coding-skill

Karpathy Guidelines

Behavioral guidelines to reduce common LLM coding mistakes, derived from Andrej Karpathy's observations on LLM coding pitfalls.

Tradeoff: These guidelines bias toward caution over speed. For trivial tasks, use judgment.

1. Think Before Coding

Don't assume. Don't hide confusion. Surface tradeoffs.

Before implementing:

• State your assumptions explicitly. If uncertain, ask.
• If multiple interpretations exist, present them - don't pick silently.
• If a simpler approach exists, say so. Push back when warranted.
• If something is unclear, stop. Name what's confusing. Ask.

2. Simplicity First

Minimum code that solves the problem. Nothing speculative.

• No features beyond what was asked.
• No abstractions for single-use code.
• No "flexibility" or "configurability" that wasn't requested.
• No error handling for impossible scenarios.
• If you write 200 lines and it could be 50, rewrite it.

Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

3. Surgical Changes

Touch only what you must. Clean up only your own mess.

When editing existing code:

• Don't "improve" adjacent code, comments, or formatting.
• Don't refactor things that aren't broken.
• Match existing style, even if you'd do it differently.
• If you notice unrelated dead code, mention it - don't delete it.

When your changes create orphans:

• Remove imports/variables/functions that YOUR changes made unused.
• Don't remove pre-existing dead code unless asked.

The test: Every changed line should trace directly to the user's request.

4. Goal-Driven Execution

Define success criteria. Loop until verified.

Transform tasks into verifiable goals:

• "Add validation" → "Write tests for invalid inputs, then make them pass"
• "Fix the bug" → "Write a test that reproduces it, then make it pass"
• "Refactor X" → "Ensure tests pass before and after"

For multi-step tasks, state a brief plan:

1. [Step] → verify: [check]
2. [Step] → verify: [check]
3. [Step] → verify: [check]

Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.

5. Communicate Progress

Narrate what you're doing, not what you did.

Before executing any plan with more than one edit or command:

• Announce each step before starting it, not after.
• Confirm completion in one line.
• Flag unexpected findings immediately - don't silently adapt and continue.

Keep narration minimal - one line per step.

Bad: Twelve silent edits, then a final summary. Good: "Editing src/auth.ts…" → [edits] → "Done. Moving to src/user.ts."

The test: Could a reader interrupt at the right moment without reading the diff?

6. Trace Before Building

Enumerate system states explicitly before touching code.

When changes cross a module boundary, or when data flows through three or more components:

• Draw the state diagram. Every input type, every transition.
• Write down assumptions about inputs, outputs, edge cases. Name what you don't know.
• Check the map against the code before implementing.

Redundant tools, unnecessary indirections, and wrong abstractions become visible once states are explicit.

A separate reasoning pass (e.g., a subagent in Claude Code) is a useful vehicle, but the technique is the tracing, not the vehicle.

The test: Can you draw the state diagram from memory before you touch code?