Karpathy builds HackerTube video recommendation site with Claude Code in two hours
Karpathy built the video recommendation website HackerTube in about two hours using Claude Code and Claude Opus 5.5, using Claude to analyze 43 million Hacker News comments from the past 20 years and filter out 6,454 videos that were consistently recommended. The site requires no registration or login, uses a mid-video frame rather than the creator's thumbnail for thumbnails, and publicly releases a ZIP package of the comment analysis data for others to adapt themselves. He said he only manually created the Google Cloud API key, started the VPS, and purchased the domain, while Claude completed the rest of the code, deployment, and security hardening.
Why read this
AI score 69
The original article discloses the specific division of labor in building the recommendation site from scratch with Claude Code, which can be used to observe the practical boundaries of AI programming tools in personal projects.
Topics
Body · AI translation
Karpathy stated on social media that compared with YouTube's existing recommendation mechanism, HackerTube can provide a "significantly better viewing experience."
Unlike video platforms that rely on users' viewing history and personalized recommendation algorithms, the core idea of HackerTube is to use real community discussion records to find video content that has been recommended and recognized over the long term.
The project's data source is Hacker News comments from the past 20 years. Karpathy had Claude analyze 43 million of those comments and ultimately filtered out 6,454 videos that had been continuously recommended, thereby building a curated video library.
HackerTube also has several design choices that differ from traditional video platforms. First, users do not need to register or log in to an account; usage state is stored directly in the local browser session. Second, the site no longer uses the thumbnail images uploaded by YouTube video creators, but instead captures a representative frame from the middle of the video as the thumbnail.
Karpathy believes this can avoid video thumbnails that have been repeatedly A/B tested and specifically optimized to attract clicks, so that users focus more on the content of the video itself.
In addition, HackerTube also provides a ZIP archive of the underlying comment analysis data, allowing other developers to download it and hand it to Claude for analysis, and to create different versions of video recommendation sites according to their own needs.
What is more noteworthy than HackerTube itself is the development process Karpathy disclosed.
According to him, the entire project was mainly completed using Claude Code, with the Claude Opus 5.5 model and the reasoning effort set to Medium. He used the Mac version of the Claude Code desktop application.
Karpathy said that he only needed to converse with Claude to have the AI complete almost all of the development work.
Throughout the process, there were mainly only three operations he handled personally: creating an API key through the Google Cloud interface in order to query YouTube video metadata; starting a VPS server for deploying the website; and purchasing the website domain.
After the server was ready, Karpathy provided the IP address to Claude and asked it to update the code, strengthen security protection, complete deployment, and optimize performance. According to him, Claude then completed this work.
He finally raised a question: "What other possibilities have we not yet thought of?"
HackerTube is not only a redesign of YouTube's recommendation mechanism, but also demonstrates that AI programming tools are changing the way individual developers build internet products. However, the roughly 2-hour development time disclosed by Karpathy is a personal project practice and does not mean that similar products can all be completed at the same speed in terms of data processing, long-term maintenance, and security assurance.