The AI Design Stack We Actually Use

Every few weeks someone on the team sends around a new “best AI tools for designers” roundup, and every few weeks it says roughly the same thing: fifteen logos, a star rating, a line about pricing. None of it tells you anything about how the tools actually sit together in a working day. So instead of adding another list to the pile, here’s what’s genuinely open on our screens right now, and — more usefully — why some tools we tried didn’t survive past the second week.

The part before any pixels get placed

Most of these roundups start with the tool that makes the screen. We start earlier than that, because the biggest time sink was never drawing — it was the hour of scattered notes, half-formed briefs, and reference images living in six different places before a project even starts.

We moved that whole stage into a shared AI canvas that can read a messy pile of notes and pull out a rough brief and mood direction. It’s not glamorous work and it doesn’t produce anything client-facing, but it’s the difference between opening Figma with a clear point of view and opening it to stare at a blank frame. If you only take one thing from our setup, it’s this: put the AI tool where the ambiguity is worst, not where the output looks most impressive.

Where the actual design happens

Figma remains the surface everything else orbits around, and its built-in AI has gotten good enough that we reach for external UI generators far less than we did a year ago. It knows our components, it respects existing auto-layout, and it doesn’t invent a new visual language every time you ask it for a variation. That consistency is worth more than raw generation power — a tool that produces something spectacular but unrelated to your system creates more cleanup work than it saves.

For quick multi-screen exploration early in a project — before a system exists to constrain anything — we’ll sometimes use a separate prompt-to-UI generator to throw five directions on the wall fast. It’s disposable work by design. Nothing from that stage goes to a client untouched; it’s there to argue with, not to ship.

The handoff, which used to be the slow part

This is where the past year has actually changed how we staff projects. We used to budget real time for a designer and developer to sit together and translate a Figma file into working front-end code, line by line, with a fair amount of back-and-forth about spacing and states that got lost in translation.

Now a coding agent takes the design file and our existing codebase and produces a working first pass on its own — component names matched, states scaffolded, spacing pulled from our tokens instead of eyeballed. It’s not the final code. It’s a draft that a developer refines instead of a page they build from nothing. That single shift has done more for our project timelines than any generation tool we’ve adopted, because it collapses a step that used to require two people’s calendars to align.

The small tools nobody talks about

Buried under all the flashy generators are a handful of narrow tools that quietly do one job well: an AI color tool that builds palettes with contrast ratios already checked, a type-pairing tool that saves the fifteen minutes of scrolling through Google Fonts trying combinations by eye. Neither one makes headlines. Both get used almost daily, because they solve a small, specific annoyance instead of trying to replace a whole stage of the process.

That’s turned into something close to a rule for us now: the tools that stick around are the ones aimed at a single real friction point. The ones that get quietly dropped are the ones promising to replace an entire discipline in one prompt box. They demo beautifully and then don’t survive contact with an actual brief.

What didn’t make the cut

Worth saying plainly: we tried more tools than we kept. A couple of all-in-one platforms that promised research, wireframing, and copy in a single workspace ended up producing work at a shallower level than doing each piece separately with a tool built for it. Breadth wasn’t the win it was pitched as. We’d rather have four tools that are each excellent at one thing than one tool that’s fine at four things.

Where this leaves us

None of this is a permanent stack. Six months ago half of it didn’t exist, and six months from now some of it will probably be replaced by something better. What’s stayed consistent is the filter we run every new tool through: does it remove a real bottleneck we’ve actually felt, or does it just look impressive in a demo. Everything above passed that test. Most things don’t.