Why Does the New Claude Code UI Look More Rudimentary?

Claude Code rewrote its desktop app in April. Since its release, public opinion has been split.

Many people’s most direct reaction was: Why does it look “cheaper”? Fewer icons, more whitespace, colder colors, overall tone like an alpha-stage engineering tool without finished visuals.

“Looking more rudimentary” isn’t the problem. The new Claude Code UI did indeed shrink fonts, reduce icon usage, and reduce page colors. And Chat / Cowork / Code all changed.

But why?

The New Additions

To answer this, let’s look at what Anthropic’s product team added to Claude Code.

The new desktop Claude Code reorganized the entire interface into a few things:

A left sidebar centrally manages multiple sessions, filterable by status, project, environment. Sessions are automatically archived after PRs are merged or closed (configurable in Settings);

The right side is a set of draggable panes: terminal, preview, diff viewer, chat, file editor—arrange them into a grid as you like;

Added a side channel called side chat (/btw), opened with ⌘+;. It reads context from the main thread but doesn’t write back to it; Information density is cut into verbose, normal, summary levels, you decide how much of Claude’s reasoning process to see;

The bottom has a usage button, showing context window usage and session usage simultaneously.

claude code desktop redesign

Connecting these details, the new UI’s focus has shifted from the chat box to the task panel. Anthropic itself defined the user’s position in the announcement as “orchestrator seat”—you become the scheduler managing multiple parallel tasks. In terms of product, Anthropic seems to emphasize that users should be using the product around “orchestration”.

“Orchestration” adds a lot of product complexity.

I believe many people haven’t noticed that Anthropic’s Code already has so many complex features.

In fact, Anthropic made all these UIs look the same: more restrained, more structured, visually colder.

Why couldn’t it be a fancier orchestration interface?

Why AI Really Needs This UI

The answer is simple: Anthropic’s own models aren’t great at writing UI — and this fact is conversely deciding what the product should look like.

A widely circulated blog post《Every AI Coding Agent Is Blind to Your UI》directly calls out Claude Code, Cursor, Copilot, Windsurf—all file-first AI coding agents share the same blind spot: they can read source code, trace imports, run grep, but cannot see runtime. Specifically, computed style, component ID, responsive branches, stacking context, hover/focus/animation things are not in the source code. They only exist in that instant after browser rendering.

https://frontman.sh/blog/introducing-frontman/

The author defines this as an “architectural gap”, and makes a rather strong judgment: stronger models won’t fix this problem. GPT-5 won’t fix it, Claude’s next generation won’t either, because they are all reading the same source code that doesn’t contain runtime information.

So, what Anthropic is doing is an “AI-friendly UI”:

• Less visual noise, less likely for the agent to mess with places it shouldn’t;

• Clearer functional zones (sidebar, pane, editor, preview), it knows better which context it’s working in;

• More explicit state (context usage, session status, diff), it can make decisions based on facts, less guessing;

• Context can be artificially isolated (structure like side chat), the main task’s reasoning chain is less likely to be polluted by temporary issues.

Structurally, every panel looks like an independent “Small App”:

Hwang@hwwaanng·May 26, 2025 Normal data structure, layered abstraction, implementing a feature often involves different files across different data layers, as shown in Figure 1. A reasonable architecture can integrate different business abstractions, leading to good results for project quality and requirement iteration.

But Vibe coding needs the structure in Figure 2. Minimize interaction between files of one feature and another, this way AIShow more22143.6K

Reading this, it’s easy to think I’m opposing “pretty UI”. Actually, it’s more like pretty UI is being recalibrated under new premises: complex floating layers, shadow overlays, implicit animation states are details for human users, but cognitive noise for agents. When your product has a “colleague who can’t see the web”, UI optimization goals are re-weighted.

“Seeing like an agent”

Models “can’t see”, so UI must reduce pressure for them. Anthropic is indeed using the principle of “designing from the model’s perspective” internally.

In another blog post. On April 10—4 days before the UI redesign went live—Anthropic published an engineering blog titled《Seeing like an agent: how we design tools in Claude Code》. It discusses the Claude Code team’s methodology when designing tools: think from the model’s own perspective, where it has visibility on information and where it doesn’t, which tools are useful at first but become constraints after the model gets stronger.

The article ends with a very direct phrase—“try to see like an agent”. https://claude.com/blog/seeing-like-an-agent

This methodology was originally only used to explain tool design: whether to make AskUserQuestion, Task tool, Grep. I admit this is inference, product design philosophy is rarely written explicitly in announcements, it’s usually hidden in the similarity between two things. Constraints on the outside, intent on the inside, matches perfectly. See like an agent.

Inspiration

Reduce unnecessary page styles, define clear and refined rules, limit AI’s autonomous play, and you can have better collaboration with the Agent, making team iteration more efficient.

However, perhaps the software industry will become more boring because of this.

Does Aesthetics Still Matter

A year ago, people were dissing blue-purple gradients as AI Slop, today people say something like Anthropic’s beige is the new AI Slop

Paul Bakaus@pbakaus·May 3this is 2026 AI slop.

ironically, in part introduced by anthrophic’s frontend-design skill, which replaced one set of model defaults…with another (fraunces for editorial anyone? warm brown italics?)

it’s incredibly hard to avoid, and build software and skills that unlock trueShow more392034550K

The standard for “beautiful” is actually being constantly rewritten. When everyone reaches the average level, the original beauty is no longer beautiful.

Does aesthetics matter? Of course it matters. The new UI really looks colder, but many people don’t like it.

Today, whether Claude Code or Cursor / Codex, all support model reading Preview capabilities. In the foreseeable future, model capabilities will have better vision, after all the specialized student Gemini has already shown good visual understanding capabilities.

Software is not designed for AI, graphical user interface is always for humans. Design has always been “human-centric”, not LLM/Agent or AI-centric.

After all, the ones paying are always people like you and me.