Your "Taste" Is Now a Technical Skill — What the Codex Lead Gets Right About Designers Who Build

Andrew Ambrosino, the person responsible for turning OpenAI's Codex desktop app into what he calls "the best desktop app that has ever existed, full stop," just went deep on Lenny's Newsletter about what product work actually looks like inside OpenAI right now. The headline finding: nearly 100% of OpenAI employees — not just engineers — use Codex weekly. If that's already true inside the company that makes the tool, the rest of us are a few months behind a wave that's already broken.

The thesis nobody is saying plainly enough

Here it is: the bottleneck in building has moved from "can you write code?" to "do you have good taste?"

Ambrosino frames taste not as an aesthetic preference — liking clean type or nice colour palettes — but as a professional capability: the ability to make sharp, early decisions about what a product should feel like before a single screen is built. He points to research on what's sometimes called the "primal mark" — the idea that the earliest choices in a creative process have an outsized influence on everything that follows. In an AI-first workflow, where you can go from prompt to working prototype in an afternoon, those first decisions land faster and lock in harder than ever.

For designers, this is genuinely good news. Taste, judgment, and the ability to hold a clear vision — these are the things design training actually builds. The skill gap between you and a senior engineer just got a lot narrower. The taste gap between you and someone without a design background has arguably gotten wider.

What "roles are collapsing" actually means (and doesn't mean)

Ambrosino describes his team at OpenAI as having collapsed traditional role boundaries — designers writing code, engineers making product calls, PMs shipping features directly. But he's careful about one thing: eliminating the concept of roles entirely is a mistake.

What's changing is not that specialisms stop mattering. It's that the handoff model — designer throws a spec over the wall, engineer builds it, PM files the ticket — is dying. What replaces it is closer to what he calls a "zone defense": everyone covers their primary area, but anyone can and does cross over when the moment calls for it.

For you, as a designer starting to build, this means the goal isn't to become a full-stack engineer. It's to be fluent enough in the tools that you can close the loop yourself on small things — a layout change, a copy tweak, a logic adjustment — without waiting for a handoff that might never come.

How to actually use Codex as a non-engineer

Codex (accessible at chatgpt.com/codex) is a coding agent — meaning it doesn't just suggest code snippets, it can work on tasks somewhat autonomously in the background while you do other things. Think of it less like autocomplete and more like a junior developer you can brief and then check in on.

Ambrosino uses it to run his own internal workflows — the article doesn't detail every specific workflow, but the broader pattern is clear: he delegates implementation tasks to Codex and focuses his own attention on the decisions that require judgment. That's the exact division of labour designers should be aiming for.

A practical starting point: instead of trying to learn syntax first, try briefing Codex the way you'd brief a collaborator. Describe the outcome you want and the feeling it should have, not just the functional requirement. "Build a simple task list that feels calm and doesn't interrupt me" is a better brief than "make a to-do app." The taste you bring to that brief is your competitive advantage.

There's also a hint in the source material of something bigger coming: a vision for a "home base" that would coordinate work across ChatGPT, Codex, and the tools people already use day-to-day. An early version of this appears to exist under the name Atlas (at chatgpt.com/atlas), though details remain sparse. Watch that space.

Why timing matters more than you think

One of the more interesting admissions in the interview: Ambrosino believes the Codex desktop app would have failed if they'd launched it last November instead of in February. The models weren't capable enough for the product to actually deliver on its promise.

This is worth sitting with. A lot of the frustration designers feel when they try AI coding tools — "it doesn't understand what I mean," "it keeps breaking things" — is partly a model capability problem that is actively being solved, not a permanent ceiling. The tool you tried six months ago and bounced off may be a genuinely different experience today. The on-ramp is getting less steep in real time.

The grounded takeaway

None of this means the path from designer to builder is instant or frictionless. Learning to work with an AI coding agent — writing good briefs, reviewing its output critically, knowing when to push back — is its own skill that takes reps. And the "roles collapsing" framing can obscure real complexity: there are still things that require deep engineering knowledge, and pretending otherwise gets people into trouble.

But the direction of travel is clear. Designers who develop genuine taste and the fluency to act on it directly — without always needing an engineer in the loop — are going to be unusually valuable. The tools are catching up to that possibility faster than most people realise.

Start small. Brief Codex like a collaborator. Trust your taste. That last part, it turns out, is not soft advice — it might be the most technical thing you do.