karpathy-guidelines

Avoid LLM coding mistakes by enforcing cautious reasoning, minimal implementation, and verifiable outcomes. This approach requires explicit assumptions, surgical changes, and goal-driven execution with tests to ensure modifications are correct and not overcomplicated.

Writing safer code with surgical changes requires modifying only what is necessary, aligning with existing style, and cleaning up only artifacts your changes introduce. This minimal implementation approach prevents speculative features and unnecessary refactors.

Enforce assumption checking by requiring the LLM to explicitly surface ambiguities, tradeoffs, and assumptions before coding. This cautious reasoning mechanism ensures the AI thinks through the plan and produces a minimal, test-backed, surgical patch instead of guessing.

Verify LLM generated code by converting work into testable success criteria and looping until verified. This goal-driven execution ensures bugs and refactors are validated against explicit success checks, confirming the code functions as intended.

Yes, you can use this approach for code review and debugging. It enforces cautious reasoning and verifiable outcomes during writing, reviewing, and refactoring, ensuring unsafe assumptions and overengineering are caught across the development lifecycle.