Ask Claude to build a landing page and the finished site comes back with pictures on it: a logo, a few illustrations, a hero image. That result confuses people for a solid reason. Anthropic says it directly in its documentation: Claude is an image understanding model. It reads pictures. It doesn't make them.
So where do the pictures come from? A coding agent gets them three ways, and knowing which way a picture was made changes what you ask for next time.
Way 1: It calls an image service
When a site needs a photograph, the agent writes a few lines of code that send a description to an image generation service, and that service's image model draws the picture and hands back the file.
The division of labor is the whole trick. Claude wrote the request and handled the plumbing. The image model made the pixels.
This is also why asking Claude by itself to "generate an image" tends to go badly. There's no image model inside Claude to do the drawing, so the request has nowhere to land. When an image tool is connected, the same request tends to work, because now there's a specialist on the other end of the line.
Tools connect to agents through what the AI world calls MCP, short for Model Context Protocol. The practical version: agents can be given tools, image generation is one of them, and connecting one usually starts in the tool or connector settings. The exact setup depends on the provider.
Way 2: It writes the picture as code
For logos, icons, and diagrams, the picture on an AI-built site is often not a picture at all. It's an SVG file: a text file that describes shapes. "A circle at these coordinates, filled with this color, with a line under it." The browser reads the description and draws it.
This route plays to a coding agent's home advantage. Writing text is one of Claude's strengths, and an SVG file is nothing but text. So the agent can design a logo the way it writes everything else, and it can revise it the same way: "make the icon rounder" is a text edit, not a redo from scratch.
SVG has two bonus features worth knowing. It scales to any size without blurring, because the browser redraws it from the description at whatever size is needed. And there's no image-service fee and no wait for a generation, because the file is written as code.
We use this route for our own graphics. Diagrams and illustrations for our posts get written as code, rendered by a browser, and shipped. The social preview card for every post on this site is drawn the same way: code describes the design, a renderer draws it, and no image model is involved.
Way 3: It points at pictures that already exist
The third route is the simplest: the agent links to or downloads an image that someone else already made. Free stock photo libraries like Unsplash and Pexels are the usual sources, along with placeholder images that stand in during a build.
The sources depend on how the agent is set up and what it can reach. If you want stock photography, point the agent at a library whose license fits your use, since it might not pick a legally suitable image on its own. Links also break: a stock photo URL can expire or move, which is why an AI-built site occasionally shows a gap where an image used to be.
Which one to ask for
The three paths map cleanly onto requests:
- A photograph of something specific (a person at a desk, a landscape, food): connect an image generation tool first, then ask for the image. Without the tool connected, ask for a stock photo instead.
- A logo, icon, or diagram: just ask, no setup needed. Naming the format helps: "as an SVG file" tells the agent to take the code route.
- A picture for right now, quality to be decided: ask for a placeholder, and swap in the finished image later.
The one request worth rewriting is "make me an image" said to a bare coding agent. It fails not because the agent is limited but because the request sends the job to the wrong desk. Route the same want to a connected image tool, or ask for it as code, and it works.
FAQ
Can ChatGPT generate images if Claude can't?
Yes. ChatGPT's app bundles an image model, so typing "draw me a cat" works inside ChatGPT without extra setup. Claude's app doesn't bundle one as of this writing. Both coding agents can reach image models through tools, so with setup they land in the same place.
What is an SVG file?
A text file that describes a drawing: shapes, positions, and colors written out in code. Any text editor can open one, and browsers and most design tools can render it. Because it's text, an AI agent can write one, and you can ask it to tweak the file the way you'd tweak a sentence.
Why is my AI-built site showing broken images?
Usually because an image was a link to somewhere else, and that link died. Ask the agent to download the images into the project, or to redraw simple ones as SVG. That removes the risk of an outside URL expiring, though files inside the project can still be misnamed, so it's worth a click through the pages afterward.
How do I connect image generation to Claude?
Look for the connector settings in your coding agent (the MCP settings) and add an image generation tool from a provider that offers one. Each provider has its own setup steps and pricing, so follow that provider's documentation. Once the connection is configured, the agent can call the tool mid-project, within whatever permissions you granted.
Is it safe to use AI-generated images on a business site?
The image service's terms of service decide this, and they differ between providers, so check the commercial-use terms before putting results on a site that earns money. Stock libraries have their own licenses too; check the terms for the specific library and image before using a download commercially.
Want to go further?
- How to Use Claude AI: A Beginner's Guide covers what Claude is good at and where it needs help.
- What Is an AI Agent? Explained Simply covers tool use, the idea underneath everything above.