How to Use Karpathy's Method Instead of Prompting Claude

TL;DR

Karpathy’s method for getting faster results with Claude uses three layers: a detailed spec, a verifier, and a supportive environment. Start by uncovering the real goal, divide work into tightly scoped checkpoints, and explicitly review key decisions so the AI makes fewer assumptions. Then apply measurable verification criteria, tools such as a Claude MD file, and rule-based guardrails. Read on for the practical workflow behind each layer.

Transcript

I just listened to Andrea Cararpathy speak at AISN 2026 and I learned something that I wasn't expecting. Almost everyone is prompting Claude wrong. So I decided to dig deeper and see exactly how Carpathy, the former head of AI at Tesla, uses AI in 2026. And it turns out that Karpathy's method for building 10 times faster can be broken down into thr... Read More

Key Insights

• Karpathy's method is based on three layers: spec, verifier, and environment.
• A spec translates human goals into a format AI can understand and act upon.
• Verification involves setting precise evaluation criteria to ensure AI outputs meet expectations.
• Using multiple AI models can enhance verification by providing diverse perspectives.
• Creating a supportive environment involves setting up tools and systems for efficient AI interaction.
• A Claude MD file can automate verification steps in AI processes.
• Building a personal LLM knowledge base helps organize and utilize training data effectively.
• Establishing clear rules for AI tasks ensures critical processes are safeguarded against errors.

Questions & Answers

Q: What is Karpathy’s three-layer method for using Claude?

The method consists of three layers: the spec, the verifier, and the environment. The spec conveys your goal and context, the verifier evaluates the output, and the environment supplies the tools, knowledge, and rules needed to support the process.

Q: How do you create a detailed spec for Claude?

First, uncover the actual goal rather than merely naming a task; for example, identify the conclusion or decision an end-of-month report should support. Ask Claude to interview you to identify that goal, then turn the information into a precise, tightly scoped spec.

Q: Why is a detailed spec better than relying only on Claude’s plan mode?

Karpathy describes plan mode as useful but too high-level for this purpose. A detailed spec goes deeper by giving the agent the context, decisions, scope, and checkpoints it needs to build the intended result.

Q: What is agile speccing with AI?

Agile speccing breaks a large task into smaller, compartmentalized specs instead of assigning everything at once. Each unit should have a tight scope and clear checkpoint so you can review the output, adjust it, and repeat.

Q: How can you reduce incorrect assumptions in Claude’s output?

Be precise, because every assumption the AI makes creates another opportunity to drift from the desired result. The transcript recommends making yourself verify key decisions explicitly so important details are not missed.

Q: Why is verification important when using AI?

AI can perform well on measurable work but may miss context-driven requirements, so its output needs explicit evaluation criteria. A verifier checks the result against predefined expectations, and multiple AI models can provide additional perspectives during that review.

Q: How does a Claude MD file support the verification process?

A Claude MD file can embed verification instructions and protocols into the workflow. This helps apply the same checks consistently across tasks and reduces the chance that a verification step will be overlooked.

Q: What belongs in a supportive AI environment?

A supportive environment includes the tools and systems that help the AI access relevant resources and follow established processes. The existing framework highlights a personal LLM knowledge base for organizing training data and rule-based guardrails for protecting critical tasks.

Summary & Key Takeaways

• Karpathy's method involves three layers to optimize AI use: a detailed spec, a verification process, and a supportive environment. The spec ensures AI understands the goal, while the verifier checks the accuracy of outputs. The environment provides tools and systems to facilitate efficient AI interaction.

• A spec is crucial for translating human goals into a format AI can understand. Verification involves setting precise evaluation criteria to ensure AI outputs meet expectations. Using multiple AI models during verification can enhance accuracy by providing diverse perspectives.

• Building a supportive environment includes setting up tools and systems for efficient AI interaction. A Claude MD file can automate verification steps, while a personal LLM knowledge base organizes training data effectively. Establishing clear rules for AI tasks ensures critical processes are protected.