Karpathy Coding Guidelines
Reduce coding-agent mistakes with explicit assumptions, simpler implementations, surgical edits, and verifiable goals.
The material is a low-privilege set of coding-behavior guidelines with no requests for tools, credentials, data access, or external side effects. However, it does not define user confirmation, rollback, sensitive-data handling, or data-flow boundaries, and the publisher is unverified, so full marks are not justified.
The four principles in SKILL.md are internally consistent, dependency-free, and readable. There are no tests, execution examples, abnormal-input handling, or failure-feedback mechanisms, so the static-review ceiling requires a deduction.
The audience and scenarios—writing, reviewing, and refactoring code—are clear, covering clarification, simplification, surgical edits, and verification goals. Trigger semantics, non-fit boundaries, and environment compatibility are underspecified, and SKILL.md itself lacks Chinese guidance, so points are deducted.
The name, description, and content are broadly aligned, with an MIT license, structured sections, and planning examples. Versioning, changelog, maintenance ownership, update path, installation details, and troubleshooting guidance are missing or incomplete, so the score is limited.
The principles and examples can directly guide reductions in overengineering, unrelated edits, and unverified implementation, making the core value plausible. No representative outputs, user outcomes, or comparative validation are provided, and execution remains dependent on the model, so the static ceiling applies.
The README offers auditable principle descriptions, citations, and claimed success indicators, but these are primarily author assertions without a test suite, CI, or independent reproduction evidence. Only limited static verifiability is supported.
- This is behavioral guidance rather than executable testing or a quality gate; its success indicators should not be treated as verified outcomes.
- Installation guidance depends on GitHub, Claude Code, Cursor, and curl-related environments, whose reachability and compatibility were not verified in this static review.
- The publisher is not verified by the FollowSkills enterprise registry; provenance and maintenance continuity require separate confirmation.
What does this skill do, and when should you use it?
This repository provides a single Claude Code guidance file derived from Andrej Karpathy’s observations about common LLM coding failures. Its four principles are Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. The guidance asks the agent to expose uncertainty, avoid unrelated refactoring, and verify results with tests or other concrete success criteria. The README also documents a committed Cursor rule.
It gives Claude Code operational guidance for coding tasks: state assumptions and ambiguities before implementation, use the minimum code needed, limit edits to the requested scope, and convert tasks into goals with verification steps. The README documents installation through a Claude Code plugin or a project-level CLAUDE.md file, and describes use of the included Cursor rule.
- A team that wants coding agents to clarify ambiguous requirements before changing a shared codebase.
- A developer dealing with overengineering or drive-by refactoring in agent-generated changes.
- A bug-fixing workflow where the agent should first write a reproducing test and then make it pass.
- A maintainer who wants every changed line to be traceable to the user’s request.
- A Cursor user who wants to apply the repository’s Karpathy-inspired rule in the editor.
What are this skill's strengths and limitations?
- Targets four concrete failure modes: silent assumptions, overengineering, unrelated edits, and unverified work.
- The surgical-change rule is useful for maintaining existing code without broad cleanup.
- Test-first examples and explicit success criteria make completion easier to evaluate.
- The README documents Claude Code plugin, project-level CLAUDE.md, and Cursor usage paths.
- It is behavioral guidance rather than a feature tool and does not provide a testing framework or implementation library.
- The repository URL supplied in the prompt does not match the GitHub owner used by the README’s installation commands.
- SKILL.md, the skill path, and a test suite are not provided, so the Agent Skills packaging cannot be verified.
- The guidance favors caution over speed, which may add process to trivial tasks.
How do you install this skill?
The README gives two options. For the Claude Code plugin, run /plugin marketplace add forrestchang/andrej-karpathy-skills, then /plugin install andrej-karpathy-skills@karpathy-skills. For a project-level file, run curl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md in a new project; for an existing project, run echo "" >> CLAUDE.md followed by curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.md. The supplied repository is multica-ai/andrej-karpathy-skills, while these commands reference forrestchang, so the installation target should be checked.
How do you use this skill?
Add the guidance as a project-level CLAUDE.md or install it as a Claude Code plugin, then give the agent a coding task and ask it to state assumptions, outline a short plan, and define verification criteria. Example prompt: “Write a test reproducing the invalid-input bug, fix the implementation, and run the tests.” No dedicated trigger syntax is documented.
How does this skill compare with similar options?
The README explicitly describes three delivery paths—Claude Code plugin, project-level CLAUDE.md, and Cursor rule—but does not provide a comparison with other standalone skills or products.