ABAB News Logo
Back to Flash News

Andrej Karpathy: Building HackerTube and TristanTube quickly with Claude

Andrej Karpathy posted his weekend project HackerTube (hackertube.org) around November 1. He said that because he was dissatisfied with YouTube recommendations, he used Claude (Opus 5.5, through the Claude Code desktop app, Medium effort) to filter about 20 years of Hacker News comments (about 43 million comments) and screen out 6,454 videos that were repeatedly recommended. Building it took about 2 hours, and his manual involvement was limited to creating a Google Cloud API key to obtain YouTube metadata, starting a VPS, and buying a domain name; the rest was deployed, hardened, and optimized by Claude.

HackerTube does not require login (status is stored in the browser's local session), thumbnails are representative frames from the middle of the video rather than optimized images from the uploader, and it provides zip files of comment analysis data for others to use Claude to generate their own versions. Karpathy said the viewing experience is significantly better than YouTube recommendations.

On November 7, Karpathy further released TristanTube (karpathy.ai/tristan-tube/). He cited Tristan's YouTube channel ranking list published this week, converted it into a site containing 47,000 videos, and added TV mode. The version reportedly was done in one shot with Claude Code, taking about 3 minutes.

Original source:

  1. Title: Andrej Karpathy X posts on HackerTube and TristanTube Publisher: Andrej Karpathy (@karpathy) Publication date: 2026-11-01 and 2026-11-07 URL: https://x.com/karpathy/status/2119114537442189652（及相关前帖） Source type: original social media post Verification: Read the post content provided by the user in this round; cross-checked that the site was accessible and consistent with media reports Supporting content: the HackerTube filtering process and figures, building time and manual steps, TristanTube conversion and the 3-minute statement

ABAB AI Insight

Karpathy used community recommendation signals (HN comments) instead of algorithmic recommendations as the basis for content filtering, which is an individual practice rather than a platform-level product. HackerTube delegated large-scale text filtering and site building to Claude, while he himself only handled account and infrastructure access, reflecting a conversational development path. Open data (zip files) allows others to replicate or customize it.

TristanTube further demonstrates the rapid conversion of an existing list, reducing the time from hours to minutes. Both rely on external video hosting (YouTube), while the site itself provides browsing, recommendations, and state management. Build speed depends on model capability and prompt quality, while actual maintainability, copyright compliance, and long-term stability are not explained in the post.

Such projects show that individual developers can use coding agents to build usable content sites in a short time. The effect depends on source data quality, filtering accuracy of the model, and subsequent maintenance costs; this is currently a public personal experiment.

ABAB News · Cognitive Laws

  1. Community comment signals can replace algorithmic recommendations to generate video directories.

  2. Coding agents compress site building from weeks to hours or minutes.

  3. Open data allows others to quickly generate their own versions based on the same source.