How We Built a Brand
AI Can Run

Introduction

This is issue #004, and it's the one about brand. We're still in open beta, onboarding 20 people a day, and this week our free Chrome extension crossed 6,000 users, which is starting to send real traffic back to the site. The public launch is still a few weeks out.

Here's the thesis we keep coming back to. Building a product has never been easier, and everyone around us is shipping fast and automating everything. So when the product and the speed are table stakes, what's left to compete on? Two things: how human you are in public, which we covered in #001, and the brand, what people feel and remember after they've seen you once. This issue is how we built ours, and the part that's actually new: we built it so AI can run it. Jim led this project, so a lot of what follows is me walking through his work.

Prefer to watch or listen? Here's the full episode

From Omnia to Kai: why we rebranded before launch

The name we started with was Omnia. It means everything in Latin, and the idea was simple: an assistant that handles everything for you. It came out of a 5-day hackathon in Zurich where the only thing we optimized for was speed and shipping. We didn't think about what the brand should make people feel, we just needed a name and a site, fast. We even slapped on the same little circle icon every AI company uses. We knew from the start it was an internal, get-going name.

Before, and after.

So a month or two in, we made the call to build a real brand. Not a rebrand exactly, more like deciding for the first time what brand we actually wanted. And here's where we did something a little against the usual startup advice. Plenty of people will tell you not to pay for branding early, not to spend on a logo, just ship. We think that misreads what a brand is. It's way beyond the logo and the name. It's the system, and the feeling people walk away with. We decided that was worth real time before launch.

To lead it, Jim brought in Dave, a branding specialist we found on LinkedIn after watching a video of him rebranding a toilet-paper company. Two things made him the right call. He's genuinely creative, and he doesn't come from software or tech, so he isn't boxed in by what every other SaaS brand looks like. We wanted a shot at something that didn't feel like everyone else, and hiring outside the tech bubble was part of that bet.

The brief: brand instructions for agents, not a 50-page PDF

This is the part Jim got right early. After a couple of calls with Dave, he sent him a message that set the whole tone: the way we work is AI-native, not just the product but everything internal, so what comes out of this project can't be a 50-page PDF of colors and illustrations that nobody opens. It has to be instructions agents and LLMs can read and build on.

Dave's reaction was the tell. He'd been trying to push his own clients toward exactly this, and we were the first to come to him already asking for it. So the brief became: a memorable brand that makes people feel something, documented so agents can do more of the production themselves instead of it all running through us.

The creative process itself was normal-good. Dave started with how we wanted people to feel, and the word we kept landing on was familiar, like Kai is someone you've met before, an old friend you haven't seen in a while. He worked from a few reference brands we liked, including PostHog, which has its own very un-tech identity and a mascot. He pitched 3 or 4 names before Kai stuck, then did the same with the visual direction.

The mascot has a good origin story. Dave first showed two separate brand directions. We liked the world of the first one, the hand-drawn illustrations, and the logo of the second, a darker shape that looked a bit like someone waving their arms. So we merged them, and that's how the little green guy you see now came about.

Turning a brand into something agents can run

Once the identity was locked, the real project started: making it agent-friendly. This is the bit I haven't seen another company document, and it's the reason the issue exists. The whole thing fits in one picture: the brand on the left as a folder of files, and the pipeline on the right that turns those files into assets.

The brand, as a pipeline.

The brand folder

We asked Dave to do something most brand people never do: write down, in plain text, how he thinks when he makes a creative. Not just the finished poster, but why he had the idea and how it came to life. That turned into a folder of about 13 markdown files, one per topic. Color palette, typography, the logo system, when to use which logo, tone of voice, and so on.

Each file is the rules plus the reasoning, and crucially the exact CSS token each decision maps to, with the app equivalent next to it. The first color rule, for example, literally says to use cream and never plain white, then explains why. It's the same content you'd get out of a three-month branding engagement and a fat PDF. It's just written so that when an agent reads it, it follows the rule, and when we ask it to design a landing page, it pulls straight from the file.

What the system actually looks like.

The asset pipeline

Dave also hand-made the assets, somewhere around 16 to 20 icons plus a few posters, all drawn in Adobe Illustrator. A calendar, a clock, each one hand-drawn in cream with a thick border and a big blue panel behind it. Beautiful, but he's a freelancer, not a full-time hire, and we wanted to keep his expertise without keeping him on the team forever.

So we used his hand-made set as the training data and built our own pipeline on top. The approach is a LoRa, a low-rank adaptation model: you train it on a folder of reference images, and then it generates more in the same style from a prompt. We built ours with fal.ai, then wired it to a skill in Claude. Now the whole thing runs from plain English. Here's the loop:

1. Ask, in normal words, for something like a hand-drawn illustration of a teacup.
2. The skill finds the pipeline, hits the fal.ai LoRa, and comes back in about 2 to 3 minutes with 4 variants.
3. A composite step merges the illustration onto the brand's blue panel, and that's the finished asset.

The variants aren't all perfect, one of the teacups had a weird bit of coffee interaction we'd throw out, but usually one of the 4 is clean and on-brand. From there Claude can even write the copy to go with it, in Kai's voice, because the tone-of-voice rules are sitting in the same folder.

One prompt, four variants.

A good example of the output in the wild: last week we invited the first batch of users in with an email built entirely this way. A door swinging open, Kai standing in the middle, and the words "you made it, come on in." The door and the mascot are AI-generated and composited, and the typography is on-brand: when text is written that particular way, it means Kai is the one speaking. That went to 1,000 people on the waitlist.

Kai opening the door.

Making Kai move

The last layer is animation. For a while the mascot was static. We tried to make the SVG move with Opus 4.8 and couldn't. Then Fable 5 came out and just did it. We built a skill on top, so now we run /kai-animate, describe the motion in a sentence, and get a Lottie back. As a test we asked for a breakdance with a split and a jump, and it did it. The plan is to have Kai actually moving around the site, reacting to the blocks, instead of sitting still.

The playbook, if you want to copy it

Strip it back and the method is simple, and it's the same one we use for everything now:

1. Hire an expert for the job. Design, paid ads, outbound, branding, whatever it is, get a real professional first. AI amplifies expertise, it doesn't replace the need for it.
2. Document the manual process. Sit with them and write down how they think and why they make each call, not just the output.
3. Build the AI system on top of the documentation. The hand-made assets are your training data, the written reasoning is your rules, and together they let the work scale without the person.

The honest caveats

In the spirit of the series, the parts that are messier than the headline.

• It hasn't paid off yet. We've put weeks into this brand and it has returned nothing measurable so far. We believe it pays back over years, not weeks, but right now it's faith and a thesis, not a number on a dashboard.
• One-shot generations are good, not finished. The fal.ai variants and the first animations are impressive out of the gate, but the best ones still take iteration on the prompt to go from good to great. The pipeline removes the grunt work, not the taste.
• The website doesn't match the brand yet. We have the design system and the animations, but the current site, the meetings page especially, still reads like a decent-but-generic SaaS page. Making it match what the brand promises is the actual work, and it's not done.
• "First to do this" is a guess. We haven't seen another company document a brand as agent-runnable rules, and the timing lines up with what AI can suddenly do. But we can't prove we're first, so treat that as a hunch, not a claim.

What's next

• Redesign the whole site around the system. This week and next, the job is taking the design system and the Kai animations and making hirekai.ai look the way the brand promises, starting with the pages that feel most generic today.
• Keep onboarding, in waves. Still 20 a day in open beta, with internal debate about when the bigger launch lands, somewhere across the summer or September.
• The Lisbon hackathon. In about 3 weeks the team gets together for a big sprint. The marketing focus isn't locked yet, but the standing mission is unchanged: get more people to Kai.

If you want to follow along, the whole series lives at How We Grow. Sign up and each issue lands in your inbox, along with early access to what we're building.

FAQ