Skip to content

Andrej Karpathy shipped a video site in two hours

Blog cover with the official CommandLyne logo: the headline Karpathy shipped a video site in two hours in white type beside a finished app window with a blue play button and a video grid, fed by three step blocks on a dark navy grid.
CommandLyne TeamAI Product & Engineering

Nov 2, 20265 min read

Ask a team to build a video platform and you are budgeting a quarter. Andrej Karpathy did it on a Saturday morning without opening an editor.

On October 31, 2026, Karpathy published HackerTube (opens in a new tab), an alternative YouTube front end built entirely from videos the Hacker News community has recommended over the years. He described the brief he handed the model: "Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News."

What the agent actually did

The scope is the interesting part. This was not a scaffold or a clickable prototype.

StepWhat ran
Source dataDownloaded the full Hacker News comments dataset
FilteringRan Sonnet across the comments inside a Claude Code Workflow
CleanupRemoved spam and low-signal recommendations
Output40,000 comments narrowed to 6,454 recommended videos, from 2007 to 2026, across 150 topics
InterfaceBuilt a YouTube-style UI with keyboard shortcuts
RecommendationsCustom "For You" feed with no user accounts, storing state in the local browser session
DeliveryBought a domain, pushed to a VPS, secured, hardened and optimised it

Elapsed time: roughly two hours.

WHAT THE HUMAN DID            WHAT THE AGENT DID

1. Created a Google Cloud      Downloaded the dataset
   API key for YouTube         Filtered 40K comments
   metadata                    Ran Sonnet over them in a Workflow
2. Booted a VPS                Built the UI and the feed
3. Bought the domain           Deployed, secured, hardened, optimised

Three steps on the left. Everything else on the right.

The part worth reading twice

Karpathy added a note to the post that is more useful than the launch itself. He works in Claude Code on the desktop, with Opus 5.5 at Medium effort as his default. Then this:

"I don't actually do anything resembling traditional engineering anymore (read/write code, run commands in a terminal, or etc.), I only talk to Claude and have it do everything."

He kept the three tasks that need a human identity to exist: an API credential, a server, and a domain. Everything between them was a described outcome.

Why this shape travels to enterprise work

A side project and a quarterly business review are not the same thing, but the shape of the work is. Both start with a described outcome and a set of systems that have to be read, reasoned over, and written back to.

That shape is what CommandLyne (opens in a new tab) is built around. You describe the outcome in plain language and it becomes a multi-step workflow across the tools you already use, without flowcharts or scripting. In the Agents and Workflows (opens in a new tab) editor, a request like "every Monday at 9 AM, summarize last week's Jira sprint activity" becomes a scheduled workflow spanning Jira, Drive, Confluence and Slack.

Three capabilities from the Karpathy build map onto it directly:

  1. The data pull is a task, not a chore. Long research and file processing runs in Background Tasks (opens in a new tab), where active and recent work is visible without blocking the conversation.
  2. The output is an artifact. Canvas (opens in a new tab) is a side workspace for structured and interactive content beyond a normal chat reply: flowcharts, architecture diagrams, data visualizations, UI mockups and decks.
  3. The repeat runs itself. Scheduled Tasks and Routines (opens in a new tab) automate recurring work like daily email summaries, reports and scheduled research.

Where governance enters

Karpathy's build had one user, one server and no review board. Enterprise work carries a different set of expectations, and that is where a described workflow needs a governed one around it.

CommandLyne pairs the automation with controls that hold up in an audit:

  • Scoped tool access, so an agent reaches only the tools each task needs
  • Human sign-off, so sensitive actions wait at an approval checkpoint before they execute
  • Full audit trail, recording every prompt, data query and tool execution, plus who approved it
  • VM-level isolation, giving every user a dedicated virtual machine with no shared runtime memory

The result is the same described-outcome workflow, with the parts your security team will ask about already in place.

Assisted codingDescribed-outcome automation
Who writes the stepsYou, one at a timeThe agent, from the outcome
What you reviewEvery line as it appearsThe artifact, then the approval
Where it runsYour editorBackground tasks and scheduled routines
What happens on repeatYou run it againThe routine runs on its schedule
Who can see itYouScoped access, with a full audit trail

What to take from it

The lesson is not that a weekend project replaces an engineering team. It is that the decomposition step, the part where a goal becomes an ordered plan across systems, has moved from the human to the agent. Once that step is handled, the ceiling on what one person can ship stops being their typing speed.

Start with the work you already repeat. Describe the outcome, let the agent build the plan, and keep your hand on the approval step. The two-hour build is the far end of the curve. The Monday morning report is where it starts paying you back.


Try CommandLyne to turn a described outcome into a governed, multi-step workflow, or book a demo and we will build the first one with you.

Try CommandLyne (opens in a new tab) | Book a demo (opens in a new tab)

See CommandLyne in action

Book a demo to see how CommandLyne brings your agents, models and workflows together under one governed workspace.