Why Hasn't Codex Shipped Its Own Claude Design?为啥 Codex 还不推出类似 Codex Design 的产品？

Why hasn't Codex launched a product like Claude Design yet?

Anthropic recently launched Claude Design, and besides coding it's the agent I use most and have recommended many times. It's genuinely impressive: describe the app you want in one sentence, and it directly generates an interactive prototype where every click responds — if you don't look closely, you'd think you were operating a real app.

A reader asked: why hasn't Codex launched something like Codex Design yet?

Simply put, GPT-5.5's model capability isn't there yet. But to explain why clearly, you first need to understand a key distinction.

The two layers of an agent: model and harness

Many people lump Codex, Claude Design, GPT-5.5, and Claude Opus 4.8 together, but they're actually two completely different layers.

Claude Design and Codex are the "product layer," what the industry calls the harness — prompts, toolchains, UI interaction flows, all engineering-level stuff. Claude Opus 4.8 and GPT-5.5 are the "model layer," the brain that actually does the work.

An analogy: the harness is the kitchen, with pots and pans (tools) and recipes (skills); the model is the chef. Swap the chef in the same kitchen and the dishes turn out completely different.

Once you understand this distinction, everything that follows makes sense.

The harness isn't the barrier

The harness layer of Claude Design isn't technically complex. With a bit of effort you can reverse-engineer it — the prompts and tool code are almost all obtainable. I've already done this; the result is baoyu-design (https://github.com/JimLiu/baoyu-design), which lets you run Claude Design's skill on other models. There's no secret on the engineering side.

What really creates the gap is the model behind it.

High-fidelity interactive prototypes are hard because of the model

The name "Claude Design" is easy to misunderstand — people assume it delivers static design images like Figma or Photoshop. In fact what it delivers goes further than Figma: a high-fidelity interactive prototype that merges the design and the prototype — you don't just see the design, you can directly operate it.

This demands a lot from the model.

Here's an example. Say I want to build a client similar to X/Weibo. Getting a model to draw a nice-looking static interface — many models can do that. But making that interface interactive is far more complex: switching between different timelines, displaying different types of tweets (text, images, video), turning the heart red on a like, removing a deleted tweet from the list, clicking into a detail view from the list and back while preserving state.

To pull this off, the model has to think through the entire data structure and state management before it even starts drawing the UI: what a tweet looks like, how many kinds of timelines there are, what state each button is currently in, how the states link together. This is systems architecture work, not UI-drawing work.

What Claude Design demands of a model is excellent UI/UX design ability combined with systems architecture design ability at the same time — missing either one and the result suffers badly. This is also why I've previously argued against producing plain HTML mockups alone — that's just static UI design, without UX interaction baked in.

If you're able to, try testing it yourself and see how it feels. For example, use this prompt:

Design a X Client for Mac, similar to Tweetbot for Mac from Tapbots

Giving the same prompt to Codex also produces something — viewable, and simply interactive. But comparing the two reveals the gap: the list scrolls, but the sidebar isn't clickable; the like button doesn't respond. It takes several rounds of back-and-forth iteration to reach a barely passable level.

What Claude Design produces is completely different. Switching from the timeline to the notifications page, clicking into details from a list and back — it's smooth throughout, and state is preserved the whole time. If you don't look closely, you'd really think you were operating a highly polished app, even though all the data is mocked.

Claude Opus 4.8 has clearly undergone extensive training and optimization for scenarios like this involving design and architecture.

The output is code itself

Look at what Claude Design produces, and pay attention to the data.jsx file inside it. It defines the entire design's data structure very clearly, mocks up a complete dataset based on that structure, and then builds the UI on top of that data using React.

The design output itself is code (React, CSS, JSON), not Figma or PSD files — any developer who gets it can immediately see the button's corner radius, the primary color, the spacing, and implement it directly in their own stack. A later design change? A quick git diff shows exactly what changed. The communication overhead between design and development is reduced to a minimum.

To put it more precisely, the communication overhead between the design agent and the development agent is now very low. Nowadays it's a human directing an agent to design, and a human directing an agent to write code.

How to make the most of Claude Design

Many people don't know how to use Claude Design well. It's actually a bit like vibe coding: you have a basic idea, let it produce a first version, then use chat to direct the agent to help you revise it — after a few adjusted versions, your own thinking becomes clear.

The whole adjustment process is remarkably magical, with a kind of "speak and it shall be done" feeling — however you want it changed, it can pretty much deliver. That's also why I'm so hooked on Claude Design right now — the feedback loop is so fast it's addictive.

Here's one more small tip: don't state overly specific requirements — instead, state your goal, what you're aiming for, and let it improvise freely. This often produces better results, since it has been trained on nearly all publicly available UI designs.

Back to the original question. Codex hasn't launched a similar design product because GPT-5.5 still can't handle this work. Many models can draw a good-looking interface, but the hard part is thinking through the data structure, state management, and interaction logic before you even start, then delivering a complete, interactive prototype in one shot.

Right now, only Claude's models can do this. As for how long the lead will last, it depends on how fast OpenAI's or other companies' future models evolve.