I wanted to share a clip of how I design using only LLMs.
My thesis is that designers who go out of their way to have agents design for them, while tracking every design decision and critique along the way, will have an unfair advantage as frontier models continue to improve.
The idea is that this history of decisions becomes a growing dataset of your taste. The more data an agent has on what you like, dislike, and why, the better it gets at predicting and producing designs you'd actually want.
This means less time pushing pixels or tweaking shaders, and more time obsessing over the small details or solving harder product problems. As agents get better at making decisions you'd otherwise make yourself, you can delegate more work, tackle more things at once, and ultimately ship faster.
Here's how I'm approaching this right now:
1) Start with a design task. Create a spec or prompt for whatever I'm working on (e.g. a developer docs block or marketing page).
2) Have agents present work in phases. I use a personal design dashboard that I can plug into any frontend. Agents present their design explorations in a feed.
3) Make agents predict my choices. Before I select anything, I ask the LLM to predict which designs I'll prefer and why.
4) Select and critique. I choose what I like, explain what I don't, and specify what needs improvement.
5) Have agents reflect on my feedback. This is important. I ask them to compare my actual choices against their predictions, evaluate where they were right or wrong, and understand why.
6) Track preferences in a Kanban. Agents sort my design decisions into columns based on how they think I'm ranking the work. This gives me a dashboard of how they perceive my taste. It's useful because something I consider a 9/10 on Monday might become a 5/10 when a better design comes along Wednesday.
7) Continuously refine the taste log. Agents run daily evaluations of my decisions. At the end of each design cycle, they analyze what actually shipped and how it compares to earlier iterations. Over time, this builds a more complete picture of my preferences that can carry into the next project.
8) Reevaluate with better models. Whenever a more capable frontier model comes along, I have it analyze the entire taste log and the system itself to identify ways to improve the process.
The goal is to build a system that gets better at understanding my taste with every design decision, while giving me more time to focus on the parts of design that actually need me.
1
3
173
design has never been more fun.
it’s also getting harder.
we can generate infinite interfaces at the exact moment we’re losing certainty about what an interface should even be.
the hard part is deciding what to keep.
design is getting easier to produce.
good layouts, clean interfaces, decent typography, polished motion. the baseline keeps rising, and tools keep making execution faster.
almost everything is already available on the internet now.
design resources.
assets.
templates.
typefaces.
references.
components.
systems.
inspiration.
access is no longer the advantage.
what matters is what you do with all of it.
how you combine things.
how you interpret them.
what you ignore.
what you change.
what you make your own.
and most importantly, whether any of it creates actual impact.
that is where the real differentiators start:
craft, taste, and point of view.
craft is caring enough to notice what most people skip.
taste is knowing what to remove, what to exaggerate, what deserves attention, and when something is already enough.
point of view is harder.
it is the reason your work feels like yours.
the choices you keep making.
the things you refuse to do.
the references you pull from.
the problems you believe are worth solving.
the way you interpret the same brief differently from everyone else.
ai can generate options.
templates can give you structure.
systems can make production faster.
the internet can give everyone access to the same ingredients.
but access does not create great work.
judgment does.
when everyone can find the same resources and use the same tools, technical ability becomes less scarce.
what becomes scarce is knowing how to use them with intention.
the future of design will not be won by whoever has the biggest resource folder or produces the most screens.
it will belong to people who can take what is available to everyone and turn it into something only they would have made.
something useful.
something memorable.
something that actually moves the business, product, or culture forward.
the ingredients are everywhere now.
the difference is you.
1
4
265
A couple months ago we shipped Payva 2.0 for merchants.
Now it's the customer's turn.
We rebuilt the customer experience around the question: how do we make tracking an order feel obvious?
What's new:
• A full timeline for every order, so you always know where your payment stands
• Manage your payment methods in seconds
• Add multiple cards and bank accounts to your wallet
• Pull your documents in a couple clicks, not a support ticket
More insight. Less friction. And a lot more coming.
Log in now at app.payva.com
#Payva #BNPL #Fintech #ProductUpdate #CreatorEconomy
congrats to @putrikarunian on the launch of @lunagraphHQ 🚢
if you like html design canvases like @paper, I definitely recommend checking this out
it's happening 🥹
@lunagraphHQ is open
not just a design canvas, but a visual editor for React
this means you can design the full experience: from exploration to defining every state, interaction, and flow. and when it feels right, hand devs nicely packaged code, UX and all
come design with code this weekend 🌸
(sound on 🔊)
2
4
454
elastic table apron
live width sensing → badges reshape, nothing ever clips
made w/ @claudeai
animated w/ @motiondotdev
designed w/ @paper
table apron
bottom count pill + clearable filter badge
pins to table edge or viewport
made w/ @claudeai
animated w/ @motiondotdev
2
3
766
filter caddy🏌️♂️
auto bookmark your table views
made w/ @claudeai
animated w/ @motiondotdev
table apron
bottom count pill + clearable filter badge
pins to table edge or viewport
made w/ @claudeai
animated w/ @motiondotdev
1
8
781
supercards: paste one txt file into any LLM and your answer comes back as a clean, one-idea-per-block breakdown instead of a wall of text.
built for my ADHD. still rough, curious what you get.
the prompt:
"supercard me on something id find interesting: berafoot.com/llms.txt"
1
6
506






