Karpathy Method

🗒️ Description

Austin Marchese distills how Andrej Karpathy actually uses AI in 2026 into a stacked, repeatable framework: most people are “prompting Claude wrong,” and the fix is three layers — the spec, the verifier, and the environment — topped by one thing worth focusing on: understanding. It’s a practical, tactical companion to the Agentic Engineering interview and the Software 3.0 thesis, aimed squarely at Claude Code users.

The opening hook is the same car-wash example from the Karpathy interview: ask any frontier model whether to drive or walk to a car wash 50 m away and they all say walk — because AI is “brilliant at what can be measured” but has no signal for context. The three layers exist to feed AI the context and verification it can’t generate itself.

🧩 The three layers

Layer 1 — The Spec

The spec is how you deliver your understanding to Claude in a format it can use. Karpathy finds plan mode “too high-level” — the real move is working with the agent to design a very detailed spec. Three rules:

• Uncover the goal, not the task. “Create an end-of-month report” is a task; the goal is the decision the report drives — something AI can never decide for you. Tactic: tell Claude to interview you to extract the goal.
• Be agile, not waterfall. People dump everything on the agent at once (waterfall). Instead, work in tight scopes with clear checkpoints — review, adjust, repeat. Tactic: tell Claude to “bias towards smaller and more compartmentalized specs.”
• Be precise and use your brain. Every assumption AI makes is a chance to drift. When Claude drafts a spec, think critically about it. Tactic: “Make me verify key decisions explicitly to ensure nothing is missed.”

This connects directly to Specification-Driven Development and Karpathy Skills (which targets the same four pitfalls: assumptions, overengineering, scope creep, weak goals).

Layer 2 — The Verifier

Sits on top of the spec. Built on Karpathy’s animals vs. ghosts mental model: humans are animals (driven by intrinsic motivation — “learn this or you’re fired” works), but AI is a ghost — Marchese reframes it as a robot librarian that can only answer from the books in its library and doesn’t know when a book is missing, so it confidently makes things up. Yelling, pleading, or “make it better” doesn’t work. The only real lever is verification. Three places to apply it:

• Set evaluation criteria up front. Before Claude touches anything, define what “good” looks like precisely. Not “make this report look good” but “three sections, each ending with a recommendation.” (Same precision principle as Layer 1.)
• Use a second model as critic. A second “robot librarian” from a different library can catch what the first got wrong. Tactic: install the Claude Code Codex plugin and route final output through Codex so both systems must agree.
• Pull external signal. Bring in ground truth: connect Claude to the deployment system to confirm a deploy actually succeeded; feed historical reports as a format reference.

Boris Cherny (creator of Claude Code): if Claude has a feedback loop, it will 2–3× the quality of the final result.

This is the practical core of Agentic Engineering’s “verifiability is the constraint” thesis.

Layer 3 — The Environment

The spec and verifier need somewhere to live — the workshop. Most people rebuild the workshop from scratch every session. Build one that compounds:

• A proper CLAUDE.md. Auto-injected on every prompt — the first thing Claude reads. Cover how the repo works, custom skills and routing, knowledge architecture, and key working rules. “Make this your environment. It’s your world, and AI is living in it.” See Harness Engineering and Context Engineering.
• An LLM knowledge base. A folder system you ingest your own training data into, structured so Claude knows where information lives. “Your data is your moat.” This is exactly Karpathy’s viral LLM Knowledge Bases concept — and exactly what this vault (Brain) is.
• Skills. Anything you do repeatedly → a custom skill. “The best way to find a leak in a hose is to run water through it” — the more you use skills, the faster they improve and compound.
• Rules and guardrails. A CLAUDE.md line (“don’t edit this folder”) is a request — Claude can still ignore it. For critical constraints, enforce at the tool level with a PreToolUse hook that blocks write/edit on protected files. Bucket actions into always do / ask first / never do.

📖 Takeaways

• The framework stacks: spec (deliver your understanding) → verifier (the only lever that actually moves quality) → environment (where it compounds over time).
• Precision recurs at every layer — every assumption you remove is drift you prevent.
• Prompt-level guidance is a request; hooks are a rule. Enforce the things that can’t be wrong at the tool level.
• The one thing to focus on in the age of AI: “You can outsource your thinking, but you can’t outsource your understanding.” All three layers are downstream of your understanding of the goal and the bigger picture.

✍️ Quotes

“AI is brilliant at what can be measured, but for context-driven things… it has no signal to act on.”

“We’re not building animals, we are summoning ghosts.” — Andrej Karpathy

“Your data is your moat.”

“You can outsource your thinking, but you can’t outsource your understanding.” — Andrej Karpathy

🔗 Related

• Agentic Engineering — the source Karpathy interview this video distills (verifiability, animals vs. ghosts, “can’t outsource understanding”)
• Software 3.0 — Karpathy’s broader paradigm thesis as a business playbook
• Karpathy Skills — a CLAUDE.md targeting the four LLM-coder pitfalls; Layer 1 in practice
• Specification-Driven Development — the spec-first discipline behind Layer 1
• LLM Knowledge Bases — the “build your own knowledge base” core of Layer 3
• Context Engineering / Harness Engineering — designing CLAUDE.md, skills, and hooks (Layer 3)
• Progressive Disclosure — index-first knowledge architecture, how this vault’s environment is built
• Brain — this vault, a working instance of the Layer 3 LLM knowledge base

📖 Further reading/watching

• 📺 Source: Stop Prompting Claude. Use Karpathy’s Method Instead. — Austin Marchese, 13:18
• The original Karpathy @ Sequoia interview the three layers are drawn from

---

Template: knowledge_note_info