AI-generated SVG illustrations: why I use Quiver instead of Claude and ChatGPT for website logos, icons and illustrations

If the AI illustrations on your website look fake, the problem is the format. Logos, icons and illustrations need an SVG, a vector image. And when it comes to SVGs, Claude and ChatGPT are genuinely bad. For this job I use Quiver, which generates SVGs with a specialized model and connects to Claude Code.

Everyone talks about AI slop, but almost always about text and interfaces. Hardly anyone talks about vector graphics, which is exactly what gives a website its face.

Why use an SVG instead of a ChatGPT image?

An image generated by ChatGPT is a photo: you use it as it is. If you want a different color or different text, you have to regenerate it from scratch and hope everything else stays the same.

An SVG is code. Claude can edit it whenever you want, without regenerating anything:

• it changes the color or adapts it to your brand palette;
• it removes or adds text;
• it stays sharp at any size, from the smallest icon to a full-screen banner;
• it weighs almost nothing, so your site loads faster.

My rule is simple. For realistic photos, ChatGPT is fine. For logos, icons and illustrations, SVG wins every time.

Why do Claude and ChatGPT draw SVGs so badly?

ChatGPT and the other image generators think in pixels, tiny colored dots. The result is a photo, even when it looks like a drawing.

Claude can write SVGs, because an SVG is text. But Claude is a general-purpose model: it writes the coordinates of curves and shapes without seeing the drawing they produce. In practice it draws blind, and it shows.

I ran a simple test. I asked Quiver to draw the front squat technique and gave Claude the exact same prompt. In the video the difference is obvious at a glance: Quiver’s illustration is clear and polished, Claude’s isn’t.

What is Quiver and how does Arrow work?

Quiver is the product of QuiverAI, a lab that does nothing but vector graphics: it generates, edits and animates SVGs for product, brand and marketing work. Inside it runs Arrow, their model.

Arrow writes the SVG code directly, skipping pixels entirely, and picks curves, points and colors the way a designer would. It’s trained only for this, which is why the output is clean.

And there are two more things:

• a Max version of the model, built for technical drawings full of detail, like the front squat from the test. Today the family is called Arrow 2 and Arrow 2 Telos, where Telos is the variant for complex briefs;
• reference images: you give it up to 4 and Quiver copies their style, keeping the palette and the line work.

And no, they don’t pay me. They don’t even know who I am. I found it, I liked it and I’m sharing it.

How do you connect Quiver to Claude Code?

The part I care about most is the MCP server. MCP is the protocol that lets an agent like Claude Code use external tools. Quiver has an official one, and in the QuiverAI documentation the command for Claude Code is this:

claude mcp add --transport http quiverai https://app.quiver.ai/mcp

Then open Claude Code, run /mcp and log in.

From then on you tell Claude Code things like "make the logo for my app" or "create the icons for this page". Claude Code generates them with Quiver and puts them straight into the project. With the same connection you can also turn a regular image into a vector or animate an SVG you already have.

For anyone working with AI agents, this changes the workflow. Graphics join the project as an ordinary task: the agent generates them and saves them where they belong, with no detour through another tool and no manual import.

How much does Quiver cost?

When I tried it, in September 2026, there was a free plan with about 10 images a week, enough to test it properly. Today Quiver’s pricing page starts with the Go plan at $8 a month, with a 14-day trial.

If your site has icons and illustrations that all look the same, try redoing one as an SVG with Quiver and compare it with the one you have now: you’ll see the difference for yourself.