01
Read the Hacker News archive
The workflow started with public Hacker News comments spanning 2007–2026 and looked for YouTube links discussed by the community.
Build story · Primary-source guide
How Andrej Karpathy built HackerTube with Claude
Karpathy says he directed Claude Opus 5.5 to turn years of Hacker News video recommendations into a working YouTube-style product in about two hours. The interesting part is not only the speed: the project joins data extraction, model-based filtering, product design, and deployment in one agent-directed workflow.
Accuracy note: the approximately two-hour build time is Karpathy's own account, not an independently timed benchmark. His post identifies Opus 5.5 as the primary coding agent and describes a Sonnet workflow used in filtering.
The pipeline
02
Separate recommendations from noise
Claude filtered spam and judged the role of each link, narrowing the material to about 40,000 recommending comments rather than treating every mention as an endorsement.
03
Create a structured video library
The resulting catalogue contains 6,454 videos across 150 topics, with summaries and recommendation context layered over YouTube metadata.
04
Build, deploy, and open the data
The agent built the interface and helped deploy it to a VPS. The data export was packaged so another agent could turn it into a different interface, newsletter, or analysis.
What Karpathy still did himself
In the process note, Karpathy says his manual work was limited to creating a Google Cloud API key for YouTube metadata, starting a VPS, and buying the domain. He then gave the server address to Claude and asked it to update, deploy, secure, harden, and optimize the application.
That makes HackerTube a useful example of agent-directed software development: the human supplies taste, scope, credentials, and judgment while the agent carries out most implementation work.
Read the sources, then explore the result
The launch post is the best source for the creator's process. The Hacker News thread records feature requests, criticism, and rapid changes made after launch. Explorer adds a smaller, source-visible path through the public data rather than replacing either source.