Karpathy says Claude built HackerTube from 43 million Hacker News comments

• HackerTube collects 6,454 video picks from 43 million Hacker News comments, with a roughly two-hour Claude build reported in its creator's announcement.
• URL filters, regexes and spam removal narrowed the archive before Sonnet reviewed the remainder, according to a follow-up clarification.
• Recommendations need no account, and the underlying analysis is available as a ZIP for custom versions, as described in the launch post.

HackerTube uses the frame halfway through each video as its thumbnail, a wonderfully blunt design choice. There is a wrinkle in the curation: Kingy.ai's inspection of the export found that 1,658 videos, or 25.7% of the catalog, had no positive recommendation label among their retained comments.

Claude split the data work into stages:

1. Download the Hacker News comment archive.
2. Narrow it using URL filters, regexes and spam removal.
3. Dispatch the remaining material to Sonnet through a Claude Code Workflow.

Karpathy says the Workflow used his subscription tokens. His build note describes Sonnet identifying comments that recommend YouTube videos, followed by assembling the catalog and interface.

HackerTube creator Andrej Karpathy describes a conversation-only build: Claude Code's desktop app for Mac, with Opus 5.5 at Medium effort. He says he neither wrote code nor ran terminal commands.

His three manual tasks were:

1. Create a Google Cloud API key for fetching YouTube metadata.
2. Boot a virtual private server.
3. Buy the domain name.

Once the server was running, he gave Claude its IP address. Claude handled deployment, hardening and optimization, according to his account.

The live site keeps its recommendation state in the browser rather than requiring a user account. Alongside the midpoint thumbnails, Karpathy's build note and launch screenshot identify:

• A catalog organized across 150 topics.
• Keyboard shortcuts.
• A dedicated Longs view.
• Liked, History and Watch later navigation.

Videos qualified through recurring discussion, according to Kingy.ai's analysis of the project documentation:

• Mentions from at least three different commenters.
• Mentions across at least three different months.
• Up to 20 retained mentions per video.

Sonnet's classification distinguishes recommendations from other kinds of references. A repeatedly linked video can therefore enter the catalog even when its retained mentions are critical rather than enthusiastic.

The 20-mention cap also leaves room for recommendations outside the exported sample. Kingy's count measures the supplied model labels, not an independent assessment of the videos.

Karpathy offered the comment-analysis data as a ZIP for agents to turn into custom versions. The download is available through the site's About page.

According to Kingy.ai's inspection, the package contains:

• Video records and a topic taxonomy.
• Labeled mentions, including Hacker News comment IDs and source URLs.
• Related-video data and embeddings.

The export leaves out some metadata used by the live site, including video durations and view counts. Karpathy's build obtained additional metadata through the YouTube Data API.