Summary
This video from Nick Saraev rounds up how Andrej Karpathy has described using AI agents in his own work over the past two weeks. According to the video, Karpathy uses the Claude Code Mac desktop app with Opus 5.5 on medium effort as his default, and follows a rule of not reading or writing code or running terminal commands himself, instead talking to Claude and letting it handle the rest.
The habits covered include asking models to explain things in ASD-STE100, a strictly controlled technical English, keeping prompts short because piling on requirements can contradict earlier instructions, and requesting diagrams, HTML pages or explainer videos instead of reading long text. Examples include a one-line request to check recent cyber threats that triggered sub-agents, and deploying a project by giving Claude a server IP and a list of verbs.
The video also covers Karpathy’s skepticism of public benchmarks and his preference for Epoch’s ECI, his quirky tests such as a land-or-water world map built from model answers, and HackerTube, a Hacker News-based site he built after finding YouTube thumbnails to be slop. It closes on his delegation setup of Opus orchestrating Sonnet sub-agents and his habit of sharing a data source and prompt instead of an app.
📺 Source: Nick Saraev · Published November 05, 2026
🏷️ Format: Roundup
