How We Built Prism Break With AI: Names, Jail Bars, and Knowing When to Push Back

The AI built a full color-matching puzzle game. The human caught the things only a human would notice: a name that tried too hard, a visual metaphor that looked like a bug, and the complete absence of mobile controls.

April 16, 2026 · Vibe Arcade

Prism Break taught us about aesthetic judgment — that AI can build a mechanically sound game faster than you'd expect, but it has genuine blind spots when it comes to how things look and feel to a human sitting in front of the screen.

This is the story of a match-3 puzzle that went through a bad name, an ugly visual metaphor, a performance problem, and a missing input method before it became the game you can play today. The AI built it. The human caught what it couldn't see.

Want to play first, then read the build story?

▶ Play Prism Break Now

The Naming Problem

When we asked the AI to build a color-matching puzzle game, its first move was to name it. That's a reasonable instinct — you need to call the thing something while you're building it. The name it chose was "Synthwave Gems."

Vibe Arcade has a retro-futuristic aesthetic — neon colors, dark backgrounds, the whole 80s-arcade-meets-modern-browser vibe. The AI had clearly internalized this. "Synthwave Gems" was its attempt to stay on brand: retro music reference, gem imagery, the general feeling of belonging on the site. It sounded like it was generated by an algorithm told "name things in a retro-futuristic way." Which is exactly what had happened.

The feedback was blunt: stop forcing names to conform to the theme. Any name that makes sense for the actual gameplay works.

It came back with "Prism Break." Much better. It evokes the core mechanic — light, color, splitting, shattering. It sounds like a game you'd want to play rather than a genre tag someone slapped on a thumbnail. The AI's first instinct was brand consistency, when it should have been thinking about the player. Once we told it to think about the player, it nailed it.

The First Build

The AI shipped a working 8x8 grid of colorful gems. Click one, click an adjacent gem to swap. Match three or more and they shatter, remaining gems fall under gravity, new gems cascade in from the top. Each cascade wave multiplies your score — a single well-placed swap can chain into thousands of points.

Three game modes arrived out of the gate: a 60-second Timed mode, a 30-Moves mode for strategic planning, and a Daily Challenge that gives every player the same seeded board each day. The mechanics were solid and the scoring felt right. But the AI had made visual and interaction decisions that needed intervention.

The Jail Bars Disaster

Match-3 games need special gems — gems created by matching four or more in a row that have extra powers when matched again. In Prism Break, matching four creates a striped gem that clears an entire row or column when activated. The AI needed a way to show which direction each special gem would clear.

Its solution was to overlay jail bars on the gems. Thin dark lines laid over the gem graphics — horizontal bars for row-clearing gems, vertical bars for column-clearing gems. The logic was sound. Bars running in the direction of the clear would communicate the mechanic clearly.

It looked terrible.

The gems in Prism Break are bright, glowing, colorful shapes. Laying dark lines over them made the special gems look like they had a rendering glitch. The bars competed with the gem colors instead of complementing them. Worse, they made the special gems look less appealing than the regular ones — the opposite of what you want. Special gems should look powerful. These looked broken.

The AI's reasoning was correct but its aesthetic judgment was wrong. Overlaying opaque dark elements on luminous colored shapes kills the visual appeal. A human sees this instantly. The AI didn't.

The Fix: Stripes That Belong

We pushed back and discussed alternatives. The solution we landed on was striped patterns painted inside the gem shapes themselves — not overlaid on top, but integrated into the gem's visual identity. Horizontal stripes for gems that clear their row. Vertical stripes for gems that clear their column. The stripes follow the gem's silhouette, so they look like a property of the gem rather than something stuck on top of it.

The difference was dramatic. The striped gems look like powered-up versions of the regular gems. They pulse gently with a glow effect, drawing the eye without competing with the base colors. The directional information is just as clear — you can instantly tell which way a special gem will fire — but now the gems look like something you want to activate rather than something that went wrong.

The AI chose a technically valid approach that happened to look bad. The human caught it, and the collaborative discussion produced a better solution. The AI brought the implementation skill — painting SVG patterns inside gem shapes, animating the glow, getting the timing right. The human brought the judgment call that dark bars on bright gems is an aesthetic non-starter.

Performance Tuning

With the visual indicators fixed, we turned to performance. The AI had been generous with effects — particle bursts when gems shattered, glow layers on every gem, animated transitions on cascades, floating score numbers, screen shake on big matches. Each effect looked good in isolation. Together, during a cascade chain where dozens of gems are shattering and falling simultaneously, they bogged down the framerate.

This took a few passes. The approach wasn't to strip out effects entirely — the game's visual energy is part of what makes it satisfying — but to figure out which effects were pulling their weight. Some glow layers got simplified. Particle counts got reduced. Animation durations were tightened so effects resolved faster and overlapped less. The end result looks nearly as good as the maxed-out version but runs smoothly during long cascade chains on mid-range phones.

Performance tuning is an area where AI tends to add rather than subtract. It's good at generating effects, less good at evaluating whether the cumulative cost is worth the visual payoff. The AI can execute the fixes once you identify them, but it needs a human watching the screen to say "that cascade just dropped frames — dial it back."

Mobile Adaptation

The original interaction model was straightforward: click a gem to select it, then click an adjacent gem to swap. This works perfectly on desktop. A mouse cursor is precise, clicks are intentional, and the two-click pattern feels natural with a pointer device.

On mobile, it felt clumsy. Tapping a small gem to select it, then tapping the adjacent gem to swap — the two-tap interaction added friction that desktop players never felt. It wasn't broken. But it felt like a desktop game being tolerated on mobile rather than a game designed for touch.

The feedback was simple: add swipe gestures. The AI implemented touch controls so that dragging a gem in any direction initiates a swap. One gesture instead of two taps. The board suddenly felt responsive and direct in a way the click-click pattern couldn't match on a touchscreen.

This is another case where the AI built something that worked but didn't notice the gap. "Functional" and "good" aren't the same thing, and the difference was obvious the first time a human tried to play it on a phone. The AI doesn't pick up a phone and play its own game. The human does.

What This Taught Us About AI's Blind Spots

Prism Break is the clearest example we have of the collaboration pattern that defines how we build games. The AI built a complete match-3 game with cascade scoring, three game modes, special gem mechanics, and particle effects. Mechanically sound from the first build.

But every problem we fixed was something a human noticed by looking at the screen or holding a phone. The name sounded algorithmic. The jail bars looked like a bug. The effects were too heavy. The mobile controls were an afterthought. These aren't logic problems. They're taste problems, feel problems, context problems. They require knowing what it's like to be a person using the thing you made.

The AI ships fast. The human catches the gaps. That's the loop. The AI doesn't get defensive when you tell it the jail bars look terrible — it just proposes alternatives. The human doesn't have to build the striped gem system from scratch — they just have to know that stripes inside the shape are better than bars on top.

The naming conversation still makes us smile. "Synthwave Gems" is exactly what an AI would name a game on a retro-futuristic gaming site — technically appropriate and completely uninspired. "Prism Break" is what happens when you tell the AI to stop fitting in and start thinking about what the game actually is. Sometimes the best creative direction you can give an AI is permission to stop being so agreeable.

Play Prism Break — try Timed mode for fast-twitch cascades, or take on the Daily Challenge and see how your strategy compares to everyone else's on the same board.