Quick Answer: A Hacker News video recommendation site is a YouTube-style catalog built from links that HN commenters have shared. HackerTube, built by Andrej Karpathy with Claude, pulls about 40,000 comments since 2007, filters them down to 6,454 YouTube videos, and serves them through a YouTube-like interface with no user accounts.
Hacker News Video Recommendations: Inside HackerTube
Late one Friday, the author opened YouTube and found a feed of exaggerated thumbnails and recycled topics. Instead of fighting the algorithm, they asked Claude to build something different: a YouTube clone made from videos that Hacker News users had already vouched for. The result, HackerTube, went from idea to live site in roughly two hours, according to the original write-up. The interesting part isn't the interface. It's the decision to treat human comment threads as a recommendation engine.
What HackerTube Actually Is
HackerTube is a static-feeling video browser organised around topics rather than a personal feed. The catalog contains 6,454 YouTube videos spread across 150 topics, covering 2007 to 2026. Each video earned its place because at least one commenter linked it in a discussion, which is a stronger signal than a view count. A view count tells you a video was clicked. A comment saying "this explains the thing I've been stuck on for a week" tells you someone used it.
The site adds a few deliberate touches. It uses a "50% thumbnail" rather than the uploader's thumbnail, a choice the write-up mentions without fully explaining. It includes keyboard shortcuts, a "ForYou" section that works without an account, and an export of the site's data so others can rebuild their own version.
How the Pipeline Turned Comments Into a Catalog
The build followed a sequence that any data-oriented developer would recognise, even if the author never wrote a line of code by hand:
- Pull the raw comments. The author downloaded the full HN comment archive.
- Filter for YouTube links. Comments containing YouTube URLs were isolated first.
- Classify with an LLM. Claude Sonnet ran inside a Claude Code workflow to separate genuine recommendations from spam and noise.
- Enrich with metadata. The YouTube Data API supplied titles, channels, and other video details, using an API key the author created in Google Cloud.
- Deploy and harden. A VPS hosted the app, and Claude handled deployment, security, and optimisation.
The post's own numbers show the funnel: roughly 40,000 comments recommending a YouTube video survived filtering, and those comments resolved to 6,454 unique videos. The write-up says Claude handled nearly all of the engineering. The human steps were getting an API key, provisioning the server, and buying the domain.
Why the Filtering Was the Real Work
Here's where most people underestimate the project. Anyone can scrape links. The hard part is deciding which links count as a recommendation.
Raw link extraction picks up everything: "this is wrong, see [video]," drive-by promotional posts, and angry rebuttals. An LLM classifier can separate those cases reasonably well, but it's not perfect. A comment dismissing a video can look identical to one endorsing it when you strip context. That's a failure mode a practitioner has to plan for: false positives quietly add weak videos to the catalog, and nobody notices until someone clicks through to a dud.
The sensible mitigation is to keep the source comment alongside each video so anyone can audit why it's listed. It's also worth sampling a few dozen entries by hand before trusting the pipeline. Classification is a cheap step to get wrong and an expensive one to discover late.
The Counter-Intuitive Part: No Accounts, No Watch History
The popular advice for building recommendations says you need user accounts, behavioural tracking, and a watch-history model. Without those, the argument goes, you can't personalise anything. HackerTube quietly challenges that.
Its "ForYou" recommendations store state in the browser's local session rather than on a server. Nothing gets tied to an identity, and nothing needs a login flow, a password reset, or a privacy policy that reads like a legal trap. The trade-off is that your recommendations reset when you switch devices or clear storage. For a site where the recommendation quality comes from the corpus and not your personal history, that's a fair deal.
This matters because it reframes personalisation. Much of YouTube's recommendation power comes from behavioural data on huge populations. A curated corpus with strong editorial signal can deliver useful results with almost no user data at all. The catch is that you're trading breadth for taste, and taste comes from whoever commented.
HackerTube vs. YouTube's Default Feed
| Dimension | YouTube default feed | HackerTube |
|---|---|---|
| Recommendation source | Behavioural data across a massive user base | HN comments that link a YouTube video |
| User state | Tied to a signed-in account | Stored in local browser session, no account |
| Thumbnails | Uploader-supplied | Uses a "50% thumbnail," per the write-up |
| Catalog size | Effectively unbounded | 6,454 videos across 150 topics |
| Data portability | Limited to Google Takeout exports | Site data exportable to rebuild elsewhere |
The comparison isn't a verdict that HackerTube is better. It's a different tool for a different audience: people who want a curated, quieter feed of content that engineers have already found useful.
Where This Approach Breaks
Responsible readers should weigh the limits before copying the design.
Selection bias is baked in. Hacker News skews toward English-speaking software developers and founders. A catalog built from its comments will reflect those interests, and topics outside that world will be thin or absent. That's a feature if you want developer-adjacent content, and a serious gap if you don't.
Commenters aren't a random sample. People comment when they're engaged, which means outrage and strong opinions get more links than quiet, solid content. Some of the best videos may never get mentioned at all.
API quotas constrain refreshes. The YouTube Data API uses a daily quota system, with a default allocation of 10,000 units per project per day. Large-scale metadata refreshes can hit that ceiling, so you'll want to cache results and schedule updates rather than re-query on every page load.
Video availability changes. Videos get removed, made private, or age-restricted. A catalog built in one year will decay unless something checks links periodically.
Licensing and embedding rules apply. YouTube's API terms and embed policies govern how you display content. Check the current developer policies before you publish a derivative site, especially if you plan to monetise it.
How to Build Your Own Version
If the idea appeals to you, the transferable pattern is simpler than the full project. Start small:
- Pick one community with a strong link-sharing culture, such as a subreddit, a Discord, or a forum.
- Export a bounded slice of comments, say one year, so your classifier has manageable volume.
- Run an LLM classifier with a clear prompt that distinguishes recommendation from criticism, and log the reasoning for each decision.
- Store the source comment and permalink with every video so you can audit results.
- Cache YouTube metadata and schedule refreshes to stay under your quota.
Expect the first version to surface odd videos. That's the signal that your classifier needs tightening, not a sign the idea failed.
Frequently Asked Questions
What is HackerTube?
HackerTube is a YouTube-style website that lists about 6,454 videos recommended in Hacker News comments from 2007 to 2026. It organises them across 150 topics and works without user accounts. The site was built with Claude, and its creator has published the data so others can build their own versions.
How do you build a YouTube clone from Hacker News comments?
Download the comment archive, filter for YouTube links, and use an LLM to separate genuine recommendations from spam or criticism. Enrich the surviving links with YouTube Data API metadata, then display them in a video-browsing interface. Store each source comment alongside its video so you can audit the results.
Why does HackerTube need no user accounts?
It stores personalisation state in the browser's local session instead of on a server. That means no login, no password, and no identity tied to viewing behaviour. The trade-off is that preferences don't follow you across devices or survive clearing your browser storage.
When should you trust a curated recommendation feed over an algorithm?
Trust a curated feed when the community that produces the signal has strong domain knowledge and the topics are narrow enough for that community to cover well. Algorithmic feeds work better when you need breadth across many interests and have enough behavioural data to personalise. Neither is universally better; the right choice depends on your audience and goals.
Want to see the idea in practice? Open HackerTube, pick one topic you already follow, and compare its top videos with what YouTube serves you on the same subject. Then pass this to someone who's tired of the same recycled recommendations, and ask them what they'd build with the export.