Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

Designers Wired Claude Code and ChatGPT Together and Cut the Development Stage by 60%

Asking an AI to "make it look nice" with no foundation gives you bland results, so the design team gave the model a framework: prototypes, Figma mockups, and a logo. Where Claude had already built the prototypes, the design was applied to them in a single pass, and the development stage shrank by roughly 60%. The remaining 40% — small things like a misaligned button — is easier to finish by hand. The takeaway: first give the AI the stack and constraints, then ask for creative work on a finished base.

AI-processed from Habr AI; edited by Hamidun News
Designers Wired Claude Code and ChatGPT Together and Cut the Development Stage by 60%
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The AI replaced dashboard images with SVG, and those had to be fixed by hand
Key metric #2The hardest part was building tables with filters, modals, and roles, not the dashboard
Key metric #3Claude Code coded a landing page from 1920 and 360 pixel mockups in minutes

Almost every designer has asked an AI to "make it look nice" and gotten a bland, template-like result. In the third part of a series on Habr, a designer from a product team shows how to set up the work differently. The headline number: in a project where the service prototypes had already been built in Claude, the development stage shrank by roughly 60%.

"Development" here means turning mockups into working markup and connecting the backend. The team handled the remaining 40% manually: moving a button through an AI costs as many tokens as it does time. Claude Code is Anthropic's tool for working with code, and MCP is the way the model gets access to external services such as Figma.

This is personal experience across three projects, not a laboratory comparison, but it contains many practical details.

Option 1: prototypes already exist

The first project was a service fully laid out as prototypes inside Claude. They were gray and impersonal, but they contained all the logic. The team had only two design mockups in Figma. The service had two directions: a personal account area for users and a public section for customers.

The personal account went smoothly. The AI was given the UI kit, colors, fonts, sizes, and a couple of behavior patterns, and it applied the styling to an existing system. The public section was harder: its prototypes had been drawn by hand in Figma. So the prototypes were first transferred to the AI: it was given access via MCP or API and a link to the frame. According to the author, this now works noticeably better than it did three months ago, even with a plain PNG.

The order of steps matters. First you set the constraints: the mockup is 1920 pixels wide, and sizes, spacing, fonts, and grid are taken from Figma. Then you check that the prototypes match the mockups. Only after that do you add the visual direction: "here are two mockups, apply this design to the remaining prototypes and use the UI kit." The stack (Vue, libraries, fonts, icons) was set in advance.

There were pitfalls too. The AI replaced some dashboard images with SVG files, although the team needed a live system with markup patterns. The developer explained that difficulty for a designer and difficulty for markup are different things. A flashy dashboard is simple to build. A modest table with filters, data loading, modals, states, and roles takes far more work. That's why it's better to agree on the page order with development early. Everything ready was immediately uploaded to GitHub, where developers reviewed the code.

Option 2: only a concept exists

In the second project there were no prototypes and no analytics. The author started with Claude Design: it produces pages, but the result looked like a prototype rather than a finished design. Ideas for the personal account, cart, layout, and mobile adaptation were usable, but not as creative work. They dropped it because it multiplied separate entities within the service.

Claude Code proved more systematic. It has separate pages, its own system, a readme, and an md file with the change history. That is context the model relies on as the project's knowledge base. Claude Code doesn't generate images itself, so for creative assets other models are connected via API: GPT, Kling for video, and Nano Banana. Codex from OpenAI works similarly, just inside the GPT ecosystem.

The token advice: a solid model such as Sonnet or Opus is enough for routine work. Top-tier models are brought in selectively for complex creative tasks, or they'll eat through the limits in a couple of sessions. Presentations, the author observed, come out worse than landing pages.

Option 3: from scratch, from logo to interface

The third case was an internal product: an idea, functionality, and a non-working prototype with tangled logic. A request to "make it look nice" with no foundation again produced a bland result. What worked was a chain that started with the logo:

  1. In ChatGPT the author described the product's ideology, received six logo variants, and picked one.
  2. Based on it, the model assembled a visual brand. The context in the chat was preserved.
  3. Then a landing page was made from the brand, and the main page interface was made from the brand and the landing page. The needed visual came out on the second iteration.

In the author's experience, GPT is good for interface visuals but tends to overload images with detail, so it is explicitly told "remove the excess." Fixes can be made selectively by leaving comments on the image. The basic Go tier wasn't enough, so they took Plus.

Everything was assembled in Claude Code, with a separate session for the landing page and for the service. It laid out mockups at 1920 and 360 pixels (the width of a computer and a phone screen) in a matter of minutes, and several animations were added in about fifteen minutes. For the animation, the author found a reference on Pinterest, asked for the video to be broken down frame by frame, generated similar images via API, and assembled the clip.

What follows from this

The author pays for two AIs: one produces the creative work, the other builds the markup. That way tokens aren't burned on a single model and the work runs in parallel. The limitations are honest: this is one person's experience with one team, and apart from the 60% estimate there are no precise measurements. Part of the work still falls to developers. The working rules are simple. Set the framework and stack first, then the visuals. Keep project context in files. Involve development right away. And don't ask an AI for "something nice" without a brand.

Story

Key Questions Answered

What are the key details on Option 1: prototypes already exist?

The first project was a service fully laid out as prototypes inside Claude. They were gray and impersonal, but they contained all the logic. The team had only two design mockups in Figma. The service had two directions: a personal account area for users and a public section for customers. The personal account went smoothly. The AI was given the UI kit, colors, fonts, sizes, and a couple of behavior patterns, and it applied the styling to an existing system.

What are the key details on Option 2: only a concept exists?

In the second project there were no prototypes and no analytics. The author started with Claude Design: it produces pages, but the result looked like a prototype rather than a finished design. Ideas for the personal account, cart, layout, and mobile adaptation were usable, but not as creative work. They dropped it because it multiplied separate entities within the service. Claude Code proved more systematic. It has separate pages, its own system, a readme, and an md file with the change history.

What are the key details on Option 3: from scratch, from logo to interface?

The third case was an internal product: an idea, functionality, and a non-working prototype with tangled logic. A request to "make it look nice" with no foundation again produced a bland result. What worked was a chain that started with the logo: In ChatGPT the author described the product's ideology, received six logo variants, and picked one. Based on it, the model assembled a visual brand. The context in the chat was preserved.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…