A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
No releases, no versioned artifacts, no CI pipeline. The repo is fundamentally a single CLAUDE.md markdown file, making traditional software maturity metrics inapplicable. Created January 2026, last commit April 2026 — roughly three months of active development compressed into what is ultimately a configuration document. Alpha by default, though the content itself is well-reasoned and stable-feeling.
README is unusually thorough for what it is: four principles explained with motivating quotes from Karpathy, a table mapping principles to the problems they solve, worked transformation examples for Goal-Driven Execution, and clear install instructions covering both Claude Code plugin and per-project curl methods. A Chinese translation (README.zh.md) and a dedicated CURSOR.md for Cursor integration are included. No API docs needed — there's nothing to document beyond usage, and that's covered well.
No code. The repository is markdown. No dependencies, no tests, no linting, no CI. The CLAUDE.md content itself is well-structured prose with clear imperative rules, but this dimension is structurally penalized by the absence of any software artifact to assess.
Zero open issues, which is clean but uninformative for a text-only repo. A single maintainer, roughly one commit every few weeks during the active window. The ~3-week gap between last commit and appraisal date suggests the guidelines have stabilized rather than been abandoned. No PR history visible. Responsive enough given the simplicity of the artifact.
126,805 stars and 12,887 forks in under four months is a strong viral signal — this trajectory rivals major tool launches. The /plugin marketplace integration with Claude Code legitimizes it as an ecosystem artifact rather than a gist. Forks at ~10% of stars is a high conversion ratio for a config file, indicating developers are actively adapting it rather than just starring it.
Overall: 3.0/5
Category: AI Coding Agent Skills Known alternatives in vault: mattpocock--skills (4.1/5, same category); humanlayer--12-factor-agents (3.7/5, AI Agent Engineering Principles — adjacent) Differentiation: This repo is Claude Code-specific and operationally immediate: drop one file, change agent behavior now. mattpocock--skills appears more general-purpose across agents and tools; 12-factor-agents addresses architectural principles for building agents rather than prompting existing ones. The Karpathy provenance gives this unusual epistemic authority — it is framed as empirical observations from a practitioner known for LLM literacy, not aspirational best-practices. The four-principle taxonomy (Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution) is tighter and more actionable than most alternatives. Gap or crowd: Category currently has one repo (mattpocock--skills). Adding this moves coverage from thin to adequate and provides a direct comparison: general agent skills vs. Claude Code-specific behavior shaping. Not crowded.
Score: 5/5 Harvestable: The CLAUDE.md itself is the primary artifact — deploy verbatim. The Goal-Driven Execution pattern (transform imperative tasks into declarative goals with verification loops) is directly extractable as a meta-skill template. The Surgical Changes principle is a high-signal heuristic for any PAI hook that invokes code-editing agents. The four-principle framework can be adapted into a skill definition schema. Integration path: Trivial for Claude Code projects — curl CLAUDE.md into project root or install via /plugin command. The principles can also be incorporated into system prompt templates for non-Claude-Code agent invocations, or surfaced as a checklist skill that runs before any code-modification task is dispatched. Overlap with existing: mattpocock--skills covers adjacent territory (agent skill design); humanlayer--12-factor-agents overlaps on behavioral constraints for agents. Neither addresses Claude Code's specific failure modes (silent assumption-making, overcomplication, orthogonal edits) with this precision. Adoption cost: trivial — one file, zero dependencies, zero build steps.
The star count is anomalous for a single-file text repo and warrants scrutiny, but the fork ratio (10%) and the existence of plugin marketplace integration suggest genuine utility-driven adoption rather than pure viral curiosity. The Karpathy attribution functions as a strong prior: developers already trust the source's diagnosis of LLM pitfalls, reducing the persuasion cost for adoption. The four principles map cleanly onto the most common failure patterns observed in LLM coding agents, making this immediately useful for any PAI that relies on Claude Code for automation tasks. The lack of a license is a minor friction point for redistribution but irrelevant for personal vault use. Consider pairing with mattpocock--skills for a two-source view of agent skill design when the vault matures.