Let's start with the mindset problem, because no skill sticks without it: designers need to stop thinking AI is replacing them and start realising it can empower them. AI cannot just create. It needs your input to create. It's a tool. A spectacular one, but a tool, and the value in any tool-based profession has never been the tool. It's the human holding it. Nobody looks at a beautiful building and congratulates the scaffolding.

With that installed, here's the practical question: what should you actually learn? Because the honest answer is not "this week's hot app". New AI design tools launch weekly and most will be gone by Christmas; chasing them is a treadmill dressed up as professional development. Underneath the churn, though, four durable skills keep compounding. They transfer across every model, every app, and every hype cycle. I use all four daily, first as a designer and now in product management, and they are, as far as I can tell, the whole game.

1. Brief it like a professional assistant, not a vending machine

The single biggest lesson from years of daily AI use: the more detailed and focused your instructions, the better the output. Every disappointing AI result I've ever produced traces back to a lazy prompt. Every impressive one traces back to a proper brief.

Here's the reframe that unlocked it for me: don't treat AI like a tool you operate, treat it like a professional human assistant you're delegating to. Ironically, this machine works best when you brief it exactly the way you'd brief a capable colleague. Context, goal, constraints, audience, format, examples of good and bad. Designers should be brilliant at this. We've spent careers writing creative briefs and complaining when clients give us vague ones. Prompting is the same discipline pointed the other way. If you've ever received "make it pop" and despaired, congratulations, you already understand prompt quality. Don't be the "make it pop" client to your own AI.

My running description of AI is that it's a five-year-old adult: fully formed and articulate on the surface, but only ever as good as the information provided and the guardrails set. Sometimes you genuinely have to hold its hand to get the output you expected. You know the joke that the smarter a person is, the less common sense they have? AI is the definitive proof. Brief accordingly.

Start today: take a prompt that disappointed you recently and rewrite it as if delegating to a new team member on their first day. Background, objective, constraints, what "good" looks like. Compare the results. That gap is the skill.

2. Narrow the scope: build yourself a team of personas

Here's the counterintuitive one. AI is a hugely capable generalist, and that's precisely the risk: ask a do-everything tool a broad question and you get a watered-down, everything-flavoured answer. The fix is deliberate scope narrowing, meaning constraining the AI's context so it responds like the specialist you actually need.

My method: I keep a library of predefined text files that act as agents. They're reusable persona definitions I load in depending on the job. I have one for a UI expert, a user researcher, a front-end developer, a back-end developer, an analyst, even a product manager. Each file defines who the AI is being, what it cares about, what lens it critiques through, and what kind of output I expect. When I want a design review, I don't ask a general assistant "any thoughts?" I brief my UI expert. When I want holes poked in a plan, the analyst does it.

Two things happen when you work this way. First, the output quality jumps, because the model isn't averaging across every possible reader. It's answering as someone, for something. Second, and nobody warns you about this, you get a strange multidisciplinary education. Writing a good back-end developer persona forces you to understand what back-end developers actually care about. Building the team teaches you the team.

Start today: write one persona file for the discipline you most often need feedback from. A few hundred words: role, expertise, priorities, pet hates, output format. Save it, reuse it, refine it. Then build the next one.

3. Stay the human in the loop, because the loop is the job

Every impressive AI output you've ever seen had a human somewhere behind it doing the unglamorous work of catching what was wrong. In my daily use, AI output almost always needs my input, from small tweaks to correcting large misunderstandings. I've had to remind it of things we'd already agreed. I've watched it be confidently, fluently incorrect. None of this is a scandal; it's the nature of the tool, and it defines the third skill: judgement at speed.

When generation is nearly free, taste becomes the bottleneck. The scarce ability is no longer "can you produce fifty options?" The machine does that before your coffee cools. It's "can you evaluate fifty options quickly, spot the two worth pursuing, catch the subtle error in the plausible one, and say why?" That's critique, which design school actually trained us for, now running at industrial throughput.

This is also why "human in the loop" deserves to be more than a compliance slogan. The loop isn't a safety measure bolted onto the process. The loop is the process. A human in the loop is always needed, no matter how good the models get, because a business doesn't just need output; it needs someone accountable for the output. The machine produces. You are responsible. That word, responsible, is your job security, so get comfortable exercising it: verify claims that matter, challenge outputs that feel off, and never ship anything you couldn't defend in a room without mentioning the AI.

Start today: next time AI gives you something 90% right, don't just fix the 10%. Articulate what kind of error it was. Misread context? Missing constraint? Invented fact? You'll start seeing the patterns, and your briefs (skill 1) improve automatically.

4. Build it into your workflow, don't just visit it

The final skill separates people who use AI from people who've adopted it: workflow design. Knowing where AI belongs in your end-to-end process, and, just as importantly, where it doesn't.

My own progression is a decent map. I started, back in the Microsoft Copilot days, with documentation. The busywork layer. Useful, but a visit, not a residence. The step change came from wiring AI into the actual creative pipeline: using Claude to research, to draft briefs and proposals, to analyse data, and then the big one, using it to prototype ideas I can put in front of real people for feedback. Lately I've been pushing further with Claude Code, extending those prototypes toward working software, with the explicit aim of shrinking the design-to-dev gap. That handoff has been the most reliable source of pain in every team I've ever worked in. (The full audit of what this did to my week is in AI didn't take my job, it took my busywork.)

Notice the direction of travel: from AI as a writing assistant, to AI as a research assistant, to AI as a making tool that collapses stages of the process into each other. Each step needed skills 1 to 3 already in place. The briefing, the scoped personas, the judgement. Workflow design is where they compound.

And notice what's not in my workflow: I don't hand AI the parts where the value is human contact. Research with actual users, stakeholder conversations, the judgement calls. Knowing where the tool stops is as much a workflow skill as knowing where it starts.

Start today: map your process end to end and mark each stage: automate, assist, or human-only. Then wire AI into one "assist" stage properly. Not as an occasional visit, but as the default way that stage now works.

Tools are verbs. These are grammar.

Notice that none of the four skills is "learn tool X". Briefing, scoping, judgement, workflow design: they're the grammar underneath every tool. Learn the grammar and every new verb is a weekend, not a crisis. Skip the grammar and every product launch feels like an extinction event. (And if you want the historical evidence that it never is one, I've lived through a few: this isn't my first extinction event.)

The tools will keep churning. The four skills will keep compounding. Spend your five hours a week accordingly.