Karpathy Built His Own YouTube With Claude Code in 2 Hours. Here Is the Workflow You Can Copy.

Friday, 11 PM. Andrej Karpathy opens YouTube to wind down.

What he gets is, in his own words, “an enshittified wall of slop.”

Screaming thumbnails. The same tired topics. A feed built for the median viewer that does nothing for Karpathy.

💥 Master AI & Tech Skills in 3 Months!
🔥 Get Up to 50% OFF | ⏰ Limited-Time Offer
👉 Enroll Now & Start Learning

So on Saturday morning, Karpathy did what most developers only tweet about.

He built his own YouTube.

About two hours later, HackerTube was live: 6,454 videos that Hacker News users kept recommending between 2007 and 2026, sorted into 150 topics, with no login.

Karpathy says he no longer writes or runs the code himself. Claude Opus 5.5, working inside Claude Code, did that part.

Most coverage stopped at “AI built a website in two hours.” The useful part is how the work was split, because that split is a pattern any developer can copy this week.

Even without paying for Opus.

🎯 What HackerTube actually is

HackerTube is not a list of links. HackerTube is a working YouTube-style site with:

• Videos picked by humans, sorted by AI. Every video was recommended by real people in Hacker News threads.
• Honest thumbnails. The site shows a frame from the middle of each video instead of the clickbait image the uploader chose.
• A “For You” feed with no account. Recommendations live in your own browser session.
• Keyboard shortcuts.
• A full data export, so your own agent can build a custom version.

That last feature matters later. Keep the data export in mind.

🧠 The workflow, step by step

Karpathy ran the Claude Code desktop app on a Mac, with Opus 5.5 at medium effort as the default. Here is the order the work happened in, based on his blog post.

Step 1: One plain-English goal. Karpathy asked Opus to write a YouTube clone from videos recommended on Hacker News. No spec doc. No wireframes.

Step 2: Get the raw data. Opus downloaded the full Hacker News comment archive from a public Hugging Face dataset. The archive holds about 43 million comments from 20 years.

Step 3: Cheap filtering first. Claude used URL matching, regular expressions and spam rules to throw out every comment that clearly was not a YouTube recommendation.

Step 4: The smart model, only where needed. The comments that survived went to Sonnet, inside a Claude Code Workflow, to judge which ones were real recommendations. That left roughly 40,000 comments pointing to 6,454 videos.

Step 5: Build, ship, secure. Claude wrapped the data in a YouTube-style interface, deployed the site to a server, then hardened and optimized the setup.

Plain code did the bulk work. The model only judged the leftovers

And Karpathy himself? He did three things by hand:

1. Created a YouTube API key in Google Cloud
2. Booted a VPS
3. Bought the domain

That is the entire human part.

📊 The detail most headlines got wrong

“Claude read 43 million comments” makes a great headline.

That is not what happened.

Karpathy later clarified that the archive was cut down with plain code (URLs, regex, spam removal) before any language model saw it. Sonnet only judged the leftovers.

This detail is the real lesson, and the reason the build was fast and affordable:

Boring code does the bulk work. The expensive model does the judgment work.

Most developers who burn through their Claude limits do the opposite. They paste everything into the smartest model and hope.

💡 The pattern you can steal

Look at what stayed human in Karpathy’s build: an API key, a server, a domain.

Every one of those three involves money, identity or access. Everything else was handed off.

Get Ayesha Mughal’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

That gives you a clean rule for your own projects:

• Keys, payments and accounts stay with you.
• Filtering, sorting and cleanup go to plain scripts the agent writes.
• Judgment calls go to the model, in small batches.
• Building and deploying go to the agent, with you reviewing.

✅ The free version: build your own “Tube” this weekend

Karpathy used Opus 5.5, which needs a paid Claude plan. You can run the same workflow for $0 with three swaps.

Be clear on one thing first: the free version copies Karpathy’s workflow, not his brain. Opus 5.5 is the reason his build took two hours. A free model will get there slower, with more babysitting.

Swap 1: Keep Claude Code, change the brain.

Claude Code is the agent. The model behind the agent can be swapped. Ollama (version 0.14 or newer) speaks the same API format Claude Code expects, so one command connects the two:

ollama launch claude

Ollama asks you to pick a model, configures Claude Code and opens Claude Code for you. Prefer a manual setup? Set two variables and run any model you have pulled:

export ANTHROPIC_AUTH_TOKEN=ollama
export ANTHROPIC_BASE_URL=http://localhost:11434
claude --model qwen3-coder

If you read my Codex CLI free setup, the idea is the same: a strong agent running on a free model.

Want results closer to Karpathy’s without Opus prices? Claude’s $20 Pro plan includes Claude Code with Sonnet. That is the best value path. The free path is for learning the workflow first.

Swap 2: Skip the 43 million comments.

You do not need to repeat Steps 2 and 3. HackerTube’s data export already includes the video records, labeled mentions, topics and related-video data, plus a DATA.md guide written for agents.

The data uses the ODC-By 1.0 license, so credit HackerTube on your site. YouTube’s own metadata is not included, so your agent will use the video IDs and YouTube’s free API quota to fetch titles and stats.

Or point the same workflow at a different source: your favorite subreddit, a newsletter archive, your own bookmarks.

Swap 3: Free hosting instead of a VPS.

The finished site is mostly static pages and JSON. GitHub Pages, Vercel or Cloudflare Pages will host that for free. No domain needed to start.

The prompt to paste into Claude Code:

OBJECTIVE
Build a YouTube-style video discovery site from the HackerTube
data export in ./data. Read ./data/DATA.md first.
CONSTRAINTS
- Static site only: HTML, CSS, vanilla JS. No backend, no accounts.
- Store likes and watch history in localStorage.
- Write plain scripts for all filtering and sorting.
  Do not process the data by reading it yourself.
- Credit HackerTube (ODC-By 1.0) in the footer.
OUTPUT FORMAT
- A ./site folder ready to deploy on GitHub Pages.
- A README with setup and deploy steps.
GOALS
- Topic pages, search, and a "For You" feed based on local history.
- Keyboard shortcuts: J/K to move, Enter to play.
- Works well on mobile

Notice the constraint in the middle. That one line forces the free model to copy Karpathy’s best move: scripts for the bulk work, the model for the thinking.

If this prompt saves you a weekend, follow me. I share one free AI coding setup every week, and Thursday’s goes deeper on this exact Ollama setup.

🚨 What the free version will not give you

Honest trade-offs, so nobody feels tricked:

• Speed. Karpathy’s two hours came from a top model. Free models are slower and need more back-and-forth. Plan for a weekend, not a morning.
• Reliable file edits. Smaller local models sometimes fumble tool calls. If edits keep failing, switch to a bigger model or to one of Ollama’s cloud models, which come with free usage limits.
• Hardware. Local models eat RAM. On a 16GB laptop, stick to smaller coding models or use the cloud option.
• Security work. Karpathy had Claude harden a real server. A static site on free hosting has far less to secure, which is one more reason to start there.

📌 The real takeaway

The headline says AI built a website in two hours.

The real story is smaller and more useful: one person kept the keys and let the agent do everything else.

You can run that same split tonight, on a free model, with Karpathy’s own data.

So tell me: what would you build your own “Tube” from? A subreddit, a newsletter archive, your bookmarks? Drop it in the comments. I read every one.

If this saved you $20 of tokens, hold that clap button for a second. 👏

Hit Follow for one free AI coding setup every week. Next up on Thursday: Claude Code with no router, Windows included.

Find me on LinkedIn for quick findings between posts.

I’m Ayesha Mughal. I build AI agents for clients and write about the tools that actually ship. Need an agent built? ayeshamughal2162@gmail.com

Someone has to test the AI hype before your codebase does.

Thank you for being a part of the community

Before you go:

👉 Be sure to clap and follow the writer ️👏️️

👉 Follow us: Medium

Disclosure: This post includes affiliate and partnership links.