Autonomos was built with AI at basically every step. The design system, the copy, the website, the agent workflows themselves. Claude wrote most of the code. Claude wrote most of the product copy. Claude helped design the dashboards.
Here's what we'd tell someone trying to do the same thing.
What worked really well
Code generation. Claude Code built our entire Next.js app from design specs. Components, pages, data layer, deployment. It's genuinely good at this and it saved us weeks.
First drafts of everything. Product copy, blog posts, email sequences. AI first drafts are a massive time saver. They're rarely perfect, but they're a much better starting point than a blank page.
Design system consistency. Once we defined our tokens and rules, AI kept everything consistent across five dashboards. Colors, spacing, typography — it doesn't drift the way a human designer might over a long project.
What didn't work
The copy sounded like AI. Our first pass on all the product copy was technically correct, well-structured, and completely lifeless. It read like a marketing textbook. We had to rewrite a lot of it to sound like actual humans talking.
If you're building with AI, plan to rewrite the copy. Not revise it — rewrite it. The structure AI gives you is useful. The voice is usually wrong.
The icons we first picked were random. We grabbed Unicode symbols for our agent characters. They looked meaningless. We had to go back and design actual icons that matched what each character does — crosshair for the finder, crown for the judge, eye for the pattern-spotter.
Details like that matter more than you think. "Close enough" reads as careless.
What we learned
AI is a great employee but a bad founder. It executes well. It doesn't have opinions. When we let AI make creative decisions, the result was generic. When we made the creative decisions and let AI execute them, the result was specific and interesting.
Speed is the biggest advantage. We went from design system to deployed website in days, not months. That velocity let us iterate faster than we could have with a traditional build process.
You still need taste. AI removes the execution bottleneck. It doesn't remove the need for someone to know what good looks like. If anything, it makes taste more important, because the bottleneck shifts from "can we build it" to "should we build it this way."
The honest summary: AI got us 80% of the way to a finished product in 20% of the time. The remaining 20% — the personality, the voice, the details that make it feel like humans made it — still required humans.