how we design and build with AI: the tools behind our work

people ask me which AI tools we use.

the honest answer: fewer than you'd think, used deeply.

here's the whole kit, and how it fits together.

claude code, running fable 5

almost everything starts here.

claude code is an AI coding agent that works inside a real codebase. we run it with fable 5.

for client work, i clone the client's github repo, and the agent gets direct access to the codebase and the front end. so when i have design feedback, it becomes code. it goes to a branch, and the CTO and the tech team review it like any other pull request.

the old way was do-it-with-you: async comments, annotated screenshots, review calls, and a two-week lag.

the new way is do-it-for-you. design judgment travels all the way to the product, and engineering reviews instead of translating. faster for us, better for the client, and kinder to the engineers' time too.

figma, through MCP

MCP, the model context protocol, is a standard way to plug tools into an AI agent.

with the figma MCP connector, claude code reads the real design files: the components, the tokens, the spacing. an agent that can see the design system builds with it.

that matters more every month. on zaki AI, the design system is more than 120 components, in light and dark, right to left in arabic. the agent works inside those rules from the first line of code.

mobbin, for patterns

the mobbin MCP gives the agent access to real app screens and ux flows from some of the best products out there.

when we design an onboarding or a paywall, we start from patterns people already know. then we make them ours.

references make the first draft better. taste still decides what ships.

swiftui for iOS, react for the web

for native iOS apps, we build in swiftui. tabilog uses mapkit for its globe and storekit for purchases. haku health uses healthkit to read your sleep from apple health each morning.

for web apps, we work in react with shadcn/ui and tailwind, as on the zaki web app. zaki's phone app runs on expo.

cloudflare, quietly underneath

haku health's AI reader runs on cloudflare workers, with claude reading plates, barcodes, and labels. this website runs on cloudflare too.

it's fast, it's affordable to start, and it scales quietly. that's exactly what an early product needs.

a mac mini that never sleeps

the newest part of the kit is a little mac mini, running AI agents around the clock.

when a founder books a launch with us, an agent reads the booking, studies their site and the sites they love, and prepares a kickoff pack and a first draft. it's ready for my review before the call.

other agents handle research, drafts, follow-ups, and repetitive admin. everything waits for my review before it goes anywhere. updates reach my phone on telegram through openclaw, so i can review from anywhere.

it feels like hiring a tireless junior teammate for the price of electricity.

what stays human

AI speeds up layouts, scaffolding, routine integrations, first drafts of copy, and testing.

experienced people make the product decisions. we own the permissions, the payment failures, the edge cases, the quality, and the release checks.

in one year, the tools changed completely. the job got bigger. design, build, and sell, by the same small team.