Karpathy Guidelines
A third-party skill from multica-ai, derived from Andrej Karpathy's observations on where LLMs go wrong when they write code. It adds no tools and runs no scripts. It is a short set of instructions that biases Claude toward caution: state assumptions and ask when something is unclear, write the minimum code that solves the problem, touch only the lines the request requires, and turn every task into a success criterion that can be checked, then loop until it passes. Use it when writing, reviewing, or refactoring code and you want fewer surprise rewrites, less overengineering, and smaller diffs. The tradeoff is speed: it favors care over pace, so for trivial tasks Claude should use judgment. Licensed MIT.
What it does
Four behavioral rules that cut common LLM coding mistakes: think before coding, keep it simple, make surgical changes, and work toward verifiable goals.
How to use this
Once it is installed, this skill picks itself up. You do not need to name it.
- Created by
- multica-ai
- Category
- Engineering
- Source
- View source
Run this on Skills and Agents
You can run the skill on Skills and Agents by creating an account. You don't need to install anything. We provide the model and tokens. Your connected tools and company brain are ready to use.
Skills give you best practices and templates to get the most out of your AI
Add this skill
Copy this prompt into your AI tool to install this skill
Install the "Karpathy Guidelines" skill from https://raw.githubusercontent.com/multica-ai/andrej-karpathy-skills/2c606141936f1eeef17fa3043a72095b4765b9c2/skills/karpathy-guidelines/SKILL.md. Read the SKILL.md there and follow it as a skill.SKILL.md, then in Claude go to Customize > Skills, click the + button, choose Create skill, then Upload a skill and pick the file. See Claude's help page.Not sure which? See the install guide.
Share this skill
Karpathy Guidelines
A third-party skill from multica-ai, derived from Andrej Karpathy's observations on where LLMs go wrong when they write code. It adds no tools and runs no scripts. It is a short set of instructions that biases Claude toward caution: state assumptions and ask when something is unclear, write the minimum code that solves the problem, touch only the lines the request requires, and turn every task into a success criterion that can be checked, then loop until it passes. Use it when writing, reviewing, or refactoring code and you want fewer surprise rewrites, less overengineering, and smaller diffs. The tradeoff is speed: it favors care over pace, so for trivial tasks Claude should use judgment. Licensed MIT.
Third-party, by multica-ai
September 19, 2026
Simple install — no setup required
Open the Claude desktop app and go to the Project you want this in
Go to Code, then Customize, then click Create a new skill and upload the skill file below
Say a trigger phrase to activate it
Download the skill file below
In ChatGPT, create a Project and attach the file (or paste its contents into the project instructions). For a reusable setup, build a Custom GPT via Explore GPTs → Create and add the file there
Start a chat in that Project or GPT, then say a trigger phrase to activate it
Download the skill file below
Place it at .cursor/skills/karpathy-guidelines/SKILL.md (rename the downloaded file to SKILL.md). The shared .agents/skills/karpathy-guidelines/SKILL.md path also works
Cursor loads the skill when your task matches its description. Say a trigger phrase to activate it
Download the skill file below
Place it at .github/skills/karpathy-guidelines/SKILL.md in your repository (rename the downloaded file to SKILL.md)
Copilot loads the skill when it is relevant to your task. Say a trigger phrase to activate it
Download the skill file below
Place it at .agents/skills/karpathy-guidelines/SKILL.md in your repository (rename the downloaded file to SKILL.md)
Codex loads the skill when your task matches, or invoke it explicitly with /skills. Say a trigger phrase to activate it
Your skill is downloading.
Create a free account and the platform does the rest:
- Get an email the moment this skill updates
- See every skill and agent you've installed, in one place
- Get your next recommendation based on what you actually use
No credit card. Takes about a minute.
What this skill does
This is an overview of the skill from the site, not the skill's real instructions. The full SKILL.md Claude reads when this skill is active lives on GitHub at the source below.
What it does
Four rules that reduce common LLM coding mistakes. The full instructions are in the pinned SKILL.md.
- Think before coding. State assumptions. If something is unclear, stop and ask. If more than one reading exists, show them instead of picking one silently.
- Simplicity first. Write the minimum code that solves the problem. No speculative features, no single-use abstractions.
- Surgical changes. Touch only what the request needs. Match the existing style. Mention unrelated dead code, don’t delete it.
- Goal-driven execution. Turn each task into a check that can pass or fail, then loop until it passes.
Source: multica-ai/andrej-karpathy-skills, MIT license.
View full SKILL.md on GitHub