Summary
Nate Herk shows how to turn the teaching approach of Andrej Karpathy into a custom Claude agent using Claude Code. The agent is not a voice clone or impression. It captures how Karpathy explains things, following seven rules drawn from his writing and lectures, such as building the smallest version first, predicting results before running code, showing the broken version, and never handing over code that has not been run.
The build runs in four steps and six prompts. First, parallel agents collect public material, including YouTube captions pulled with the YouTube Transcript API and yt-dlp, and X posts since 2023 fetched through twitterapi.io for about a dollar. Next, the material is organized into an Obsidian-based wiki using Karpathy’s own LLM wiki memory pattern, with sources, topics, principles and methods linked together. A further step extracts the rules, each backed by an exact quote and its source, so the agent does not invent claims.
Herk explains why this matters: Claude often overexplains or sounds right without being right, and Karpathy himself has said Claude Code prefers writing code to teaching. The same four-step method can be applied to any expert, author or coach, giving builders a way to put real expertise into their own skills, agents and workflows.
📺 Source: Nate Herk | AI Automation · Published October 05, 2026
🏷️ Format: Hands On Build
