Build with AI · 2 min read

Building with AI, from first prompt to tested design.

AI makes the pixels faster. I spend the time it saves on what it can’t do: sitting with users, reading the room in a crit, and earning trust with compliance.

Where
Work and side projects
Tools
Claude, Figma, Mobbin, Obsidian
The rule
AI drafts, never decides

Ads before launch

Two AI reference sheets: a made-up office worker from the front, back and close up, and a phone showing a food-rescue app from nine angles

References

Two sheets keep it consistent.

One sheet fixes the person, the other fixes the phone and the real app screen. Every clip starts from both, so the same face and the same app turn up in every shot.

AI-generated

First cut

Filmed like a real customer.

A handheld selfie at a stall at night, like showing a friend a good find. Fifteen seconds to check the order works: the deal first, then how it works.

AI-generated

Full ad

A 30-second ad, before launch.

Price, food, the app, then the reason to care: less food thrown away. Made in an afternoon, so the message gets tested before anyone pays for a shoot.

What studies found

Faster with AI

Results from controlled studies, by kind of task

Product designFigma study, 2026~20% less time
Consulting tasksOrganization Science, 202625.1% faster
WritingScience, 202340% less time
CodingGitHub Copilot study, 202355.8% faster

Published research, shown for reference. The consulting study also found AI made people 19% less likely to solve a task it wasn’t suited to. Sources: Stewart et al. (2026)↗, Dell’Acqua et al. (2026)↗, Noy and Zhang (2023)↗, Peng et al. (2023)↗.

The flow

1 · Prompt

Start from real references.

The prompt starts in Figma, with real patterns from Mobbin as references, so the first draft already fits the product.

2 · Generate

Options in minutes.

One brief turns into a handful of directions. Choosing between them is still the design work.

3 · Test and synthesise

Learn from people sooner.

The best options go in front of users early. AI personas run a first, synthetic pass, so time with people goes to the harder questions. AI sorts the notes into themes, and the team decides what they mean.

My design harness

A harness is everything around the model: the process it follows, the tools it can reach, the rules it can’t break and the checks before anything gets to me. The model is the same one everyone uses. The harness is what makes it design my way.

The process

The steps, in order. My design-process skill takes every brief from screenshots to a tested redesign, one stage at a time.

The tools

What it can reach. Figma for designs, Mobbin for references, Obsidian for writing, and a browser to check what gets built.

The references

A benchmark to beat. Fresh examples from current apps make sure the basics match the best, so the effort can go into what makes the design its own.

The rules

Taste, written down. Every correction becomes a rule it checks before handing anything over, so it only needs saying once.

The reviewers

A second opinion before mine. A critic flags anything that looks generic or AI-made, and every screen is checked at each size, in both themes.

Show, then ship

Nothing goes live unseen. It shows me the work and waits for a yes.

The last call

What to keep, what to cut and what ships is always mine.

How I design

Four agent cards: Designer in Figma, Writer in Obsidian, Developer in HTML and CSS, Reviewer as a critic persona

A team of AI agents, each with one job.

A designer in Figma, a writer in Obsidian, a developer in code, and a critic that flags anything that looks AI-made.

A checklist titled “Rules every agent checks”, saved as memory

Taste becomes rules they can’t break.

Every correction becomes a rule the agents check before handing anything over, so it only needs saying once.

Figma first, then code that checks itself.

Nothing gets built until the design is approved. The build is then checked at every screen size in both themes, and fixed before it comes back.

Skills I built

A Claude skill is a short file that teaches Claude how I work. It reads the first line of every skill with each brief, opens the one that matches, and follows it. The process is written once and used on every project.

The whole redesign process, in one file.

Capture what exists, name the problems, research references, explore wide, decide narrow, test, then go round again until it holds up. The file holds the order and the rules. The decisions stay with me.

Read the doc →

Where it saves time

Edge cases

Every state, drafted first.

A bill can be due, paid early, split into months or overdue. AI drafts every state, so the time goes to the ones people actually hit.

UX writing

Copy that starts as a draft.

Errors, warnings and confirmations all need clear words. AI writes the first pass, and people decide what ships.

Onboarding prototype flow where a failed MyKad scan loops back to just that step

Prototyping

Prototypes in hours.

Cheap prototypes mean rougher ideas get tested earlier, like retrying one failed step instead of starting over.

What stays with me

Framing the problem

Deciding what’s worth solving at all.

Judgment and taste

Choosing between options, and knowing when good enough isn’t.

Trade-offs

Balancing risk, regulation, the business and the people using it.

Accountability

Owning what ships, whoever drafted it.

Like my projects? Email me jordychew93@gmail.com or LinkedIn↗