Karpathy Built HackerTube With Claude in About Two Hours: How It Works and What It Does Not Say

Andrej Karpathy published HackerTube, a YouTube-style site at hackertube.org listing 6,454 videos recommended in Hacker News comments from 2007 to 2026 across about 150 topics, with Claude Opus 5.5 and the Claude Code desktop app doing nearly all the work in about two hours, per his blog post. Karpathy said he handled only three manual steps — creating a YouTube API key, booting a VPS and buying the domain — while Sonnet ran over roughly 40,000 filtered comments inside a Claude Code Workflow; the post drew more than 2,000 points on Hacker News, where commenters asked about token and dollar cost, which the post does not state, and requested sorting by upload date.

Andrej Karpathy has published HackerTube https://hackertube.org/?ref=explainx , a YouTube-style website that lists 6,454 videos Hacker News commenters keep recommending. By his own account on his blog https://karpathy.bearblog.dev/hacker-tube/?ref=explainx , Claude Opus 5.5 and the Claude Code desktop app did nearly all of the work in about two hours, while he handled three small chores: an API key, a server and a domain. The post drew more than 2,000 points on Hacker News https://news.ycombinator.com/item?id=50319510&ref=explainx . It is a useful case study for builders because the recipe is short, the data is public, and the write-up is candid about what he did and did not do. It is also thin in places, and those gaps matter if you want to repeat it. This post lays out both. TL;DR | Question | Answer | |---|---| | What is it? | A YouTube-style site at hackertube.org built from videos recommended in Hacker News comments | | Size | 6,454 videos, 2007 to 2026, about 150 topics | | Model | Claude Opus 5.5 at medium effort, with Sonnet over the comments in a Claude Code Workflow | | Tool | Claude Code desktop app for Mac | | Build time | About 2 hours, per the author | | Manual steps | YouTube API key, a VPS, a domain name | | Cost | Not stated | | Open source? | No repo linked; data export offered | How was HackerTube built? Karpathy's account gives the pipeline in five steps. We are relaying them from his post, not from our own run. 1. Download all Hacker News comments from a Hugging Face dataset. 2. Filter the comments and remove spam. 3. Narrow to roughly 40K comments that recommend a YouTube video. 4. Wrap the resulting videos in a YouTube-style interface. 5. Add a domain, deploy to a VPS, then secure, harden and optimize the code. Opus 5.5 at "medium effort," which he says is his default, led the work. Sonnet ran over the comments inside a Workflow, a Claude Code feature for fanning a task out across many small model calls. A cheaper model for bulk classification and a stronger model for planning and code is a common split. Our post on Claude Opus 5.5 and Karpathy's Three.js world https://explainx.ai/blog/karpathy-opus-5-lord-of-the-rings-threejs-world-2026 covers an earlier build of his with the same model family. What did the human actually do? Three things, by his telling: - Create a YouTube API key in the Google Cloud console so the site can fetch video metadata. - Boot a VPS and hand Claude its IP address. - Buy the domain name. Everything else, including hardening the server, was done by talking to Claude. That is the part worth pausing on. Giving an agent root-level access to a rented server is convenient and also a real risk surface; we return to that below. What does the site do? Per the post, HackerTube replaces uploaded thumbnails with the frame from the middle of each video, which strips clickbait art. It has a "ForYou" feed with no accounts, with state kept in the browser session, plus keyboard shortcuts. It also exports its data so another person can give it to an agent and build a variant. The framing is a reaction to a "wall of slop" on YouTube on a Friday night, in his words. Recommendation feed shown as a river of cards with one picked, illustrating how HackerTube curates videos The image is only an illustration of curated feeds. The underlying idea is simple: use a human signal, repeated recommendations in a technical community, instead of a watch-time algorithm. What did Hacker News say? The thread is large, so these are individual commenters' reports, not verified findings. - Cost: The first top-level comment, from user sc970, asks how many tokens or dollars https://news.ycombinator.com/item?id=50319736&ref=explainx a build like this takes. The post does not answer it. - Recency sorting: Several commenters, including mixtureoftakes and kiuiras, asked for a sort by upload date. beaconflint pointed out https://news.ycombinator.com/item?id=50319965&ref=explainx that "recency" is ambiguous, since a 2012 talk linked last month might be exactly the find. - Source code: vfclists and saikiran91 asked where the source is and whether it is open source. - Hit rate: reorder9695 said https://news.ycombinator.com/item?id=50319961&ref=explainx the front page was surprisingly close to what they actually want to watch, better than YouTube's personalized feed for them. - Terms of service risk: GodelNumbering said https://news.ycombinator.com/item?id=50319962&ref=explainx they had heard YouTube can be quick to ban third-party use of its content. - Playlists: twalichiewicz suggested publishing the curated list as auto-maintained YouTube playlists, calling the curation pipeline the valuable part. The mix is typical: enthusiasm for the idea, requests for features, and questions about reproducibility. What does the write-up leave out? Reading the post against the claims, the gaps are specific: | Gap | Why it matters | |---|---| | No token or dollar cost | You cannot budget a similar build | | No filter accuracy | How spam was detected and how well Sonnet classified recommendations is not measured | | "Quality" is subjective | The curation signal is repeated recommendation by a few users, as the post puts it, not a measured quality score | | No dataset license or YouTube API terms discussion | Reuse of comments and video metadata has rules | | One author, one run | It is an anecdote, not a reproducible method | Also note the number trail. The blog post describes the narrowing to about 40K comments that recommend a video, then 6,454 videos; the Hacker News submission itself says 39,834 comments. Neither gives the size of the starting comment dump, so treat any secondhand figure for it with care. Why a human-signal feed instead of an algorithm? The design choice is the interesting product idea. Recommendation systems on large video platforms optimize for watch time, and many users report that this drifts toward repetitive or low-quality material. HackerTube takes the opposite approach: a video earns a place because a person in a technical forum took the time to recommend it in a comment, sometimes years ago. That signal has strengths and limits. It favors talks, lectures and long-form explanations that people want others to see, which is why commenters noted lectures by well-known computer scientists on the front page. It also reflects the tastes of a specific community, mostly programmers, so topics outside that bubble are thin. And because the pool is built from a fixed dataset, it ages: a commenter on the thread asked whether the site will be updated with new videos and comments, and the post as we read it does not promise that. For builders, the transferable lesson is that "curation by community endorsement" can be built from any public text corpus where people link to things: forum threads, issue trackers, newsletters, even changelogs. The hard part is no longer the code. It is deciding what counts as a recommendation, how to weigh repeated mentions against one enthusiastic comment, and how to avoid rewarding spam. Karpathy's post says the filtering removed spam and kept comments that recommend a video, but gives no precision or recall figures, so anyone copying the idea should measure their own classifier on a hand-labeled sample before trusting it. What can you copy from this? The pattern is more transferable than the site. 1. Start from a public dataset. The Hugging Face comment dump removed the scraping step. 2. Use a cheap model for bulk labeling and a strong one for orchestration. That is the Sonnet-in-a-Workflow pattern. See our Claude Code desktop browser coverage https://explainx.ai/blog/claude-code-desktop-browser-built-in-july-2026 for the app he used. 3. Keep the human tasks to credentials and money. API keys, servers and domains are where you should stay in control. 4. Ship without accounts. No login means no user data to protect, which shrinks the security surface. 5. Publish an export. A data file lets others fork the idea without your code. The same logic runs through his LLM wiki pattern for agent memory https://explainx.ai/blog/karpathy-llm-wiki-pattern-agent-memory-guide-2026 . A starter prompt in that spirit, which is our own wording and not from his post: Download the Hacker News comments dataset from Hugging Face. Filter spam, keep only comments that recommend a YouTube video, group videos into topics, and build a static site with a keyboard-navigable grid. Do not add user accounts. Ask me before touching any server, key or domain. Is it safe to let an agent harden your server? Karpathy says he gave Claude the server's IP and had it secure, harden and deploy the site. That worked for a single-purpose site with no logins. For anything holding user data, treat it differently: review the changes, keep SSH keys and secrets out of the agent's reach, and snapshot the VPS first. If you run agents with shell access, a guardrail layer helps; AgentBeam https://agentbeam.com/?ref=explainx is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions. Why does this story matter beyond one site? It lands in the middle of a debate about what "programming" means now. Karpathy's earlier posts, and the vibe coding explainer https://explainx.ai/blog/what-is-vibe-coding-explained-2026 , frame natural-language building as a real workflow. His move to Anthropic, covered in our report on him joining for pre-training https://explainx.ai/blog/andrej-karpathy-joins-anthropic-pre-training-2026 , adds a note of interest to his choice of tools. His public Claude Code guidelines https://explainx.ai/blog/karpathy-claude-code-guidelines-andrej-karpathy-skills show how he asks agents to behave. The takeaway is not that anyone can replicate a famous engineer's judgment in two hours. It is that a small, opinionated product built on public data is now a weekend project, and the open questions, cost, filter accuracy and licensing, are where the real work moves. Related reading Details are accurate as of November 2, 2026, from the author's post and the Hacker News thread. We did not rebuild the site.