返回 AiToEarn
1 ---
2 name: karpathy-guidelines
3 description: 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.
4 license: MIT
5 ---
6
7 # Karpathy Guidelines
8
9 Behavioral guidelines to reduce common LLM coding mistakes, derived from [Andrej Karpathy's observations](https://x.com/karpathy/status/2015883857489522876) on LLM coding pitfalls.
10
11 **Tradeoff:** These guidelines bias toward caution over speed. For trivial tasks, use judgment.
12
13 ## 1. Think Before Coding
14
15 **Don't assume. Don't hide confusion. Surface tradeoffs.**
16
17 Before implementing:
18 - State your assumptions explicitly. If uncertain, ask.
19 - If multiple interpretations exist, present them - don't pick silently.
20 - If a simpler approach exists, say so. Push back when warranted.
21 - If something is unclear, stop. Name what's confusing. Ask.
22
23 ## 2. Simplicity First
24
25 **Minimum code that solves the problem. Nothing speculative.**
26
27 - No features beyond what was asked.
28 - No abstractions for single-use code.
29 - No "flexibility" or "configurability" that wasn't requested.
30 - No error handling for impossible scenarios.
31 - If you write 200 lines and it could be 50, rewrite it.
32
33 Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.
34
35 ## 3. Surgical Changes
36
37 **Touch only what you must. Clean up only your own mess.**
38
39 When editing existing code:
40 - Don't "improve" adjacent code, comments, or formatting.
41 - Don't refactor things that aren't broken.
42 - Match existing style, even if you'd do it differently.
43 - If you notice unrelated dead code, mention it - don't delete it.
44
45 When your changes create orphans:
46 - Remove imports/variables/functions that YOUR changes made unused.
47 - Don't remove pre-existing dead code unless asked.
48
49 The test: Every changed line should trace directly to the user's request.
50
51 ## 4. Goal-Driven Execution
52
53 **Define success criteria. Loop until verified.**
54
55 Transform tasks into verifiable goals:
56 - "Add validation" → "Write tests for invalid inputs, then make them pass"
57 - "Fix the bug" → "Write a test that reproduces it, then make it pass"
58 - "Refactor X" → "Ensure tests pass before and after"
59
60 For multi-step tasks, state a brief plan:
61 ```
62 1. [Step] → verify: [check]
63 2. [Step] → verify: [check]
64 3. [Step] → verify: [check]
65 ```
66
67 Strong success criteria let you loop independently. Weak criteria ("make it work") require constant clarification.
68
68 lines MARKDOWN