The Icon Factory: 48 On-Brand Icons in Under an Hour with Claude Code + Higgsfield

I needed 48 icons for the new Talas website, one for every solution and service, visible on every page. Not stock icons. On-brand ones, matched to our design system right down to the palette and the lighting.

There are usually two ways to get that. You brief a designer and wait a week of back and forth. Or you spend an afternoon digging through icon libraries, hoping to find a set that almost matches, then fudging the rest. Neither one works when you're an AI-native agency rebuilding your own site in a sprint. And honestly, neither one is something I'd want to point to as an example of how we work.

So I built an icon factory instead. From a blank prompt to 48 finished icons saved locally and ready to drop into the site took 28 minutes and about $20. Here's exactly how I did it, because the method matters a lot more than the icons.

Why "on-brand" was the whole problem

The reason you can't just grab an icon pack is consistency. Forty-eight icons pulled from four different sources read as noise. Different stroke weights, different corner radii, different energy. A visitor won't consciously notice, but the site ends up feeling stitched together instead of designed.

For Talas, "on-brand" means a specific, repeatable spec: a palette of purple, navy blue, and coral highlights, a 3D claymorphic art style, and natural lighting with textured execution for an organic, tactile feel. That's the standard every one of the 48 had to hit, the same way, 48 times in a row. Humans are bad at that kind of repetition. It's exactly the sort of thing you want a tight prompt and a machine to enforce.

The stack: Claude for the thinking, Claude Code for the doing

The build split cleanly into two phases, and that split is really the whole trick.

First, the thinking. I worked with Claude to turn our brand guidelines into a single, tight prompt. Not a vague "make it look like us," but a spec a generator could execute the same way every time: the purple, navy, and coral palette, the 3D claymorphic style, the natural lighting and textured execution. Getting this prompt right is where the quality comes from. Everything after it is just repetition, and the prompt is what keeps that repetition consistent.

Then, the doing. I moved over to Claude Code, where I have the Higgsfield MCP connector wired in. MCP (Model Context Protocol) is what lets Claude Code actually reach out and drive Higgsfield to generate images, instead of just talking about them. I handed it one file: a markdown handover with the finalized prompt and the assignment. That file is the factory. It's reusable, it's version-able, and the next time I need a set of assets I just edit it instead of starting from scratch.

The workflow: four batches, three options each, one A/B/C choice

Here's how the 28 minutes actually went.

Claude Code generated a first round of samples against the prompt. We tuned the output on those samples specifically, dialing in background removal, edge softening, and padding, until a sample was clean enough to become the standard. Once that standard locked, the rest was just production.

The 48 icons were grouped into four batches by area of the site. For each batch, Claude Code generated three options per icon and showed them to me in a simple A/B/C format. I picked. A, B, C, straight down the list. That was my entire job: judgment, not labor. Once my choices were in, the workflow automatically stripped the backgrounds, softened the edges, and saved the finished icons straight into a local folder, named and ready for the site.

Four approval gates, one decision each. That's the part worth copying. The human stays in the loop exactly where taste matters, which is choosing, and nowhere it doesn't.

The receipts

• Time: 28 minutes, blank prompt to 48 finished icons saved locally.

• Cost: about $20, start to finish.

• Output: 48 on-brand icons, consistent across the whole site, every one chosen by me from three options.

• Reusable asset: one markdown handover file I can run again for the next set.

What I'd build differently next time

The factory worked, but two upgrades are already obvious to me.

First, I'd turn the prompting phase into a skill. Right now the brand spec lives in a markdown handover I put together by hand. The next step is a reusable skill with the Talas brand guidelines built right in, the palette, the art style, the lighting, all of it, so I can skip straight to the assignment instead of rebuilding the spec each time. The handover file was the prototype. The skill is the product.

Second, I'd add a reviewer agent. The one thing a human still had to hold in their head across all four batches was consistency. Does icon 31 actually belong to the same family as icon 6? That's a judgment a well-briefed agent can make at scale: something that reviews every generated icon against the brand guidelines and flags anything that drifts, so the whole set stays unmistakably one family while each icon stays distinct. Generation is basically solved. Automated quality control against the brand is the next frontier.

Both upgrades point the same direction, which I think is the interesting part. They push more of the brand judgment into the system itself, so the human is left with only the part that can't be automated: taste.

Why this matters beyond 48 icons

This wasn't really about icons. It was about taking a creative task that normally costs a week and a designer and turning it into a 28-minute workflow that anyone on the team can run, without dropping the bar on quality.

That's what "AI-native" actually means at Talas. Not using AI to bang out the occasional caption, but building small, reusable factories for the work that used to be slow: assets, outreach, reporting, research. The icon factory is just one of many, and the method behind it (define the standard tightly, let the machine handle the repetition, keep the human on the decisions that need taste) is the same one we bring to client work.

If that's the kind of marketing partner you want, one that builds systems instead of billing hours,