One weekend project became a template for mining HN

HackerTube started as a complaint about YouTube's recommendations. Its creator, Andrej Karpathy, says he asked an AI coding agent, Claude Code running Opus 5.5, to build a YouTube-style site from videos that Hacker News commenters had recommended. By his account, the whole thing took about two hours.

The pipeline is simple. The agent pulled every HN comment, filtered out spam, and narrowed the pile to roughly 40K comments that point to a YouTube video. The result is 6,454 videos across 150 topics, recommended between 2007 and 2026. A familiar video interface, a domain and a server did the rest.

The clones arrive

Two imitators followed. HackerBooks applies the same idea to books. It lists about 16.5k titles that at least two different HN commenters mentioned, each linked back to the original comments, with the data available to download. The post collected 1404 points and 234 comments.

The product is no longer the site. It is the recipe: scrape the comments, filter, rank, wrap in a familiar interface.

Wikihack does it for Wikipedia. It scrapes every mention of the encyclopedia on HN and ranks the articles. It offers a feed with three modes: most frequently linked, random with a tilt toward popular articles, and rising, which favours the past 12 months. It runs entirely in the frontend. It drew 197 points and 56 comments.

Why it spread

Both clone authors say they saw HackerTube and wanted the same thing for another medium. The HackerBooks author is blunt that Claude wrote almost all of it, including the data pipeline and the site, while he steered and made decisions. The effort needed to copy the idea looks very small.

That suggests the real asset is the method, not any one site. Community comment threads hold years of unstructured taste signals. A coding agent can turn them into a ranked, browsable catalogue in a weekend. Any medium that people link to can be treated the same way.

What it signals

The format has limits. HackerBooks counts mentions, so praise and criticism score the same, and its author admits some matches will be wrong. A mention count is a measure of attention, not endorsement. If the format keeps spreading, quality filtering will matter more than scraping.

Still, the pattern is clear. A single post about a personal gripe with a recommendation algorithm became a repeatable recipe: community-curated media, rebuilt as a browsable product, with an AI agent doing most of the work. If that holds, expect more of these niche, human-recommended feeds to appear as alternatives to algorithmic ones.