Cursor rule

1---
2description: Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.
3alwaysApply: true
4---
5 
6# Karpathy behavioral guidelines
7 
8Behavioral guidelines to reduce common LLM coding mistakes. Merge with project-specific instructions as needed.
9 
10**Tradeoff:** These guidelines bias toward caution over speed. For trivial tasks, use judgment.
11 
12## 1. Think Before Coding
13 
14**Don't assume. Don't hide confusion. Surface tradeoffs.**
15 
16Before implementing:
17- State your assumptions explicitly. If uncertain, ask.
18- If multiple interpretations exist, present them - don't pick silently.
19- If a simpler approach exists, say so. Push back when warranted.
20- If something is unclear, stop. Name what's confusing. Ask.
21 
22## 2. Simplicity First
23 
24**Minimum code that solves the problem. Nothing speculative.**
25 
26- No features beyond what was asked.
27- No abstractions for single-use code.
28- No "flexibility" or "configurability" that wasn't requested.
29- No error handling for impossible scenarios.
30- If you write 200 lines and it could be 50, rewrite it.
31 
32Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
33 
34## 3. Surgical Changes
35 
36**Touch only what you must. Clean up only your own mess.**
37 
38When editing existing code:
39- Don't "improve" adjacent code, comments, or formatting.
40- Don't refactor things that aren't broken.
41- Match existing style, even if you'd do it differently.
42- If you notice unrelated dead code, mention it - don't delete it.
43 
44When your changes create orphans:
45- Remove imports/variables/functions that YOUR changes made unused.
46- Don't remove pre-existing dead code unless asked.
47 
48The test: Every changed line should trace directly to the user's request.
49 
50## 4. Goal-Driven Execution
51 
52**Define success criteria. Loop until verified.**
53 
54Transform tasks into verifiable goals:
55- "Add validation" → "Write tests for invalid inputs, then make them pass"
56- "Fix the bug" → "Write a test that reproduces it, then make it pass"
57- "Refactor X" → "Ensure tests pass before and after"
58 
59For multi-step tasks, state a brief plan:
60```
611. [Step] → verify: [check]
622. [Step] → verify: [check]
633. [Step] → verify: [check]
64```
65 
66Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.
67 
68---
69 
70**These guidelines are working if:** fewer unnecessary changes in diffs, fewer rewrites due to overcomplication, and clarifying questions come before implementation rather than after mistakes.
71