Reusable coding-agent guidelines for explicit assumptions, simple implementations, scoped edits and verifiable goals.
Project overview
It provides project-mergeable instructions for common coding-agent pitfalls, with a caution-versus-speed tradeoff.
Project type
AI Coding · Prompt Engineering
Use cases
Coding & Development
Deployment
Refer to project documentation
License
License pending
Best for
Users evaluating reusable instructions for an existing coding agent’s assumptions, change scope and verification.
Key capabilities
Provides instructions to state assumptions, present alternative interpretations, push back when appropriate, and request clarification when confused.
Provides instructions to implement only the minimum requested code and avoid speculative features, unnecessary abstractions, and overengineering.
Provides instructions to limit edits to requested work, avoid unrelated refactoring, preserve existing style, and clean up only artifacts created by the current changes.
Provides instructions to turn coding tasks into explicit success criteria and verify work through tests or other checks.
Supports combining the reusable guidelines with an existing or newly created CLAUDE.md containing project-specific instructions.
Limitations and risks
The guidelines deliberately favor caution over speed.
Applying the full guideline rigor to trivial tasks may unnecessarily slow simple changes.
Getting started
Add the forrestchang/andrej-karpathy-skills marketplace from within Claude Code; Install andrej-karpathy-skills@karpathy-skills
Evidence and sources
GitHub project description: A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.
README: This installs the guidelines as a Claude Code plugin, making the skill available across all your projects.
README: "The models make wrong assumptions on your behalf and just run along with them without checking. They don't manage their confusion, don't seek clarifications, don't surface incons…
README: "They really like to overcomplicate code and APIs, bloat abstractions, don't clean up dead code... implement a bloated construction over 1000 lines when 100 would do."
README: "They still sometimes change/remove comments and code they don't sufficiently understand as side effects, even if orthogonal to the task."