Spec · verifier · environment
One guide in this compendium
Almost everyone is prompting Claude wrong. Karpathy’s 2026 method, as walked by Austin Marchese, is three layers: a detailed spec, a verifier, and an environment that compounds. Then the one thing you cannot outsource: understanding.
01 · The gap
I want to go to a car wash to wash my car, and it’s 50 meters away. Should I drive or should I walk? State-of-the-art models today will tell you to walk because it’s so close.
Marchese asked Claude, Gemini, Grok, and ChatGPT. They all said walk. The spec is how you hand the model the context it does not have. Plan mode is useful. Karpathy’s clip: go deeper, and design a detailed spec with the agent.
Dump the whole job and wait for the finish (waterfall)
Small specs, a checkpoint, review, adjust, repeat
Treat plan mode as the spec
Work with the agent to design a detailed spec
Yell “make it better” / assign a persona
Set evaluation criteria. Use a second model. Pull external signal
Start from a blank chat every time
CLAUDE.md, a knowledge base, skills, and rules the model cannot bypass
Outsource the goal to the model
You still own the understanding. The model executes inside it
02 · Three layers
Marchese’s packing list for the Karpathy clips — not Karpathy’s own branded stack. Click a layer to see what it does in the rebuilt prompt.
Deliver your understanding in a format Claude can use.
Do
Have Claude interview you for the actual goal (the decision the work drives). Bias to smaller, compartmentalized specs. Make yourself verify key decisions explicitly.
Stop
Stop handing a task name (“end-of-month report”) and calling it a spec. Stop waterfalling the whole job. Plan mode is useful; it is not the spec.
Three-layer prompt · video
Interview me to identify the goal of this project. Bias towards smaller and more compartmentalized specs. Make me verify key decisions explicitly. Outline precise evaluation criteria first. If this is a complex build, run the final by Codex so both systems agree. Verify against external signal, not your own prior output. Use this workspace’s CLAUDE.md, knowledge base, and skills. Include a verification plan before any multi-step build. Always-do on autopilot. Ask first when it is listed. Never-do is blocked at the tool.
Highlighted: the spec — Interview me to identify the goal of this project. Bias towards smaller and more compartmentalized specs. Make me verify key decisions explicitly to ensure nothing is missed.
03 · Workshop
Blueprint on the wall. Quality station by the door. The environment is the workshop itself. A long chat history is not this.
The spec is the drawing the shop works from. Goal, small slices, decisions you confirm.
Nothing leaves without a test. Criteria, a second librarian, a signal from the real world.
CLAUDE.md, knowledge on disk, skills, and hard rails. AI lives in your world — not the other way around.
Always do
Autopilot. Low cost if wrong.
Ask first
You want to double-check.
Never do
Enforce at the tool, not the prompt.
CLAUDE.md draft
# CLAUDE.md Workspace: my Claude workspace ## How this repo works This file is injected into every Claude session. Treat it as the operating manual, not a suggestion. ## Knowledge architecture Look in the LLM knowledge base folders before answering from memory. Cite the file you used. ## Skills If a custom skill matches the job, use it. Do not reinvent a handbook that already exists. ## Working rules - Before building anything multi-step, include a verification plan. - Prefer smaller, compartmentalized specs with a checkpoint I can review. - Make me verify key decisions explicitly. ## Always do - (add yours) ## Ask first - (add yours) ## Never do - (add yours) Never-do items that are critical must also be enforced with a pre-tool-use hook on write/edit — a line in this file is only a request.
04 · Rebuild
The old prompt piles persona, steps, and files into one message. The new one assumes a spec interview, a verifier, and a workshop already exist.
Interview me to identify the goal of this project. Bias towards smaller and more compartmentalized specs. Make me verify key decisions explicitly. Outline precise evaluation criteria first. If this is a complex build, run the final by Codex so both systems agree. Verify against external signal, not your own prior output. Use this workspace’s CLAUDE.md, knowledge base, and skills. Include a verification plan before any multi-step build. Always-do on autopilot. Ask first when it is listed. Never-do is blocked at the tool.
You can outsource your thinking, but you can’t outsource your understanding.
The three layers sit on your understanding of the bigger picture. You need to understand your goals, and what’s needed to direct AI to start working for you.
05 · Do this
Pick a repeating job. Fill spec, verifier, and environment. Drafts stay on this device.
Three-layer prompt
103 wordsInterview me to identify the goal of this end-of-month report — the decision it drives, not the document. Bias towards a smaller, compartmentalized spec. Make me verify key decisions explicitly. The report must have three sections, each ending with a recommendation. Outline those evaluation criteria before you write. If this gets complex, run the final by Codex so both systems agree. Match historical reports I attach; do not invent a format. Use the workspace CLAUDE.md, the knowledge-base folder of past reports, and the reporting skill. Include a verification plan before you start. Ask first if the goal could reasonably mean two different decisions.
Paste into Claude or Claude Code after the workshop exists — CLAUDE.md, knowledge base, skills.
06 · For review
This guide follows Austin Marchese’s walkthrough. Anything that is packaging, not a Karpathy primary source, stays here.
“Karpathy’s method” as Spec / Verifier / Environment is Austin Marchese’s packaging of Karpathy talks and clips. Karpathy is not shown naming these three layers as a branded stack.
Car wash, plan mode, animals vs ghosts, and “you can outsource your thinking, but you can’t outsource your understanding” are presented as Karpathy clips in the video. Wording here follows that transcript, not a primary paper.
Codex plugin, CLAUDE.md sections, always / ask / never, and pre-tool hooks are Marchese’s tactics for running the method in Claude Code / Cowork — not claimed as Karpathy’s own checklist.
The video shows on-screen “final prompts” that were not fully read aloud. Copy-paste prompts here are assembled from the spoken steps, not screenshots.