How Andrej Karpathy uses AI agents day to day

Key ideas

• Karpathy no longer reads code, writes it or runs commands. He talks to Claude and it does the rest.
• Short prompts work better as models improve, because extra requirements can steer the model away from a better answer it would have found.
• When an answer is hard to follow, ask for a different format: plain controlled English, then a diagram, then an HTML page, then an explainer video.
• For big data jobs, filter cheaply with code first, then send what's left to Sonnet subagents while Opus orchestrates.
• Share the data and the prompt, not the app, since anyone's agent can now build their own version around the data.

Andrej Karpathy builds whole projects now by only talking to Claude: short prompts, plain English or pictures when he needs to understand something, cheap filters before Sonnet does the bulk work, and the data plus a prompt instead of the app when he shares. Nick Saraev’s five-minute video collects these habits. These are my notes, with each one traced back to Karpathy’s own posts on X, his blog and Hacker News.

Some context on who he is: Karpathy was a founding member of OpenAI, led AI at Tesla, coined the term vibe coding, and joined Anthropic’s pretraining team in May 2026. The posts the video draws on run from October 1 to November 1, 2026.

Talk, don’t touch the code

His setup, in his own write-up of HackerTube: the Claude Code desktop app for Mac, Opus 5.5, and medium effort as the default rather than the maximum. He says he doesn’t do anything that looks like traditional engineering anymore. He doesn’t read code, write it or run terminal commands.

On HackerTube, the only manual work he did was in the browser: creating a Google Cloud key to look up YouTube video details, booting a server, and buying the domain (plus its DNS record, he added on Hacker News). He did those while Claude worked, which he says still felt a bit like teamwork. He also said he’d only trust this much autonomy to Opus 5.5 or better.

Short prompts and plain English

On Hacker News he wrote that his prompts keep getting shorter because the model understands what he’s doing with fewer words. Give it too much and you constrain it with requirements, when what it would have done anyway was better. He thinks the skill is moving towards fewer words, chosen well.

His examples are almost comically short. To tighten security on a project he types something like “check recent cyber stuff”, and Claude sends out a few subagents and does it. To ship HackerTube he gave Claude the server’s IP address and five words: update, secure, harden, deploy, optimize.

Going the other way, when Claude’s output is hard to read, his October 1 thread suggests asking for the explanation in ASD-STE100, the controlled English of aircraft manuals. He says models know it well and its rules make for cleaner writing. Because the full standard is strict, he sometimes asks for 80% of the way there.

Pictures over paragraphs

The same thread is a ladder, with each step better than the last:

1. Plain controlled English, as above.
2. A diagram, which is easier to take in than the same idea written out.
3. An HTML page, for something interactive and well designed, since models have become good at front-end work.
4. A custom explainer video, the format he’s most excited about: ask for a 3Blue1Brown-style video on any topic, with narration from a text-to-speech service.

His reasoning: as models do more of the legwork, more of our work moves up into oversight and understanding. Code is now cheap enough that a one-off web app or video made just to explain one thing is worth asking for and then throwing away.

Benchmarks and his own tests

Someone in the HackerTube thread asked if he still recommends the leaderboards he pointed to in a February 2025 video. He said no. He now points people to Epoch’s ECI, which combines 41 benchmarks into one capability score, and said he basically doesn’t trust benchmarks otherwise. He also linked Epoch’s reviews of individual benchmarks.

The video says he made a world map by asking a model “land or water?” at each coordinate. That eval was someone else’s: on October 1 he shared another person’s results for every Claude model and called it a cool eval. The method is simple. Give the model a latitude and longitude as text, ask land or water 16,200 times, and plot the answers as an image, one colour for land and one for water. A recognizable world map comes out, because the models absorbed geography from the internet they were trained on.

HackerTube and the filter pipeline

HackerTube started on a Friday night at 11pm, when he opened YouTube and found exaggerated thumbnails and stale recommendations. The next morning he asked Opus 5.5 for a YouTube clone made only of videos Hacker News kept recommending. It took about two hours.

The pipeline is the useful part:

• Get the raw data. Opus downloaded every Hacker News comment, about 43 million over 20 years, from a public export on Hugging Face.
• Filter cheaply first. Opus wrote filters that cut the comments down before Sonnet read any of them. The video describes these as simple filters for links and spam patterns.
• Send the rest to Sonnet. Sonnet labelled the remaining comments in a Claude Code workflow and threw out spam, leaving about 40,000 comments that recommend a YouTube video, and 6,454 videos across 150 topics.
• Build and ship. Opus wrapped it in a YouTube-style interface and deployed it.

He reported the cost in tokens: 504 million input tokens (90% cached) and 7.6 million output tokens. More than 90% of that was Sonnet labelling comments. The app and the deploy took about one million Opus tokens. It all ran on his subscription. A few touches stand out: each thumbnail is the frame from the middle of the video rather than the uploaded image, recommendations work with no account because they’re kept in your browser, and there’s a section for videos over an hour.

Share the data, not the app

He didn’t publish the HackerTube server code, which he called a vibe coded mess. Instead the site’s about page has a zip of the underlying data. His suggestion: point Opus 5.5 at the zip and ask for a YouTube-like site, which he says took about three minutes in his build. The one extra step is a free YouTube API key, since he couldn’t share YouTube’s data.

The idea underneath, as Saraev puts it, is that the app is a thin layer around the data. Once a model can build an interface on request, the data and the prompt are the part worth sharing. Karpathy’s April Sequoia Ascent talk makes a similar point: some apps stop being needed once a model can do the transformation directly.

Saraev also says Karpathy opened the HackerTube session politely, with a good-morning greeting and a closing request not to make any mistakes. I found that prompt only in a press report, not in Karpathy’s own posts.

How to try it

• Write your next prompt as one short line and see what the model does before adding requirements.
• When an answer is hard to follow, ask for it in ASD-STE100, or for a diagram or an HTML page instead.
• Before sending a big dataset to a model, have the agent write a cheap filter first, then send what survives to a smaller model.
• When you share something you built, share the data and the prompt along with it.

Key terms

ASD-STE100

Simplified Technical English, a controlled language with strict grammar rules and a small fixed vocabulary, written for aircraft maintenance manuals.

Epoch Capabilities Index (ECI)

Epoch AI's single score of general model capability, stitched together from 41 separate benchmarks.

Subagent

A separate model session that a main agent sends off to do one piece of work and report back.

Workflow

A Claude Code feature that runs a scripted batch of subagents, which Karpathy used to have Sonnet label comments in bulk.

HackerTube

Karpathy's YouTube-style site of 6,454 videos that Hacker News commenters kept recommending between 2007 and 2026.

Effort

The Claude Code setting for how long the model thinks before it answers; Karpathy's default is medium.

Quotes from the talk