Karpathy Built HackerTube in Two Hours with Claude Opus 5.5
Andrej Karpathy used Claude Opus 5.5 (Claude Code) to filter YouTube videos recommended over the past two decades from Hacker News comments and built HackerTube, a YouTube look-alike. About 40,000 recommendation comments were consolidated into 6,454 videos spanning 2007–2026 and covering 150 topics, all in roughly two hours. By hand, he did only three things: apply for a Google API key, spin up a VPS, and buy a domain; the rest was done by conversing with Claude. Once the VPS was up, he handed over the IP and let it update, deploy, harden, and optimize on its own. The site's thumbnails use the 50% progress frame, For You requires no login and stores state in the browser's local session, and the data can be exported for your own agent to make further variants.
Why read this
AI score 72
The original post documents the entire process from filtering HN comments to deploying the site live, showing the concrete path to building a replicable site with Claude Code in two hours.
Topics
Body · AI translation
HackerTube: Andrej Karpathy built a "purified YouTube" from videos recommended on Hacker News, and Claude Opus 5.5 finished it in two hours
After being driven to despair on Friday night by YouTube's exaggerated thumbnails and algorithmic junk, Andrej Karpathy used Claude Opus 5.5 (Claude Code) on Saturday to sift through Hacker News comments for YouTube videos recommended over nearly two decades, and built HackerTube, a YouTube lookalike: about 40,000 recommendation comments → 6,454 videos, 150 topics; he himself barely writes code anymore, spending the whole time talking to Claude, and manually did only three things—Google API keys, spinning up a VPS, and buying a domain—the whole process took two hours!
https://t.co/MeAdrJpISQ
# Motivation: mainstream recommendations failed him
When he opened YouTube, what he saw was "a degraded wall of junk": exaggerated thumbnails, repetitive topics, low-intelligence content. He acknowledged that this kind of algorithm may suit the "median user" on the internet, but not him. The solution was not to tweak recommendations again, but to switch the source: treat videos that "someone seriously recommended" in HN comments as a curation layer.
# How it was made
Opus pulled all HN comments, filtered them first, then used Sonnet through Claude Code's Workflow to remove junk and low-quality items, narrowing it down to about 40,000 comments "recommending YouTube videos," which were aggregated into 6,454 videos spanning 150 topics between 2007 and 2026. Then he applied a YouTube-style interface, bought a domain, pushed it to a VPS, and did security hardening and optimization. At the end he added notes on his working method: by default he uses Opus 5.5 Medium day to day; he no longer reads or edits code himself or types in the terminal, only talks to Claude; once the VPS was up, he handed it the IP and let it update, deploy, harden, and optimize on its own. He said there is still a bit of a feeling of "taking turns working with a teammate," because in between you have to wait for Claude to finish.
# A few deliberate product choices
On the content side, it bets on "someone on HN recommended it"; for thumbnails it does not use uploaded original images, but instead uses the 50% progress frame (avoiding exaggerated covers); For You requires no login, with state kept in the browser's local session; there are keyboard shortcuts; site data can be exported, making it easy to hand to your own agent to make another variant. In one sentence: curation relies on community recommendations, and the interaction is as light and reproducible as possible.