karpathy-guidelines

To enforce surgical changes during code refactoring, apply a four-step plan-driven workflow that explicitly states assumptions, scopes tasks minimally, and edits only necessary code to avoid collateral impact. This reduces common LLM overcomplication.

Surfacing hidden assumptions before writing code requires explicitly stating all expectations and tradeoffs upfront, asking clarifying questions when unclear, and converting tasks into measurable goals with clear success criteria to prevent silent misinterpretations.

Keeping LLM-generated code simple requires prioritizing well-scoped solutions with no unnecessary abstractions, ensuring tasks remain verifiable and minimal, and applying surgical edits that touch only the code necessary to satisfy the objective.

Plan-driven coding helps multi-step refactoring by converting tasks into measurable goals with clear success criteria and an explicit plan, ensuring execution remains scoped, verifiable, and minimal throughout the entire process.

LLM coding often results in overcomplicated solutions due to silent misinterpretations and unscoped execution. Applying coding discipline that prioritizes simplicity, explicitly surfaces assumptions, and restricts edits to surgical changes prevents these missteps.

You should not use an autonomous approach for code refactoring when tasks lack clear success criteria, have unstated assumptions, or require multi-step execution, as unscoped changes risk collateral impact and unnecessary abstractions.