 (prior wiki entry cited ~3.7K from an earlier @sharbel signal); current source check: 186,021. Source: github.com/multica-ai/andrej-karpathy-skills, 2026-06-01; GitHub API, 2026-07-01

Sharbel launch artifact

The @sharbel bookmark is the earliest reviewed screenshot in this cluster. It shows the README's "Karpathy-Inspired Claude Code Guidelines" page and maps Karpathy's failure observations to four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The screenshot is useful because it makes the mapping visual and teachable; it does not change the local adoption decision. Source: X/@sharbel visual artifact, 2026-04-11

Relevance

Same pattern as Kevin's skills taxonomy: upstream expertise becomes executable config. Pairs with Claude Code Harness and superpowers brainstorming/TDD skills.

Local adoption boundary: do not import this as a competing universal rule file. Kevin already owns the same invariants through Agent Ethos, Code Taste, Browser Testing Skills, Security and Review Skills, and the global AGENTS contract: state assumptions, prefer simple code, make surgical edits, and define verifiable success criteria. The repo remains useful as external evidence and a compact teaching artifact; local behavior should stay in Kevin-owned skills and rules. Source: X/@meta_alchemist, 2026-06-09


Timeline