Shannon Miwa

A note on AI and "AI First"

Designers love a new tool, especially one that opens new ways of thinking. LLMs are shiny! shiny! new! new! and the possibilities are quite real. Having spent the past year building, experimenting, and researching, I've developed a clear-eyed understanding of what they can do and where they fall short, including how they work, how they were built, and the broader economic and environmental implications of their use.

There is good and bad, and still much to discover, but what I do know is that automatically reaching for the same tool before considering the problem fully is hardly good design thinking. My first priority as a designer is deep understanding: gather context, ask questions, research, talk to people, poke around, and generally, fuck around and find out why.

I maintain a healthy curiosity and a healthy skepticism about AI-based processes. My approach to design prioritizes the need of the problem over the tooling: it is quite easy to write a prompt, harder to ask a good question, and even harder to figure out if it’s the right question.

Responsible use matters to me, not just on an execution level but on a human one. There are no clean datasets, no hallucination-free models, and all of them use vast amounts of energy, offsetting huge environmental costs to those with no say in the matter. "AI first", or default without consideration, is not a neutral act. Good design is not a generated outcome—it requires curiosity, judgment, and genuine care, and I strive to bring that to everything I do as a designer.

A Handful of Tasters

A space for experiments, not full case studies, ongoing and unfinished by design. Right now that means three parallel threads: building a card game app with Claude Code, stress-testing a real design system as an AI skill, and chasing a specific art direction across different AI design tools. Different questions, same underlying curiosity: where AI genuinely helps design and development work, and where it still falls short.

Cards with Friends

An app for playing card games remotely, built with React and Supabase using Claude Code. The real goal is flexibility: getting as close as possible to what a physical deck of cards can do, where you can play any game and bend the edit rules on the fly. Long term, the idea is to configure entirely new games through prompting, not hard-coded rule sets. (Deck illustrations are their own case study, see Standard American Playing Cards case study.)

Check out WIP app →

Tooling

Claude Code
Github
Supabase

Status

Main application functionality working
Test rooms working, history and settings added
Blackjack game fully functional
Addressing game mechanics, rules and rule changing
Improve aesthetics, less clunky UI

Building this is as much about practicing a working relationship with an AI coding tool as it is about the app itself. I want to stay in the driver's seat rather than defaulting to "yes, install this for me" on every prompt. If it's a command I already know, I type it myself instead of auto-accepting. When it comes to things like environment secrets, I ask it to walk me through how they actually work and what could go wrong if I get them wrong, rather than treating them as a black box to trust. Along the way I'm picking up real working knowledge I didn't have before: how YAML files and deployment jobs fit together, and a much better grasp of working with databases. Small habits, but the kind that keep me actually learning instead of just shipping.

BPO Manager Clickable Prototype

A prototype built after wrapping up my work at Assembled, using real production files and the actual design system to test something I don't always get to fully explore: building a design system into a working "skill" for AI tools, and seeing how far that can go with real product complexity. Less about the original design decisions (covered in the case study here), more about probing the edges of what AI can and can't do inside actual design work.

View WIP (password: bingo) →

Tooling

Claude Code
Figma

Status

Primary page layout and navigation working
Main components and colors added
Light and dark theme support
Dashboard mockup
Linking pieces together
Fully match table layout
Add heatmaps and performance pages
Hook up components and create walkthrough paths

This project is less about changing how I work and more about mapping where AI tools are actually useful in design work versus where they still fall short. Lightweight, mechanical tasks work well: randomizing values, matching up math, small time-saving cleanup. Anything requiring nuance is harder. Asking for a single line to be added or removed often takes several attempts to land right, and getting good results depends on how thoroughly the design system states are already fleshed out. The more specific and well-defined the system, the better the output. A big part of this is learning how to actually describe design changes in a way the tool can act on accurately, which is turning out to be its own skill worth building.

Portfolio Art Direction

A separate design exploration for this same portfolio content, testing four different AI design tools against one specific reference: Japanese commercial art from the 1920s and 30s. Minimal, deco, patterned, duotone, asymmetrical but balanced, mixing organic and geometric shapes with a playful color sense. The question isn't whether these tools can build a portfolio site, it's whether any of them can actually interpret a specific, unusual aesthetic rather than defaulting to their own.

Tooling

Claude Code
Claude Design
Figma Make
Loveable

Status

Initial passes, 2-3 rounds on each system
Moving most promising options forward, pushing towards uniqueness
Capture screenshots, create gallery

Claude Design and Figma Make struggle out of the box, and several rounds of prompting later, the results are still far from the reference. Neither seems to genuinely interpret the source material so much as apply a generic layer on top of a standard layout. Claude Code takes a different approach, building out a moodboard before executing, which helps early on, but it still struggles with simple pattern requests once it gets to actual execution. Lovable comes closest on pure design sensibility, offering pre-built directions that are noticeably more informed by the reference, including theming and placeholder copy that pick up on the source material in ways I didn't expect. But it struggles more with actually building out the execution.

None of the four have produced anything finished or shippable yet, and none break meaningfully away from a fairly standard portfolio layout, though Lovable gets closest. The real finding isn't which tool "wins," it's where each one's interpretation breaks down: some tools are better at surface-level styling, some are better at structural building, and almost none are good at both when the aesthetic target is specific and unfamiliar to them.

AI makes it easy to produce more work, faster. It's much harder to make that work meaningfully better. The difference comes down to editing: knowing what's actually worth keeping. But you can't edit well without reading and writing well first, reading closely enough to know what a tool actually did, writing precisely enough to give it real intent to work from. Anyone can generate more. Fewer know what to do with it.