AI researcher Andrej Karpathy built a new website on a Saturday morning and wrote that it took about two hours. HackerTube, published on 31 October, collects 6,454 YouTube videos that people recommended in Hacker News comments between 2007 and 2026 and sorts them into 150 topics. The interesting part for everyone else is his description of how he built it, because almost all of the work was done by Claude Code, and the three jobs he kept for himself are the ones anyone copying the method should keep too.
What HackerTube is
HackerTube looks like a stripped-down YouTube. Every video on it was recommended by someone in a Hacker News comment, so the feed reflects what a community of readers chose to recommend. Karpathy's post says there are no user accounts, and your history and likes are kept in your own browser session. He also added keyboard shortcuts and a data export, so other people can build their own versions from the same list. The Hacker News thread about the project passed 2,800 points and 600 comments within two days, which on that site signals a lot of interest from developers.
The tools he used
According to Karpathy's write-up, he used the Claude Code desktop app on a Mac, with Claude Opus 5.5 on medium effort as the default model. The build then ran in three stages:
- Claude pulled every Hacker News comment from a public dataset hosted on Hugging Face.
- Claude Sonnet, a cheaper model, ran over the comments in a Claude Code workflow, removed spam and narrowed them to about 40,000 comments that recommend a YouTube video (39,834, according to his Hacker News post).
- The YouTube API filled in titles, thumbnails and other details for each video.
He then gave Claude the address of a fresh server and asked it to update, deploy, secure, harden and optimise the code. In his words, he no longer does traditional engineering on projects like this and only talks to Claude.
The three jobs he did by hand
Karpathy lists only three manual steps. He created an API key in the Google Cloud console so the project could read YouTube data, he started a virtual private server, and he bought a domain name. Each of these involves an account, a payment or a credential. Keeping them in human hands is a sensible line to draw, because an agent that can create keys or spend money on your behalf is a larger risk than one that writes code you can read.
What it used, and what the write-up leaves out
The blog post gives no cost. When a commenter on Hacker News asked, Karpathy replied that he was on a subscription and that Claude had added up the usage at 504 million input tokens, 90% of them cached, and 7.6 million output tokens. More than 90% of that went to Sonnet calls labelling comments, and the app and deployment took about 1 million tokens of Opus 5.5. Other commenters put the bill at roughly $250 to $300 at API prices, an estimate Karpathy did not confirm. The recipe has other limits worth stating plainly.
- Two hours is the time for someone who knows exactly what to ask for and how to judge the result. A first-time builder should expect longer.
- Letting an agent configure and harden a public server works best when someone can check what it changed. If you cannot review that yourself, start with a hosting service that handles security for you.
- Commenters on Hacker News warned that a popular curated list invites spam and gaming, and noted that the site only adds a video once someone links it on Hacker News. Keeping it clean and current is a maintenance job the two hours do not include.
Why the method matters
We think the most transferable idea in the recipe is the split between models. Opus did the planning and the hard parts, while the cheaper Sonnet model did the bulk sorting of comments. At Anthropic's published API prices, Claude Sonnet 5.5 costs half as much per token as Opus 5.5, at $2 against $4 per million input tokens and $10 against $20 per million output tokens. With more than 90% of the tokens going to Sonnet, that choice decides most of the bill. The same pattern fits plenty of work tasks, with one capable model to plan and check and a cheaper one to process a large pile of text.
What to do this week
- Pick one list you already have and would like to browse better, such as your team's saved links, a folder of customer quotes or a spreadsheet of past campaigns.
- Ask Claude Code to plan a simple one-page site for it first, and read the plan before it writes code.
- Use a cheaper model for any step that processes many items, and set a spending limit on your account before you start.
- Keep accounts, keys, payments and domain purchases for yourself, as Karpathy did.
If you are new to Claude, the Claude guide in Learn explains the models and plans, and our piece on a malicious Claude Code extension is a good reminder to install only tools you trust.