Card Based Prompting
The Cold Start Problem: Pokémon, LEGO, and Vibe Coding
The “Cold Start Problem” is real: Staring at a blank prompt interface is paralyzing; without context or examples, we hit a local max immediately.
Mimetic learning is superior: Just as Midjourney users learned by watching the Discord feed, we need exposure to what’s possible before we can prompt effectively.
The LEGO/Pokémon realization: After a weekend rabbit hole of card openings and toy reviews, I realized that cards and bricks are just stackable components for software. (always have been)
The Prototype: I vibe coded a tool where users build single-page websites by selecting visual cards (code snippets, hex values) instead of typing into a void.
The Takeaway: We need more skeuomorphism in AI. Translating text prompts into visual, selectable “bricks” helps us make decisions faster and better.
This blog post was edited with Gemini
Card Based Web Design
One of the biggest struggles with any tool is the “Cold Start Problem.”
It doesn’t matter if it’s a canvas tool, a blank document, or a blinking prompt interface—staring at an empty screen is rarely helpful. If you don’t already know exactly what you want, you’re just sitting there, paralyzed by potential. Without a nudge to get started, you hit a local max very fast.
There have been attempts to solve this. I like the small snippets that suggest actions, or community tabs showing what others have created. But the more I think about it, the more I realize the true genius behind the launch of Midjourney wasn’t just the model—always was Discord.
I am still amazed that half of my prompting knowledge came from lurking in those chat channels. It was collective, live learning. I don’t think it would have ever occurred to me to use specific ISO settings or lighting hacks without scrolling through the feed and seeing exactly how others achieved their results.
The rise of ChatGPT had a similar effect. I had access to the OpenAI playground for ages, and in theory, I could have done cool stuff earlier. But I didn’t, because I didn’t see it. When ChatGPT emerged, we all started sharing screenshots of our weirdest prompts. That mimetic behavior created a collective vocabulary of what was even possible.
When I look at how people use AI tooling today, the gap often isn’t capability—it’s a lack of exposure.
Down the Rabbit Hole
Since I have some downtime before hunting for my next gig or project next year, I’ve been “vibe coding”—just playing around to see how fast I can create things. Last week, I built the first iteration of agen[+]cy (not a real product, just a playground), but I kept hitting that wall: I struggled to write a good prompt for what I actually wanted to build.
So, naturally, I wasted the weekend watching Pokémon card openings on Instagram Reels. Seriously, how is this so addictive? I’m not even into Pokémon. Should I open it? Should I keep it sealed? Fuck me. I fell into the hole.
But getting obsessed with cards turned out to be a great prompt in itself.
I started prompting around the idea of card reveals and generating assets on demand. Then, late Sunday night, I got lost on YouTube again—this time watching LEGO reviews (another guilty pleasure).
And it clicked: LEGOs are just components. Cards are just components.
How can I use cards as the components for a prompt?
The “Card-Based” Prototype
That led to this morning’s prototype. To get over the blank stare at an empty canvas, I changed the input mechanism. Instead of prompting a website exactly how you want it via text, you select a bunch of “cards” and use them as context for your generation tool.
It worked surprisingly well.
I built a small, feature-complete proof of concept where a user can prompt a single-page website (easy to render in Google AI Studio) by appending different types of cards. These aren’t just words; they are code examples, hex values, design tokens. The LLM is truly informed about what we want to achieve before a single word is typed.
You start the process like you’re sitting in front of a pile of LEGO bricks. You just grab a few, click them together, and see where things go. You get to a result a lot faster.
The Mini-Retro: 5 Lessons
It’s obviously not a perfect tool, but this was prototype #48 over the last two weeks. Here is what I learned:
We need more Skeuomorphism. Prompting AI through natural language makes intuitive sense, but we often lack the proper vocabulary to do it efficiently. We need skeuomorphic tricks—cards, bricks, dials—to bridge the gap and get us going.
Visual Inputs > Text Inputs. Translating prompts from text into something visual helps us make decisions faster. We need more “pre-rendered” prompts that we can select rather than write.
The “Stack of Bricks” Effect. Building a set of components extends your stack of LEGO bricks. It expands what you can do, but it also helps you rapidly learn what is possible—especially if combined with a community view.
Speed is Essential. There is a massive benefit to rapidly prototyping small, incremental ideas and testing them super early. Seeing where your thoughts go and making them real as fast as possible is necessary to escape the local max.
README & Game Design Documents First. The best design hack I’ve found lately is asking the LLM to translate my initial prompt into a README (great for breaking down technical work) and a Game Design Document (to capture core mechanics and experience loops) before writing code.
Github Link: here