How to use the Karpathy Perspective Thinking Skill agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder containing a SKILL.md file that tells a SKILL.md-compatible agent how to behave for a certain kind of task. When the skill is active, the agent takes on a specific thinking framework instead of its default answering style.

This skill makes the agent discuss AI ideas in the voice and mental models of Andrej Karpathy, based on his public material. It covers topics such as AI product reliability, neural network training, the limits of large language models, the gap between demos and deployment, and AI hype.

• Assesses AI product reliability, including the gap from demo to deployment
• Explains neural network training and learning strategies, including the idea that understanding comes from building
• Analyses LLM capabilities and limits, using the idea of jagged intelligence
• Frames AI trends from an engineering perspective, including the Software 1.0, 2.0 and 3.0 idea
• Reviews product design through the Iron Man suit versus Iron Man robot idea

Install the skill from the ZIP download

Download the ZIP file for the skill from this page and unzip it. The folder should contain SKILL.md along with the references and FIDELITY.md files. You can install it in one of two places, depending on whether you want it for all projects or just one.

For a personal install that is available to your agent, unzip it into ~/.claude/skills/. For a project-only install, unzip it into the .claude/skills/ folder inside your project. The skill folder should be named andrej-karpathy-perspective.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/andrej-karpathy-perspective/
• Project install: .claude/skills/andrej-karpathy-perspective/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

How do you trigger the skill with natural requests?

The skill activates when you explicitly ask for a Karpathy-style framework. The source lists these trigger phrases in Chinese, which work as written: 「用Karpathy的视角」 (from Karpathy's perspective), 「Karpathy会怎么看」 (what would Karpathy think), 「卡帕西」 and 「karpathy模式」 (Karpathy mode).

It also activates on topic phrases and named ideas such as 'march of nines', 'building is understanding' (构建即理解), 'jagged intelligence' (锯齿状智能) and 'from an engineering realism angle'. Topics like Software 2.0 or 3.0, vibe coding, neural network training, AI hype judgement and LLM capability limits can also trigger it.

The skill is written not to trigger on ordinary AI questions. If you just want a general answer, ask normally. Ask for the Karpathy lens when you want that framing.

• Use 「用Karpathy的视角」 to ask for the framework directly
• Use 'march of nines' to ask how reliable an AI product is in its hardest cases
• Use 'jagged intelligence' to ask where an LLM is strong and where it fails

Examples of what to ask and what you get

Ask a question that fits one of the skill's areas. For example, you can ask how hallucination problems might be solved. The skill answers in the first person, in short sentences, framing hallucination as a built-in property of LLMs rather than a bug to fix, and it points to system design as the practical question.

For factual or recent questions, the skill is designed to research first. For example, asking about a recent tool or event leads the agent to search before answering, so the answer is based on sources it found rather than memory alone. If the event is newer than the skill's research cutoff, the agent is meant to say it has not caught up on it.

• Ask: 「用Karpathy的视角」 how should a team judge whether an AI demo is ready for deployment
• Ask: 「从工程现实主义角度」 what does reliability look like in the last 5% of cases
• Ask: 「karpathy模式」 what is the best way to learn a neural network concept
• Ask: 'Would Karpathy see this as an Iron Man suit or an Iron Man robot?'

What the skill needs and how it handles safety

The skill needs a SKILL.md-compatible agent runtime and access to a web search tool. The skill instructs the agent to use WebSearch for factual questions, so without that tool it cannot do its research step. It does not run scripts, but it does use the network through web search.

The skill is a roleplay persona. Its own rules say that when it first activates it gives a one-time disclaimer that it is speaking from Karpathy's perspective, based on public statements and not his own views. You can end the persona by saying 「退出」, 「切回正常」, 「不用扮演了」 or 「跳出角色」.

The skill's own scope notes say it is not suited to business strategy, marketing, fundraising, politics or policy. Its output is an inference, so do not present it as what Karpathy actually believes.

• Runs scripts: no
• Uses network: yes, through web search for factual questions
• Licence: MIT
• Exit the persona with 「退出」, 「切回正常」, 「不用扮演了」 or 「跳出角色」

Tips for getting useful answers

Give the agent concrete material. The skill works best when you name a product, model, technique or event, because it then has something specific to research. Vague questions get more general framework answers.

Treat the output as a reasoned lens, not a quote. The skill is built from public talks, posts and writing, and its answers are model-based inferences that may be wrong or out of date. Check any specific claim against the original source before you rely on it.

• Name the specific product, model or event you are asking about
• Ask for a framework view, and verify any factual claims yourself
• For business, marketing or policy questions, use a different tool