I just finished building ajays.quest and it took me way longer than it should have.

Not because of the logic. Not because of the data. Because of the pixels. The spacing. The "feel" of a hover state. The thing that makes you look at a screen and think this was made by someone who gave a damn.

Backend work with AI? Smooth. I describe an API, Claude Code writes it, I test it, done. Frontend work? I'm in Figma for hours, iterating on layouts that "almost" work, then translating that intent to an AI that keeps giving me something slightly off.

The truth is: AI is great at logic, but taste is a different beast entirely.

Here's how I think about it.

Backend work has right answers. A function either returns the correct data or it doesn't. An API either handles the edge case or it breaks. You can describe the desired behavior precisely, and a model can execute it. Right or wrong. Pass or fail.

Frontend work has best answers. There are a thousand ways to design a button. None of them are "wrong" — but only a handful feel right. The difference between a good UI and a great one isn't logic. It's rhythm, hierarchy, negative space, and a dozen micro-decisions that compound into something that either feels polished or feels like a template.

AI is exceptional at finding right answers. But best answers require exploration, dead ends, and the kind of intuition that only develops after you've made the same mistake fifty times.

---

For ajays.quest, I ended up with a two-stage process:

Stage 1: Exploration with Variant AI

When I'm starting a new component or page, I use Variant AI to brainstorm. I'll describe what I'm trying to achieve — "a quest-style progress tracker with RPG energy" and let it generate a bunch of different directions.

Most of them are wrong. That's the point.

The goal isn't to get the final design. It's to see the edges of the possibility space. What does this look like if it's maximalist? Minimalist? Playful? Serious? Variant gives me divergent options fast, and I can cherry-pick the elements that resonate.

Stage 2: Execution with Claude Code

Once I know what I want, once I've been through Figma, refined the layout, nailed the spacing - I bring it to Claude Code.

At this stage, the problem has a right answer. I'm not asking "what should this look like?" I'm saying "here's exactly what I want, build it." Claude Code is excellent at this. I describe the component, reference the design, and it writes clean code that matches.

The mistake I made early on was trying to use Claude Code for both stages. I'd describe a vague concept and expect polished UI. It doesn't work. The model doesn't have taste, it has pattern-matching. Feed it a clear target and it hits it. Feed it ambiguity and it gives you Bootstrap.

This isn't just about ajays.quest. It's about how we'll use AI tools for the next decade.

The products that feel magical will still require human judgment when you're searching for the best answer. AI will handle the execution, the code, the optimization, the scale — but the creative direction, the "this feels right" moment, that stays with us.

For founders and builders, the skill isn't "prompting AI." It's knowing when you're solving for a right answer vs. a best answer. And having the patience to do the messy exploration before you hand it off.

I spent hours in Figma on ajays.quest that I "could have" spent prompting Claude Code. On paper, it looks inefficient. But those hours are what gave the final product its feel. The AI finished the job. But only after I'd done the work of knowing what "finished" should look like.

And that's enough.

Until next time,

Ajay

A few tools and collections I’ve built (or obsessively curated) over the years:

These are tools and services I use personally and professionally — and recommend without hesitation: