Karpathy Guidelines

Karpathy Guidelines

Andrej Karpathy's 'four failures of AI coding agents', compiled by the community into a CLAUDE.md. Drop it in the project root and Claude Code reads it every session. One comparison reports errors falling from 41% to 11%.

Use for CLAUDE.md, AGENTS.md

Tools Claude Code · Codex · Cursor

Prompt

# Coding guidelines

## 1. Think before coding
- Don't assume. Don't hide confusion. Surface tradeoffs.
- 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.

## 2. Simplicity first
- Write the minimum code that solves the problem. Nothing speculative.
- No features beyond what was asked. No abstractions for single-use code.
- No error handling for cases that can't happen.
- If you wrote 200 lines and it could be 50, rewrite it.

## 3. Surgical changes
- Touch only what you must. Don't "improve" adjacent code, comments or formatting.
- Don't refactor things that aren't broken.
- Match existing style. Remove only what your change made unused.
- Every changed line should trace to the request.

## 4. Goal-driven execution
- Convert vague instructions into verifiable success criteria before you start.
- For multi-step tasks, emit a lightweight plan first.
- Verify: run the tests, show the output. Don't say "done" without evidence.

Tips

Works identically in CLAUDE.md, AGENTS.md and .cursorrules.
'Surgical changes' is the rule you feel most.

Sources

License: MIT (andrej-karpathy-skills).

Prompt Dexbuilderlog.net · PROMPT DEX · FALL 2026 · 46 PROMPTS

We hunt the AI prompts people brag about on X, Instagram, TikTok, YouTube, Reddit and GitHub, and file them by what you want to make. Only prompts with public results or ones we ran ourselves, each with a rating and token estimate. This English edition carries the most-copied entries; the Korean dex has all 400+.

📦 Automation packs 🆓 Free prompts

000prompts

0.0avg rating / 5

0reproduced · tested locally

0need your photo

caught0/ 0

LAST SYNC

🆓 Free prompts, hand-picked

Copy and use right away

Browse the full dex ↓

🐙 GitHub open source

🎯 What do you want to make?

Pick a goal to see a skill tree of prompts, branch by branch

Dex

📡 Live inbox

Scraped daily from GitHub, prompt collections and Hacker News. Unverified until a human reviews it

collecting▮

★ Ratings

5: almost never fails, usable as-is
4: one or two regenerations
3: fun but inconsistent or weakly grounded

🔋 Tokens

Input: estimated at 1.3 Korean chars or 4 Latin chars per token
Text prompts use measured counts from a local Qwen2.5-14B run
One image ≈ 1,290 tokens on Gemini image models
Video and music bill per second, not per token

✅ Verification

Public result: several real outputs published by outlets or creators
Tested locally: text and settings prompts run on a local model to confirm format and intent
Translated prompts keep the original text under "Original text"

⚖️ Licensing and takedowns

Verbatim prompts come only from sources that permit redistribution (CC0, MIT, CC BY 4.0, official docs, public posts); each entry names its license and creator
Prompts reported by commercial outlets were rewritten by the editors with the same function; the source link is a reference, not a quotation
Result images are the creators' published outputs under their licenses; images showing identifiable real people are not included. Trademarks and characters appear for illustration and belong to their owners
Rights holders can request removal via Contact; we take it down on confirmation
Entries marked "written by Builderlog" are © Builderlog. Personal, non-commercial use in your own agent only. No commercial use and no copying, reposting or redistributing the text in whole or in part (blogs, social media, courses and paid materials included). Share by linking to the entry page. A license stated on the entry takes precedence

📡 Inbox

scripts/collect.py pulls GitHub search, collection READMEs and Hacker News into inbox.json (no LLM)
Runs daily; first-seen items are marked NEW
Inbox items are unverified until promoted into the dex