Skip to content
The Vibe Design Process
I'm Addicted To

The Vibe Design Process I'm Addicted To

A full-time designer used to be out of reach for a small team. With Claude Code or Codex, a design round now takes minutes – so you get a real brand from day one, and agents that build faster.

Screen Designs
Are Back

Two weeks ago I wrote that I thought screen designs were a thing of the past. The old way: a product manager writes requirements, a designer draws them, a developer builds them. With agents, I went straight to code. Then I built our mobile app in a weekend and designed the screens first. I haven’t stopped since.

This weekend I took it much further, on a new project, and I’m absolutely addicted to it. I call it vibe design: the same loop as vibe coding, only the agent doesn’t write your product. It draws it.

Straight to code

Describe the product, let the agent build it, fix what looks wrong.

Fast, and it looks like every other AI-built page. Each change of mind is a change in the code.

A designer or an agency

Brief, wait for the drafts, give feedback, wait again.

The real thing – and out of reach for most small teams as a full-time role.

my pick

Vibe design

An agent draws on a canvas. You look, pick and say what's off. A round takes minutes.

A designer's process at the speed of vibe coding. Nothing is built until the look is decided.

We all noticed that code became cheap. What I missed is that design rounds became just as cheap. In a coding agent like Claude Code or Codex, a new version of a logo, a homepage or an app screen takes minutes. So suddenly even a small team can work as if it had a full-time designer.

That gives you three things:

  • A real brand from the get-go. Not the look every AI-built page has.
  • Strong designs to build from. You decide what the product is while changing your mind is still free.
  • Agents that build faster. They copy a decision instead of guessing one.

One Weekend,
One Canvas

The project is Zawq, a new product I’m building: an AI content team that learns your taste. The name is the Persian and Urdu word for cultivated taste. So it had better look like somebody with taste made it.

It started the usual way. I let an agent one-shot a landing page, and it was fine. It was also a page you have seen a hundred times. Then, on Saturday at 16:50, I opened a design canvas in Claude Code and asked for “six really unique, strong, tasteful design directions”.

  1. 163

    boards on one canvas

  2. 79

    versions of that canvas

  3. ~77

    messages from me

  4. ~9 h

    of active work, across about 40 hours

  5. 0

    lines of product code

Before: the one-shot page, live until this Monday

After the weekend: homepage draft 8 – tap to scroll the whole page

By Monday morning the canvas held 163 boards on eight pages: a brand direction, a logo, a color system, eight full drafts of the homepage for desktop and phone, the main app screens, sign-in and error pages, even the emails.

And none of it is code. I typed about 77 messages. The agent drew. My active time was about nine hours, spread over the weekend. The one rule I repeated to it: nothing gets coded.

Two of the eight canvas pages, website and logo, folded into columns. Every round is still there, the first at the top.

How a Round
Works

The tool is Claude Design, inside Claude Code: a canvas where every design is a board. Under the hood each board is a small web page the agent writes, so it can draw anything a browser can show – even 3D you can turn with your mouse. The loop itself needs no special tool. It’s a conversation.

  1. 1

    I write a message

    Anything from one line to a long, dictated review. I say what I see and what I feel, not how to fix it.

  2. 2

    The agent draws

    It writes each board as a small web page and puts it on the canvas: six options when the question is open, three when we narrow down, one when we refine.

  3. 3

    The agent looks at its own work

    It opens every board, takes screenshots at desktop and phone width, and checks contrast and tap targets before I see anything.

  4. 4

    I look and pick

    By eye, fast, and by number: “Trim 2 is the one.” Usually within a few minutes.

  5. 5

    Nothing is deleted

    The new round goes below the old one. Oldest at the top, newest at the bottom. Then the next message.

Three things made this work, and I had to learn each of them during the weekend.

React, don’t specify. I’m not a designer. I can’t brief a logo. But I know in one second whether I like what I see. After twelve directions that only varied type and color, I wrote: “I feel like we lack a strong image.” After four more that felt alike: “really break out of the mold.” The agent answered honestly that it had reused one layout four times and only swapped the illustration. Sixteen directions missed before the seventeenth stuck.

Make the agent look at its own work. The first eight versions went onto the canvas without the agent ever seeing them rendered. Boards were cut off. So I told it: “You also need to take screenshots and really see if what you did has hit the mark.” From then on it checked every board before showing it.

Keep the history. Nothing on the canvas was ever deleted. That made changing my mind cheap. I had dropped a garden idea for a kitchen idea, saw the garden again in a later round, and took it back. In the morning I moved the moon away from the tree. In the evening I put it back behind it.

The Logo I Couldn’t
Have Briefed

The logo is the best proof. If you had asked me on Saturday afternoon what the Zawq logo should be, I would not have said “a bonsai in front of an orange moon”. Nobody could have briefed that. It took 64 candidates to find it.

  1. 1 · First logo round
    Sat 20:29 · rejected

  2. 2 · Second logo round
    Sat 20:37 · rejected

  3. 3 · Other names, to explore
    Sat 20:37 · rejected

  4. 4 · App icon “Harvest moon”
    Sat 23:24 · idea chosen

  5. 5 · The cloud tree
    Sat 23:32 · chosen

  6. 6 · Cloud tree 2
    Sat 23:41 · I see a bonsai in it

  7. 7 · Bonsai, in a pot
    Sat 23:53 · one of three

  8. 8 · After a reference picture
    Sun 00:28 · rebuilt in 3D

  9. 9 · Bonsai, logo version
    Sun 11:19 · “hell yes”

  10. 10 · Trim 2, rounder
    Sun 11:50 · chosen of six

  11. 11 · The final tree
    Sun 12:30 · rooted in the ground

  12. 12 · Final app icon
    Sun evening · 3D tree, orange moon

  13. 13 · Final mark
    Sun 23:00 · the logo

Thirteen stops out of the 71 boards on the logo page. Saturday evening to Sunday night.

The first twelve logos were flat marks next to the name. I wrote: “I don’t like these logos at all.” For a moment I even let it try six other names. I rejected those too and kept Zawq.

What worked came from somewhere else. The homepage already had small 3D plants on it. So I asked for app icons built from those parts, with no name at all. One of the six was a dark tree in front of a harvest moon. Three trees later there was one with foliage in levels, like clouds. The agent had listed “reads as bonsai” as its weak spot. For me that was the whole idea: a bonsai is a tree somebody shapes, patiently, with taste.

On Sunday at 11:19 I looked at the next version and typed: “hell yes, that’s the one.” The rest of the day was trimming. When it added more foliage, I took it out again, because it destroyed the clean structure. The flat logo you see at the end was drawn last, from the finished 3D tree – not the other way around.

A logo round used to mean a brief, a week of waiting and an invoice. This was one evening and one morning, and I never had to explain what I wanted. I only had to recognize it.

Decide It Before
It’s Code

Here is why this makes the build faster and not slower. Every decision below changed during the weekend. On the canvas each one was a single new board. In a built product each one would have touched many files.

Colors

First: About 20 nearly identical dark greens across the page

Then: 36 named color variables, eight of them greens

Light or dark

First: A theme switch on the page, and no light design behind it

Then: Dark only. The switch came out

Currency

First: A euro price and a currency toggle

Then: US dollars only

The app

First: Five screens of a classic dashboard

Then: Chat as the main screen, plus a side panel and a helper

Icons

First: A first set drawn by hand – fine from far away, not when you zoom in

Then: Standard outlines, with our own faceted shading inside

Section breaks

First: Eight different transitions between sections

Then: Two kept, six replaced by one landscape idea

The biggest one was the app. The first screens were a classic dashboard. I looked at them and saw the mistake at once: Zawq is an agentic product. You talk to it. So the next morning chat became the main screen. No route, no layout and no database table had to be rebuilt, because none existed.

Ask Zawq – chat as the main screen

Home – what needs you today

Review – read, grade, approve

App screens, as designs. The posts and numbers in them are sample content the agent wrote, and the icons have been redrawn since.

Late on Sunday I stopped all page work until the colors were locked. Homepage draft 7 used about 20 dark greens that were almost the same. They became eight, inside a set of 36 named colors, each pair tested for contrast. Draft 8 is draft 7 redrawn with those names only. If your app already has a style system, you know what cleaning that up in code costs.

The logo guide

Every version of the mark, moving and still, with sizes and clear space.

Tap the picture to open the whole board.

The color guide

36 variables with names, roles and the code for each.

Tap the picture to open the whole board.

The accessibility check

128 of 128 color pairs pass WCAG 2.2 AA – checked before a single component existed.

Tap the picture to open the whole board.

These boards are what the coding agents get. A style guide short enough for a briefing. The colors as code. The components with notes for the build. Every screen of the user flow as a picture. An agent that has to invent the look while building will guess, and you will correct it screen by screen. An agent that has the look only has to build it.

What’s Not
Built Yet

To be honest about where this stands: the weekend produced designs, not a product. The first piece is built since. On Monday the new homepage went live on zawq.ai. The app behind it doesn’t exist yet.

  1. Done: The brand

    Logo, colors, typography, icons, a style guide and a pack of brand files.

  2. Done: The website design

    Homepage draft 8, desktop and phone, plus sign-in, error pages, forms and emails.

  3. Done: The app design

    The main screens, as pictures. A first set, reworked once already.

  4. Done: The new homepage is live

    Since Monday, 5 October, on zawq.ai. The one-shot page is gone.

  5. Open: Building the app

    Not started. This is where I find out how much faster the agents really are with these designs.

  6. Open: Slow laptops and old phones

    The 3D scenes were only measured on my own machine. The figures for slower devices are estimates.

So “agents that build faster” is, for Zawq, still a promise. I saw it once already, when our mobile app was built from screen designs in a weekend. Now I’ll see it on a whole product, and I’ll tell you how it went.

It also wasn’t all smooth. Three logo rounds went into the bin. Three homepage drafts were so full of decoration that most of it came out again; I had to remind the agent, and myself, that a brand about taste can’t have that much going on. And one board kept a hidden 3D animation running until my machine was under full load. The agent found it and fixed it in about two minutes.

Your Vibe Design
Weekend

This works best on something new, before the first line of code. That can be a new product, or the next landing page or feature of the one you have. This is the order I’d follow.

  1. 1

    Start before the repo

    Open a canvas, not a codebase. Tell the agent what the product is and who it's for, and that nothing gets built yet.

  2. 2

    Ask for six, and for difference

    Six directions that really differ – not one layout with six color schemes. If they feel alike, say so and ask again.

  3. 3

    Find the one picture

    Type and color won't carry a brand. Look for a visual idea you can build everything on. Ours was a garden, then a tree.

  4. 4

    React, don't specify

    Pick by number. Say what bothers you. Bring references – a game, a poster, a website. You don't need the designer's vocabulary.

  5. 5

    Keep the history

    Never delete a round. The idea you dropped in the morning may be the one you come back to at night.

  6. 6

    Lock colors, then check them

    Name every color, let the agent test the contrast, then redraw the page with the named colors only.

  7. 7

    Then let the agents build

    Hand over the style guide, the components and the screens. Now the agent copies a decision instead of guessing one.

Hand this to your coding agent. It sets the rules of the game and starts by asking you questions:

I'm starting a new product and I want to design it before any code is written. Act as my designer. We work in rounds on a design canvas (in Claude Code: Claude Design; in any other tool: one HTML file per board that I can open in my browser).

Rules for our work together:

1. Build nothing in my codebase. No framework, no components. Designs only, until I say the look is decided.
2. First ask me what the product is, who it is for and what it should feel like. Ask for references I like – websites, games, posters, brands. One question at a time.
3. Then show me six directions that really differ from each other: different layout, different visual idea, different mood. Not one layout in six color schemes. Give each a number and a name.
4. Before you show me anything, open it yourself, take a screenshot at desktop and phone width and look at it. Tell me what you checked and what you did not.
5. I will react by number and say what bothers me. From my reaction, show three, then one. If I reject a whole round, ask what bothered me instead of guessing.
6. Keep every round. Put new work below the old work, oldest at the top. Never delete a board.
7. Work in this order: brand direction, one full homepage, logo, colors with names and a contrast check (WCAG 2.2 AA), then components, then the app's main screens.
8. When the look is decided, write a short style guide and notes for the build, so a coding agent can build from it without guessing.

Start with step 2.

One warning: it is addictive. A round is so fast that you can go on forever. Set yourself an end – for me it was the moment the colors were locked and the page was redrawn with them.

Want to see the canvas and ask me about it? Come to Vibe Coding Cologne on 7 October. And if you’d rather design and build your own product with a crew around you, that’s what the Pirate Forge is for.

A full-time designer used to be out of reach for a small team. Now a design round takes minutes, and the only thing you have to bring is your taste. Go draw yours.

Cheers,
Ben

Ready to Go Deeper?

Don't miss the next Insight

Join the crew and get fresh perspectives delivered weekly.

Free account. No spam. Unsubscribe anytime.

Questions & Answers

What is vibe design?
The same loop as vibe coding, one step earlier. You describe what you want in plain words, an agent draws it on a canvas, you look and react, and it draws again. A round takes minutes. The difference to vibe coding: nothing is built. You decide the brand, the logo, the colors and every screen first, while changing your mind is still free.
Do I need to be a designer, or know Figma?
No. I'm not a designer and I can't brief a logo. What you need is an opinion: you look at six options and know which one you like and what bothers you about the rest. Pick by number, say what you feel, and bring references – a website, a game, a poster. The agent does the drawing.
Which tools did you use?
Claude Design inside Claude Code: a canvas where each design is a board. Each board is a small web page the agent writes, so it can show anything a browser can, including 3D. For the 3D tree it used Three.js, and for six reference pictures of a low-poly bonsai an image model on fal. The loop itself doesn't depend on that tool: any coding agent that can write an HTML file you open in the browser, such as Codex, can work in the same rounds.
Why not just let the agent build the page and fix it afterwards?
Because every change of mind then costs code. During the weekend the app changed from a classic dashboard to chat as the main screen, the theme went from light-and-dark to dark only, the currency from euro to US dollars, and about 20 near-identical greens became 36 named colors. On the canvas each of those was one new board. In a built product each one touches many files.
How long did it take, and how much of that was you?
About 40 hours on the calendar, from Saturday 16:50 to Monday morning. My active time was about nine hours, in about 77 messages. The result: 163 boards on eight canvas pages, in 79 versions of the canvas.
Why do you keep every old version on the canvas?
Because it makes changing your mind cheap. New rounds go below the old ones, oldest at the top. I dropped a garden idea for a kitchen idea, saw the garden again in a later round and took it back. If the old boards had been deleted, that idea would have been gone.
Is the Zawq design live?
The homepage is: since Monday, 5 October 2026, zawq.ai shows the new design. No product code was written during the design weekend itself. The app behind the homepage is designed but not built yet.
Ben Sufiani, The Captain

Ben Sufiani

The Captain

Founder from Cologne with 15 years of startup experience across 9 ventures. After helping thousands master growth marketing, Ben learned vibe coding from scratch and launched CaptAIn within three months. He leads the Vibe Coding Cologne community, blending real founder experience with teaching clarity.