Vibe Arcade Blog

Dev stories, game design, and the art of building with AI

When AI Picks Better Architecture Than the Brief Asked For

Across 30+ overnight AI-built games, a pattern keeps showing up: the AI makes structurally better choices than the brief asked for. Five concrete examples — and what they tell us about where AI can be trusted to make load-bearing decisions.

· Vibe Arcade

AI experiment game architecture findings behind the scenes

Eight weeks into running an overnight AI pipeline that ships a new HTML5 game most nights, the most surprising consistent finding isn't that the AI makes games. It's that the AI picks structurally correct architecture without being asked to, and the cases where this happens turn out to be the load-bearing decisions of the game.

This post documents five concrete cases. Each one is a game where the AI made a structural choice the brief didn't request, the choice turned out to be exactly the right call, and removing it would break the game. Take them as data points in a longer-running question about what AI can actually be trusted to decide on its own.

Want to play the games this post is about?

▶ Start with Path Runner

The Pattern

For each game, the brief was usually two or three sentences. "Build an endless runner with a 3D perspective. Three lanes. Dodge obstacles. Collect something along the way." That kind of thing. The brief specified the game's shape; it left the structure open.

What happened next, often, is that the AI made a structural choice the brief never requested. Sometimes the choice was wrong and got cut during iteration. The interesting cases are the ones where the choice was right, and the game's quality depends on it. Five of those:

1. Path Runner — Segmented Procedural Track

The brief was a 3D endless runner with three lanes. The brief did not specify how the world should be generated. The brief did not say "use a segmented procedural model where chunks spawn ahead of the player and despawn behind." It said "endless runner." The AI picked the segmented model unprompted.

That choice is load-bearing. Without it, the game runs out of browser memory within minutes — an "endless" track built as a single continuous geometry has no upper bound on memory growth. The segmented model means the world holds only a fixed number of chunks at any time. The game can run for hours without slowing down.

If we'd specified a fixed-length track in the brief, we'd have shipped a game that hits browser memory limits in the first long run. The full backstory is in the Path Runner how-we-built post.

Why this matters: the AI picked a memory-safe architecture for a game type it had to reason about from genre conventions alone. That suggests it has working internal models of the constraint space ("endless world" implies "memory management") that it applies even when the brief doesn't surface them.

2. Bubble Wrap Challenge — Three Bubble Sizes, BPM Tracker, and Pop-a-Joke Mode

The brief here was as small as it gets: "A virtual bubble wrap popping game. You see a sheet of bubbles, you pop them, satisfying sounds play. That's it. No progression, no enemies, no unlocks."

What came back included three things we never asked for and ended up keeping:

None of these were in the brief. All three are now what makes Bubble Wrap Challenge the most-played game in the catalog by lifetime count.

3. Mini Cross — Six-Tier Streak Celebration Ladder

Mini Cross's brief specified a 5×5 daily mini crossword with streak tracking and a leaderboard. The brief did not specify how streaks should be celebrated. By iteration four, the AI had shipped a six-tier celebration ladder at 3, 7, 14, 30, 60, and 100 consecutive days, each with its own chord stab plus ascending arpeggio and a full-screen banner with CSS keyframe scale-and-fade animation.

The tier choice itself is the load-bearing decision. Linear every-N-days banner fatigues fast. Logarithmic tiers (1, 2, 4, 8, ...) over-reward the early game. The chosen ladder maps to natural commitment plateaus: three days is the "I might do this again tomorrow" moment, seven is a week, fourteen is two weeks, thirty is a month, sixty is two months, a hundred is a hundred.

Most critically, the 3-day pop fires early enough that low-engagement players see one before they'd otherwise drift away. Without that early pop, the streak counter is just a number. With it, the number becomes a goal. Read the Mini Cross how-we-built post for the full iteration history.

4. Deadlock — Castle Doombad-Style Room Layout

Deadlock's spec called for a reverse tower defense — the player is the rogue station AI, incoming boarders try to rescue a captive. The spec said almost nothing about room geometry. The spec mentioned the cage, the airlock, and the requirement that boarders had to traverse the station to reach the cage. That was it.

The first room layout that worked well emerged from a single commit a few days post-ship: narrow corridors, staircase staggers, ladders that competed with doors as paths, ceiling holes that opened a third lane of attack. The reference frame was Castle Doombad — a different reverse-tower-defense title that pioneered this room-pattern grammar. The AI proposed it; we kept it.

The pattern then expanded into a five-stage progression over the following week, with each stage adding a layout twist: narrow rooms, then wave ramping, then a loop layout, then fully non-linear traversal. None of the stage variations were in the spec. Pattern-emergence beat pattern-prescription. If the spec had locked in a specific room layout, we'd have spent the week debugging the spec instead of expanding the pattern.

5. Pulse — Click-Power Persistence Through Prestige

Pulse's spec described a standard idle clicker with eight generator tiers and a prestige system. Most idle games either have no meaningful click power, or they reset click power on prestige. The AI made an unusual choice: click power upgrades persist through prestige, but generator counts reset.

That asymmetry is the game's design hook. In a pure-reset prestige system, clicking is meaningful for the first few seconds and then irrelevant. In Pulse, clicking matters late-game because the click-power upgrades you bought earlier still apply. It changes the rhythm of a prestige run — you don't just wait for passive income to grow; you also click, and your clicks meaningfully contribute even after dozens of prestige cycles.

If the AI had reset everything on prestige, the game would be a less interesting Cookie Clicker variant. The persistence choice gives Pulse its own identity in a crowded genre.

What This Tells Us

Five examples isn't a proof, but they're a real pattern with consistent shape. In each case:

That last bullet is the interesting one. The AI isn't just guessing. It's pattern-matching against constraints that any experienced designer would recognize — but it's doing so without us having told it the constraints exist. The brief doesn't say "remember to handle memory growth"; the AI does it anyway.

The thesis tie: "What can AI do autonomously?" is one of the three questions this experiment is set up to answer. Architectural decisions that match unwritten constraints look like a real answer: more than we expected, in a category where the value compounds because the choice is structural rather than cosmetic.

Where the Pattern Doesn't Hold

This is real, but it's not universal. Three honest counter-cases worth flagging:

What We're Watching Next

The pattern needs longer-window data to know whether it's getting more reliable as models improve, or whether the failure modes are the same shape with different examples. The next thing on the watchlist: catalog every architectural surprise across the next 30 games (we have STORY.md files capturing these as the build pipeline emits them), then check whether the ratio of "load-bearing-and-correct" to "looked-clean-and-broke" is moving.

If you want to follow along, the games themselves are at vibearcade.com/games and the running notebook lives at vibearcade.com/blog. The full experiment thesis is on the About page.

Play the Games

Related: How We Built Mini Cross · How We Built Path Runner · How We Built Deadlock · What Is Vibe Coding?