Building this portfolio with AI agents

Pierre Byrne Designs

The pierrebyrne.com work page on desktop

This site is built in Framer, and a lot of the building is done with AI agents. I use Claude and Codex to write code components, restructure pages, audit performance and check every breakpoint. I set the direction, review every change and decide what ships.

The model I work by: “Human-defined. AI-built. Human-governed.”

Rules before tools

The agents follow a short set of rules I've written down over time:

  • Nothing invented. No made-up metrics, dates, titles or client names. If a fact is missing, the agent leaves a marked gap and asks me.

  • Real evidence stays private. Work from Surescripts and eviCore gets recreated in Figma with sample data first. I approve the recreation before it goes anywhere near the site.

  • Plain language. If a non-designer wouldn't understand a sentence, it gets rewritten.

  • Hide anything unfinished. A half-built section is worse than no section.

Every change goes on a branch

Agents never edit the live site. Each pass happens on its own Framer branch, made from the main version of the site. I look at the branch, merge it and publish it myself.

That rule earned its place the hard way. During one audit, an agent found that two earlier branches of fixes had been published but never merged into the main site. When the main site was published later, those fixes quietly disappeared from production. Better page descriptions, a faster work page, smaller images and a gentler entrance on phones: all gone. The fix was to re-apply them on a fresh branch and check the result node by node against what was lost. The lesson: published isn't the same as merged.

Checking the work, not trusting it

Agents are fast and confident, which is exactly why I don't take their word for anything. Every pass ends with evidence I can check myself: screenshots at every screen size and a crawl of every link on the site. Speed gets measured before and after. The scroll-driven stages on the case studies get tested in a small local harness, because the Framer canvas only ever shows one frame of an animation.

What it's good at, and what it isn't

Agents are good at the tedious, careful work, like changing 57 animation triggers at once or finding every place an old image still lives. They're not good at taste, and they don't know what's true about my work unless I tell them. Those parts stay with me.

From the case study

0