andrej-karpathy-skills

Prevent regressions during refactoring by enforcing a structured workflow that limits file modifications to the smallest scope and requires verifiable success checks like tests, typecheck, or build output to prove behavior remains intact.

Safe and verifiable code changes are achieved by converting requests into evidence-based success criteria, demanding observable pass/fail results from lint or build commands, and constraining edits to avoid unrelated cleanup that could break existing behavior.

Manage ambiguity in bugfixing by surfacing only materially impactful assumptions and asking for clarification when ambiguity blocks safe progress, ensuring the coding agent executes constrained, targeted edits rather than broad, speculative code changes.

Yes, debugging tasks work with this approach by applying a structured output contract that records changed files, commands run, exact pass/fail results, and remaining risks, ensuring strict validation through observable evidence and measurable success.

Avoid using automated refactoring workflows when a task requires large-scale architectural rewrites or unrelated cleanup, as this approach prioritizes minimal diffs and tightly bounded file modifications to strictly reduce regression risk.