karpathy-guidelines

To stop LLMs from making unnecessary code changes during code refactoring, apply surgical change guidelines that enforce editing only necessary lines and avoiding unrelated code. This reduces common LLM coding mistakes by enforcing caution and clarity.

The best coding guidelines to reduce AI agent prompt mistakes require explicit assumptions, limit features to avoid unnecessary abstractions, and transform tasks into verifiable goals with clear success criteria for goal-driven execution.

Review Python code generated by an LLM by applying simplicity-first guidelines that suppress unnecessary abstractions and enforce minimal changes. Require explicit assumptions and a clear verification plan as part of the review workflow.

Yes, enforcing surgical changes works for debugging sessions in Python notebooks. The guidelines apply to reviewing notebooks by editing only what is necessary, avoiding unrelated code, and maintaining a clear verification plan.

You should require an LLM to think before coding during development whenever tasks are unclear or involve tradeoffs. The guidelines mandate surfacing tradeoffs, stating explicit assumptions, and asking questions before generating code.