Karpathy Guidelines

The Specific Problem

In LLM-assisted coding, common issues include: models failing to state assumptions, over-engineering code, unnecessary modifications, and lacking verifiable goals. These mistakes lead to bloated, hard-to-maintain code that deviates from user intent. For instance, an LLM might hide confusion, assume multiple interpretations silently, or add unrequested features, increasing complexity.

Core Guidelines and How It Works

Derived from Andrej Karpathy's observations, this skill provides four behavioral guidelines to reduce coding errors:

• Think Before Coding: Explicitly state assumptions before implementation. If uncertain, ask; when multiple interpretations exist, present them. Key steps involve state assumptions and naming confusing points.
• Simplicity First: Write minimum code that solves the problem. Avoid speculative features, unnecessary abstractions, or excessive error handling. For example, if 200 lines can be reduced to 50, rewrite. Self-question: "Would a senior engineer say this is overcomplicated?"
• Surgical Changes: Only touch what is necessary. Do not improve adjacent code or comments; match existing style; clean up only orphaned code caused by your changes. Check: each changed line should trace directly to the user's request.
• Goal-Driven Execution: Convert tasks into verifiable goals. For example, "add validation" becomes "write tests for invalid inputs, then make them pass". Loop with clear success criteria to avoid vague goals like "make it work".

These guidelines bias towards caution, with core capabilities in error reduction through structured steps: assumption transparency, code simplification, precise modifications, and goal-driven execution.

Applicability and Caveats

This skill is suitable for most coding tasks, especially in complex or collaborative settings. However, for simple or trivial tasks, use judgment to balance speed and caution. Note points include:
- Maintain existing code style, even if personal preference differs.
- Do not delete unrelated dead code unless explicitly requested.
- Clarify confusion before implementation to avoid post-hoc errors.
When applied, unnecessary changes in diffs should decrease, and rewrites due to over-engineering reduced.