Karpathy Defaults to Medium effort; the Terminal Faction Is Still Writing Rebuttals in Vim

Andrej Karpathy's Claude Code setup was posted by Yuchen Jin from a HackerTube blog.

The setup is short: Mac desktop version of Claude Code; default Opus 5.5, Medium effort. Yuchen Jin's comment is even shorter: the terminal is the wrong interface; the desktop app is the best interface today.

That sentence set the comments section on fire.

A reply pointed out that in the same passage Karpathy no longer reads much code or runs many commands. The flow in the screenshot is: prepare an API key, VPS, and domain, then hand the IP to Claude and have it update, deploy, secure, harden, and optimize the code. The human is responsible for opening the entrance; the rest is handed to the agent. Under this premise, the terminal is indeed no longer a mandatory entry point.

The terminal faction has not left the stage. Some say Claude has fewer restrictions in the terminal; some say scripting still favors the terminal; some are still drafting rebuttals in Vim; and someone posted a black-and-white handshake photo captioned "Still in the terminal, nobody told me it was over."

The advantages cited by the desktop faction are very concrete: reviewing diffs, switching between multiple sessions, and using project features to coordinate multiple workflows. A more detailed point is that when the agent needs a human to make a decision, the desktop app can show the proposed change, evidence, and undo in one place. Simply wrapping the terminal scrollback in a window does not provide these advantages.

The disagreement over model choice is more interesting. Karpathy sets Opus 5.5 at Medium effort as the default. Someone in the comments joked: Karpathy defaults to Medium, while others run Max to have the model rename a variable. Some also asked what tasks are worth turning effort up for. When asked this question, Yuchen Jin replied: I have an intelligent router in my head.

A netizen said DeepSeek 4.1 is about the same as Opus 5.5 for writing code, while Opus 5.5 is stronger for long-horizon tasks, so they give planning to one model and execution to another. There is no unified answer here, only task stratification.

On the surface this debate is about app versus CLI; underneath it is about the division of labor in workflows. For everyday tasks, use the desktop app to reduce coordination costs; leave scripting and remote environments to the terminal. Karpathy's default gives a signal: the everyday entry point for coding agents is shifting from the terminal and maximum reasoning to the desktop app and medium reasoning.

Are you still in the terminal, or have you already moved the agent into the desktop app?