Andrej Karpathy builds HackerTube in two hours using Claude Code

From 'Slop' to a Custom Clone

Andrej Karpathy, a former OpenAI founding member and Tesla AI lead, expressed frustration with the current state of YouTube, describing it as an "enshittified wall of slop." Seeking a cleaner alternative, he decided to build a custom site that only features videos recommended by the community on Hacker News.

He gave Anthropic’s Opus 5.5 a high-level instruction to create a YouTube clone based on these specific recommendations. According to Karpathy’s blog post, the project was completed in approximately two hours. The resulting site, HackerTube, organizes content from 150 different topics.

The project serves as a practical example of how AI agents can now handle the heavy lifting of data collection and site construction. Karpathy noted that he no longer performs traditional engineering tasks like writing code or running commands in a terminal. Instead, he simply talks to the AI and lets it execute the work.

How the AI Handled the Build

The technical workflow relied on a combination of models and features within the Claude Code desktop app. Karpathy used Opus 5.5 on 'Medium effort' as his primary driver for the project. To clean the data, the system ran the Sonnet model over the comments in a specific Claude Code Workflow to remove spam.

The AI was responsible for downloading the Hacker News comments and filtering them down to a set of 40,000 relevant recommendations. It then populated the site with 6,454 videos. Karpathy noted that the site includes specific features like keyboard shortcuts and a way for users to export data to build their own versions.

Beyond just writing the code, the AI managed the deployment lifecycle. Once the server was live, Karpathy provided the IP address and instructed the agent to update, deploy, secure, harden, and optimize the code. The AI completed these complex infrastructure tasks automatically based on his verbal requests.

The Human Role in Agentic Work

While the AI performed the bulk of the engineering, Karpathy still had to perform a few foundational steps by hand. These included creating a Google Cloud API key, booting up the necessary server, and purchasing the domain name. These are the only three manual actions he listed in his report.

The project highlights a growing trend where the human role shifts from 'builder' to 'orchestrator.' By using a desktop app designed for agents, Karpathy was able to bypass the need to interact with a terminal. This suggests that the interface through which humans interact with AI is becoming the primary lever for productivity.

This approach aligns with broader industry movements where users are beginning to rewrite apps through agents. For example, Stripe CEO Patrick Collison recently shared that he allows users to rewrite his apps using an agent. Karpathy’s experience shows that this is moving from a theoretical possibility to a functional reality for complex web projects.

Cost and Effort as Levers

A key detail for those interested in the business of AI is Karpathy’s use of 'Medium effort' as a default setting. In a landscape where API costs are often tied to usage, the effort tier acts as a dial for the user to control the intensity of the AI's reasoning and the resulting cost.

Karpathy chose Opus 5.5 rather than the highest possible reasoning setting, suggesting that not every task requires the most expensive model. This mirrors observations from other experts who argue that most coding tasks do not need the top model at its maximum effort level. It points toward a more nuanced way of managing AI expenses during development.

The project also demonstrates the power of the 'agentic harness,' which refers to the software framework that governs how an AI interacts with a computer. Karpathy’s success suggests that the surface—the app and its capabilities—governs autonomy as much as the underlying model does. This shift may change how companies choose which AI tools to adopt for internal development.