Artificial intelligence is making it easier than ever to generate interfaces, prototypes, and entire digital products. But as execution becomes faster and cheaper, a more fundamental question emerges: what does it mean to be a designer when machines can increasingly perform the work that once defined the profession?
For John Ma, the answer lies not in producing more screens, but in designing the systems that produce them.
Credits: John Ma Substack profile image
A product designer who recently joined Factory AI after working at Meta, Ma has built his career at the intersection of product design, artificial intelligence, and emerging technologies. His experience includes roles at Scale AI and Apollo.io, working across generative AI, enterprise infrastructure, and intelligent product experiences.
At Meta, he developed AI-powered design workflows, including a shared system of agents and reusable skills intended to help designers automate repetitive checks, synthesise research, and maintain consistency across complex products.
Beyond his professional work, Ma publishes AI Artifact, a newsletter dedicated to practical AI workflows for designers and builders. His approach combines hands-on experimentation with an interest in how designers can move beyond individual tools towards more integrated systems of human and machine intelligence.
What makes his perspective particularly interesting is his willingness to discuss what does not work. Behind the promise of autonomous agents are difficult questions about adoption, maintenance, context, and the limits of automation.
In this conversation for Building Creative Machines, we explore how AI is changing the daily practice of design, why human judgment becomes more important as production becomes easier, what building in public reveals about creative work, and why the next generation of designers may spend less time operating tools and more time orchestrating intelligent systems.
You’ve been turning more of your design practice into repeatable AI workflows and agents. What has become easier to systematise, and what stubbornly remains human?
John Ma: The checks systematised first.
Is every component in this prototype actually from the design system? Does the copy match the pattern we already use for this state? What did the five competitors do here? What did the last six research sessions say?
Those used to be half a day of tab-switching. Now they’re a skill I run.
At Meta, I turned twelve of those skills and five design agents into a shared “design brain” for a team of about ten designers across eight product areas.
Plain Markdown and YAML in a synced folder, a rules.md read on every run, and a publish gate that shows you a diff and makes you type a literal “yes.” No plugin, no install.
The build took weeks. Adoption took longer.
On launch day, one designer out of ten had made it past step three of onboarding, and that one was me.
What moved it was two teammates using it on real work in front of everyone else. Not a better demo from me.
That’s a human problem, and I don’t think it ever stops being one.
The other part that stays human is knowing the domain.
I designed the tools Meta’s infrastructure organisation uses to plan capacity and place hardware in data centres. I could ship a beautiful prototype in an afternoon and still lose the room, because the technical programme managers were thinking in operational concepts I hadn’t learned yet.
No agent fixes that.
You sit in the meetings, you ask the dumb questions, you learn the territory.
The AI only accelerates you once you know where you’re going.
You experiment with many different AI tools and workflows. What makes a tool genuinely earn a permanent place in your creative practice? Can you share an experiment that didn’t make the cut?
John Ma: A tool earns its place if I come back to it every day after the first use.
The real sign is when all my knowledge is deeply encoded in it, and it feels unintuitive to switch, because the current setup generates output quality that’s so good I’d be starting from zero anywhere else.
Most tools pass the demo and fail the second Tuesday.
The ones that stay are the ones where leaving would cost me context, not just a login.
The one that didn’t make the cut: I stopped running my Telegram bots.
I had a fleet of them on a VPS, one for writing, one for design, one for repurposing content.
The noise-to-quality ratio was too high. The output became noise at this point, not customised to anything I was actually doing.
And maintaining and disconnecting the VPS was annoying in a way that compounds. Every week there’s a thing to fix that has nothing to do with design.
It did leave me with a real question I haven’t settled: cloud versus local.
You want things to keep running when the laptop is closed, which argues for cloud. But everything that made the bots bad came from them running far away from my actual files and context.
I don’t have the answer yet.
As AI makes execution increasingly cheap, where do you think the real craft of design is moving? Does human taste become more important when almost anyone can generate a polished interface?
John Ma: I prototype everything in code now.
Design changes happen in code through the Figma MCP, and the code is the prototype. Figma became the layer where I do visual polish, not the place where design happens.
That flips what a designer’s day looks like, and it moves the craft to two places.
First, the system behind the output.
A design factory needs a skill that checks every component comes from the design system, and it needs an agent that understands the laws of UX and keeps the thing simple enough to use.
Writing those rules well is design work.
If your taste can’t be written down as something an agent can follow, you don’t have taste yet, you have preferences.
Second, interaction detail and systems thinking.
That was the recurring feedback I got across every interview loop this year, and it’s where the bar is moving.
Anyone can generate forty screens.
The hard part is the one screen where chat-driven setup and visible, editable artifacts have to work together, not all form and not all chat.
I ran into that designing onboarding for an agent evaluation product, where you connect a repository, run a baseline, and land an engineer in a conversation with their own results.
The generation was the easy part.
Deciding what the engineer should see, in what order, and what they could change, took the time.
So yes, taste matters more. But taste as a verb, not a noun.
The designers who win will be the ones who can turn judgment into rules, evaluations, and review loops that other people and other agents can run.
You’ve made experimentation in public an important part of your work through AI Artifact. How has publishing your processes, successes, and failures changed the way you learn and design? Has your audience challenged any of your assumptions?
John Ma: I started the newsletter after a personal audit showed a pattern I didn’t like: I over-prepared and under-shipped.
Publishing weekly was the fix.
The Day N/100 series on X is the same medicine in a smaller dose. One insight a day, posted whether or not it’s polished.
What surprised me wasn’t one reader, it was the stats.
You get real-time feedback on what people like and dislike, and it didn’t match what I expected.
Some guides I pushed hard, like the loop engineering piece, were too complex and not well received.
Meanwhile, the original prompts I pulled out of the Muse articles, and the personal agent pieces, were much more popular.
The general audience, I realised, is not as deeply technical as I assumed.
Even my Hermes personal agent setup, which is super long, is not really useful to a normal user.
They don’t want the architecture. They want the eleven prompts.
That changed how I design, not just how I write.
If a workflow only works for someone who already thinks like me, it isn’t done.
It’s the same lesson as the team design brain: I call it a third brain, a second brain that has to work for someone who wasn’t in the room when it was built.
Readers are the harshest version of that test, because they never tell you why they left.
Looking two or three years ahead, do you imagine designers still primarily using individual AI tools, or increasingly directing systems of specialised agents? How might that change the way creative teams operate?
John Ma: They all run the systems eventually.
That’s what I saw when I joined Factory AI.
The job isn’t one workflow.
You’re building a loop of workflows that ingests signals from everywhere: user research sessions, call transcripts, Slack feedback, usage data, and then runs the different loops that turn those signals into product.
A single “generate me a screen” prompt is the smallest piece of that, not the job.
For designers who want to prepare for that future, what practical skills should they start developing today?
John Ma: Four things, in this order.
Write rules, not prompts.
A rules.md that an agent reads every run is a design document.
Learn to write one that a stranger, human or model, follows correctly.
Most designers I know can write a good prompt. Very few can write a rule that holds up on the hundredth run.
Read code well enough to review a diff.
You don’t have to write the whole thing.
You have to know when the agent’s output is wrong, and you have to be able to ship it when it’s right.
I commit frontend diffs directly now, and that single skill changed how engineers treat my work.
Learn evaluations.
Criteria grade outputs, grades cluster into issues, each issue maps to a fix.
That loop is how you’ll manage agents, and it’s how you’ll prove your design judgment at scale instead of asserting it in a critique.
Keep the human parts sharp.
Domain knowledge, running a room, getting ten people to adopt something.
Those compound while everything else gets cheaper.
Beyond the Interface: Designing the Intelligence Behind the Work
John Ma’s experience points towards a significant transformation in the relationship between creativity, technology, and professional expertise.
For decades, digital design has been organised around tools that allow humans to create and manipulate representations: interfaces, wireframes, prototypes, and visual systems. Artificial intelligence is beginning to change that relationship. Increasingly, the designer’s contribution may lie not only in producing an artifact, but in defining the intelligence, constraints, and feedback mechanisms through which artifacts are produced.
This is more than an evolution in productivity. It suggests a different understanding of creative authorship.
When a designer translates experience into reusable rules, establishes criteria for evaluating outputs, and coordinates several specialised agents, design becomes partly an exercise in constructing systems of judgment. The creative professional is no longer responsible for every individual decision, but remains accountable for the principles, boundaries, and consequences of the process.
Yet Ma’s experience also exposes a paradox.
The technical mechanisms for automating design are advancing rapidly, but the social mechanisms for making those systems useful remain much harder to reproduce. His account of building a shared design brain at Meta illustrates the distinction. The system itself could be assembled in weeks. Convincing colleagues to integrate it into their work required something less easily automated: trust, demonstration, shared understanding, and the willingness to change established habits.
There is a similar tension in his reflections on building in public. An intelligent system can be sophisticated, technically impressive, and almost entirely irrelevant to the people it is intended to help. Complexity is not the same as usefulness. A functioning workflow is not necessarily a usable workflow.
Perhaps the most provocative idea in this conversation is Ma’s distinction between taste and preference.
Traditionally, creative judgment has often been understood as something personal, intuitive, and difficult to formalise. Ma challenges that assumption by arguing that designers should increasingly be able to express their judgment through rules, evaluations, and repeatable processes.
That raises an important question: can creative judgment be fully translated into instructions?
Probably not. The most interesting forms of creativity often emerge precisely when established rules are questioned, when familiar patterns are rejected, or when an unexpected connection changes the direction of a project. A system that consistently reproduces existing standards may produce reliable work without necessarily producing original work.
But the ability to articulate judgment, evaluate alternatives, and recognise when established rules no longer apply may become even more valuable as AI systems take on greater responsibility for execution.
The future of design, then, may not be a simple competition between human creativity and machine generation.
It may be a transition from designing individual outputs towards designing the conditions under which useful, coherent, and sometimes surprising outputs can emerge.
And within that transition, the most enduring human capabilities may prove to be the ones Ma returns to throughout our conversation: understanding a domain, asking better questions, navigating complexity, earning trust, and knowing when a system needs to change.
As machines become increasingly capable of generating the artifacts of creativity, the central challenge may become something deeper:
Not simply what we can create with AI, but how we design the intelligence that participates in creation — and how we preserve the human judgment needed to give that intelligence direction and meaning.
Follow us on Youtube and LinkedIn.
About John Ma
John Ma is a product designer at Factory AI and a former Senior Product Designer at Meta, where he worked on AI-enabled internal tools, data infrastructure, and intelligent design workflows. His previous experience includes Scale AI and Apollo.io, among other technology companies.
He is also the author of AI Artifact, a newsletter exploring practical AI workflows, design systems, agents, and experimentation for designers and builders.
Follow John on LinkedIn.



So cool to see this