How a designer who can’t code built a tracing tool with Claude

AI can build tools now. So how does a designer talk to one and get something that works every time? Here’s what I figured out.

Where it started: a client sent me 30 logo PNGs and asked for SVGs

I think every designer has been here at some point.

Before AI, my routine was: drop the image into Illustrator, run Image Trace, clean up the anchor points by hand, then bring the whole thing into Figma. This time I wanted to try something different — get Claude to build me a tracing tool, and see whether it could produce SVGs that were clean enough to actually use.

The first results were not good. The logos the client sent were low resolution, so every edge came out wobbly, with hundreds of anchor points packed so tightly there was no way to fix them by hand.

Honestly, my first thought was to give up. Illustrator was still doing a better job.

I couldn’t quite let it go, though. I kept zooming into that wobbly edge, trying to work out why it looked so wrong.

Then it hit me. The AI was never looking at the shape. It just walks around the edge of the pixels and drops a point every few steps — faithfully, including all the soft grey mush along the way.

It wasn’t drawing my logo. It was copying my pixels.

So I stopped describing how I wanted the result to look, and started telling it how to decide what it was looking at. And the SVGs started coming out clean. Along the way I picked up a few things about how to talk to it, so I thought I’d share them here.

1. To an AI, a sentence like “make it smoother” means nothing. Tell it how to decide instead.

I spent ages saying stuff like “fewer points” or “make the lines smoother.” That kind of thing is too vague — the AI can only guess, which is why it worked one time and not the next.

So I tried a different way. I asked it to work out what the outline was actually made of — which is really just how we designers work anyway. When I trace a logo by hand, I sort it in my head first: straight line, arc, actual curve. Then I decide where the points go.

So I asked the AI to sort first, then draw: an outline is only ever made of three things — straight lines, arcs and Bézier curves — so use those three, and get as close to the original as you can. Once it had that rule, the number of anchor points dropped straight away.

The text inside a logo turned out to be a different problem, though.

Lettering is almost never drawn from scratch; most of the time it’s an existing typeface that somebody modified. So for that part I took a different approach: I had the AI find the closest typeface, use it as a base, and then nudge each anchor point toward the edge of the original image. The typeface gives you the skeleton, and the original gives you the position and the weight.

Both of these are things I already do when I trace by hand — sort the shapes, recognize the type. All I really did was spell them out for the AI.

2. Circling what’s wrong is the fastest way to explain it

Once it has the rules, I let the AI do a first pass, sorting every part of the outline into a line, an arc or a curve. That first pass still gets a fair amount wrong.

Here’s a common one. On a low-resolution image, a diagonal edge in pixels genuinely is a staircase — so the AI draws what it sees, and traces every single step.

This is the point where I circle things. I go straight onto the output, mark the spot, and say what it should have been: this stair-stepped bit is actually one diagonal line. It then goes back to what I circled, works out the cause, changes the code and traces again. Once it’s right, that rule stays in the code, which means I don’t have to say it again on the next logo — and the tool gets a little more accurate each time.

3. Give the AI a test, and write the answer key yourself

I wanted a tool I could rely on, not one I’d have to fix all over again every time a new logo came in.

So I made my own test images, and I already knew the answers to all of them.

I draw a clean SVG first, render it out as a PNG, and then hand that PNG to the AI to trace — it only ever sees the picture, and has no idea what the original looked like. When it’s finished, I compare what came back against the SVG I started with, and anything that’s off shows up immediately.

Because I wrote the answers myself, I always know exactly what went wrong and by how much.

There’s another thing I like about this: I get to decide how bad the image is. If I want to test blur, I add blur. Tiny text, I render it small. Jagged edges, I shrink it and blow it back up. No hunting around for sample files that happen to have the problem I’m trying to test.

4. Turning it into a tool I can just use

By now the results were consistent, so I started wondering what to turn all of this into.

I spend most of my day in Figma, so I asked Claude to turn it into a Figma plugin — no more moving images back and forth, and no need to start a conversation every time.

The plugin is fast and, more importantly, consistent: a few seconds and I have my SVG. For some complicated logos the plugin doesn’t trace them well, so I take those back to Claude, where I can work on the specific parts that came out wrong. Once I’m happy with it, I ask Claude to write the new rule back into the skill, and then update the plugin.

So this is how the two of them split the work now: the plugin handles the everyday, and Claude handles the exceptions. And once an exception has been solved and the plugin is updated, it stops being an exception.

Last thoughts

I used to move the same file through two or three apps — export here, fix it there, import it somewhere else. The tools weren’t mine. I could only work inside whatever someone else had decided to give me.

With AI, that order flips. I can start from what I need, and then build the tool that does it. I find that genuinely interesting, and it’s made me braver — I used to look at an idea and ask whether it could be done at all, and now I ask whether it’s worth doing.

AI is a very capable partner, but its weakness is that it has no idea what you care about. Which means the useful skill is being able to explain the problem clearly: taking “that looks off” and turning it into something specific, like spacing, alignment, weight or corners.

That part, designers are already good at.

Thanks for reading. If you’re building tools with AI too, I’d love to hear how you’re going about it — leave a comment, I read all of them. And if this was useful, a few claps would mean a lot; every bit of response turns into fuel for the next one 💪

How a designer who can’t code built a tracing tool with Claude was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.