How Does Andrej Karpathy Use Claude Code Day to Day?
Karpathy reportedly runs Claude Code on Opus 5.5 at medium effort, writes tiny prompts, and delegates to Sonnet sub-agents. Here is the workflow.
For Engineering
How does Andrej Karpathy use AI coding agents day to day?
According to a recent creator breakdown of his public comments, Andrej Karpathy uses the Claude Code Mac desktop app with Opus 5.5 at medium effort as his default. He avoids max effort. He reportedly does not read code, write code, or run terminal commands himself. He talks to Claude and lets it work behind the scenes. His prompts are short, he asks for diagrams over prose, and he delegates bulk work to cheaper Sonnet sub-agents.
A caveat: this account comes from one secondhand video summarizing things Karpathy said over about two weeks. It is not a primary document from him. Treat the details as reported habits, not an official playbook.
TL;DR
- Karpathy reportedly defaults to Opus 5.5 at medium effort in the Claude Code desktop app and does not use max effort.
- He follows a long-standing rule of not reading or writing code and not running terminal commands. He only talks to Claude.
- His prompts are getting shorter, not longer, because piling on requirements can make output worse when instructions conflict.
- He asks for diagrams, HTML pages, or explainer videos instead of long text, and requests explanations in a strict simplified English called ASD-STE100.
- For data work, he filters junk with scripts, then sends the rest to Sonnet sub-agents orchestrated by Opus.
- He shares the data source and the prompt instead of the finished app.
What model and effort setting does he reportedly use?
The reported setup is Claude Code in the Mac desktop app, running Opus 5.5 exclusively. The effort level is medium by default. The summary is explicit that he does not reach for max effort.
That choice is worth noting. Many builders assume the highest reasoning setting is always better. A medium default suggests he treats effort as a cost and latency dial and saves heavier settings for cases that need them. The sources do not say when, if ever, he raises it.
Why are his prompts so short?
The reported reasoning is that more requirements can make the model do worse work. People often add instructions that contradict, or partly invert, things they said earlier in the same prompt. The longer the prompt, the more likely it contains these hidden conflicts.
Two examples from the account:
- To harden a recent project, he reportedly texted something like "Check recent cyber stuff." Claude then spun up several sub-agents to investigate and strengthen the project.
- Asked how he deployed a project, he said he gave Claude the server IP address and the words "update, deploy, secure, harden, and optimize."
The pattern is a goal, a target, and a few verbs. The model fills in the plan. This works best when the model is strong and the task is common, such as deployment and security hygiene. It is a style choice for capable models, not a universal rule.
He is also reportedly polite. His prompt for a recent project opened with "Hi Claude, good morning. How are you?" and closed with "Please don't make any mistakes." Whether that line changes output is unproven, and the creator said as much.
What is ASD-STE100 and why ask for it?
ASD-STE100 is Simplified Technical English, a controlled language created for aircraft maintenance documentation. It uses a strict vocabulary and restricts sentence structure, which cuts unnecessary words and ambiguity.
Karpathy reportedly makes models explain things to him in this style. The appeal is clear. A model's default explanation tends to be padded and hedged. A constrained register forces plain, short, literal sentences. If you want to try it, you can simply ask the model to answer in ASD-STE100 style. Models will approximate the standard, though they will not enforce the official dictionary perfectly.
Does he prefer diagrams over text?
Yes, by the account. When he does not understand something, he asks for a diagram, an HTML page, or even a custom explainer video. The reasoning is simple: people absorb an image faster than a wall of text, and a strong model can generate the visual faster than you can read the equivalent explanation.
For builders, this is a cheap habit to copy. Instead of asking an agent to explain a codebase or a data pipeline in paragraphs, ask for a flow diagram or a single-page HTML explainer.
How does he split work between Opus and Sonnet?
For data projects, the reported workflow has three steps:
- Use Claude to write filters that strip out URLs, regexes, spam, and other junk.
- Take whatever remains, even if it is messy, and pass it on without polishing the format.
- Send it to Sonnet through Claude Code workflows, with Opus 5.5 (at medium effort) orchestrating a group of Sonnet sub-agents.
The stated benefit is that the work runs on his existing subscription while a cheaper model handles the volume. Opus plans and coordinates. Sonnet does the repetitive per-item work. This orchestrator-and-workers pattern is common in agent design, and it is a sensible way to control cost on large batches.
What did he build with this workflow?
The example in the account is HackerTube. After seeing a lot of thumbnail-heavy, low-quality content on YouTube, he reportedly decided to build an alternative. It scrapes Hacker News and builds an archive of videos that were repeatedly recommended there over the years. The idea is a YouTube-style site restricted to high-quality, long-lived content. He put it online.
The broader habit is to build a fix as soon as something annoys him, because the cost of trying has dropped so far.
Why does he share prompts instead of apps?
When people ask for something he built, he reportedly shares the data source and the prompt, not the app code. The logic is that the app is an abstraction layered on top of the data. If a model can generate an interface around that data on demand, each person can get a customized version without inheriting someone else's code.
This is a real shift for open sharing. A prompt plus a dataset is small, portable, and adaptable. The tradeoff is reproducibility. Two people running the same prompt will get different apps, and different model versions will drift further.
What does he think of public benchmarks?
He reportedly distrusts them. In a recent Hacker News comment, he pointed people to Epoch's ECI, described in the source as a "benchmark of benchmarks" that analyzes how models score across many tests and whether those tests deserve attention.
He also runs odd, hands-on tests. In one, he gave a model latitude and longitude coordinates and asked a single question: land or water? He colored each answer black for land and white for water and assembled a world map. The point was to show how much world knowledge is compressed inside the model. The takeaway for builders is to design your own small probes for the things you care about.
Is this workflow worth copying?
Parts of it transfer well. Parts depend on his circumstances.
| Habit | Easy to adopt? | Caveat |
|---|---|---|
| Medium effort default | Yes | Raise it for hard problems |
| Short prompts | Yes | Works best on common tasks with strong models |
| Diagrams over prose | Yes | Check the diagram for errors |
| Opus orchestrating Sonnet | Yes | Needs a task that splits into independent items |
| Never reading code | Risky | Needs strong tests and trust in your setup |
| Sharing prompts, not apps | Depends | Results vary between runs and models |
The "never read the code" rule deserves the most caution. Karpathy is an expert who can judge outcomes without reading the source. A newer builder who skips review may miss security holes or silent failures.
Frequently Asked Questions
Which model does Karpathy reportedly use in Claude Code?
The account says he uses Opus 5.5 in the Claude Code Mac desktop app, at medium effort by default. He does not use max effort. For bulk jobs, he has Opus delegate to Sonnet sub-agents.
Does Karpathy really never read code?
The summary says he has a long-standing rule against reading or writing code and against running terminal commands himself. He communicates only with Claude. This is a secondhand report, so the exact scope of the rule is unclear.
What is HackerTube?
HackerTube is a project he reportedly built and put online. It is a YouTube-style site limited to videos that Hacker News users repeatedly recommended over the years. He built it after finding much of YouTube to be low-quality.
Why would shorter prompts give better results?
The reported explanation is that people often add requirements that conflict with earlier ones without noticing. The model then tries to satisfy contradictory instructions. A short prompt with a clear goal leaves less room for those conflicts.
Where does this information come from?
It comes from a single YouTube creator summarizing Karpathy's recent comments and projects. It is not a first-party statement, so verify details against his own posts before relying on them.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles
How Much Does Gemini 4 Carbon Cost and Who Can Use It?
Coding Agent Context Rot: How to Fix It With a Handoff File
Why Does Your AI Coding Agent Say It's Done When It's Broken?
