Building a Complete Bejeweled Clone with AI in 90 Minutes

I've been following the AI-assisted coding space pretty closely, and I wanted to see just how far I could push it for game development. Could an AI build a complete, polished game from scratch in a single sitting?

Turns out, yes. I built GemMatch, a complete Bejeweled-style match-3 puzzle game, in about 90 minutes using Claude Opus 4.6 through Kiro CLI. Every single line of code was written by the AI. I didn't write or edit a single line of code by hand.

Why Godot?

I went with Godot 4.x and GDScript for this experiment. Godot's _draw() API makes it trivial to render shapes procedurally, and GDScript is simple enough that the AI can generate correct code without fighting a complex type system or engine architecture. This turned out to be a good call -- the entire game is about 1,480 lines of GDScript across six scripts, and it worked on the first or second try for most features.

The Process

I started by asking Claude to create structured specs before writing any code. This is something I've found works really well with AI coding -- if you let it plan first, the implementation goes much smoother. Claude generated three documents in .kiro/specs/: requirements (user stories with acceptance criteria), an architecture design, and an implementation task list.

From there, the build happened in five iterations:

V1 -- Basic game. A minimal working match-3 with click input, match detection, gravity, and scoring. Functional on the first run after fixing a couple of type inference errors.

V2 -- Animations and sound. Tween animations for swaps, removal, and gravity. Screen shake, score popups, and procedural sound effects -- all synthesized at runtime, no audio files.

V3 -- Full feature set. Special gems (Flame, Star, Hypercube), drag-to-swap, a hint system, level progression, combo text, particle effects, and unique gem shapes.

V4 -- Polish. Animated logo, procedural chiptune background music, and a title screen with floating gem shapes.

V5 -- Title screen and menus. A proper title screen with a 5-item menu (New Game, How to Play, High Scores, Demo Mode, Quit), full keyboard navigation throughout menus and gameplay, persistent high scores saved to disk, a How to Play showcase screen showing all gem shapes and special types, an optimal demo AI that plays as an attract screen, animated level backgrounds with 5 geometric patterns that cycle each level, and a selection ring that follows the gem during swap animations.

Gameplay shown here at 2x speed.

Each iteration was tested by launching Godot directly from the terminal via Flatpak:

flatpak run org.godotengine.Godot --path /var/home/dylan/Projects/gemmatch

Claude even handled installing Godot via Flatpak when it wasn't already on my system.

What Impressed Me

The thing that stood out most was how well Claude handled the iterative refinement. Most features worked correctly on the first attempt. When there were errors, they were almost always minor type annotation issues specific to Godot 4.6 -- not logic bugs. Claude would see the error output, understand the issue immediately, and fix it.

The procedural audio generation was particularly impressive. I just asked for "sounds" and Claude generated a complete synthesis system that creates AudioStreamWAV buffers with configurable frequency sweeps, envelopes, and noise mixing. The background music generator creates an 8-bar chiptune loop at 95 BPM with melody, bass, and pad voices. All of this runs at startup and produces surprisingly pleasant audio.

The special gem system also worked correctly on the first implementation, which I honestly didn't expect. The logic for detecting match patterns (is this a match-4? match-5? L-shape? T-shape?), creating the appropriate special gem, and then triggering special gem effects when they're subsequently matched -- that's genuinely complex game logic, and it was right the first time.

The demo mode AI was another highlight. Claude built a scoring heuristic that evaluates every possible swap on the board and picks the optimal move -- prioritizing special gem creation (5-in-a-row for Star Gems, L/T shapes for Hypercubes, 4-in-a-row for Flame Gems) and triggering existing specials. It plays impressively well and makes for a great attract screen.

The Numbers

• Total time: ~2 hours across two sessions (Feb 18-19)
• Lines of code: ~1,480 lines of GDScript across 6 scripts
• External assets: 0 (everything is procedural)
• Hand-written code: 0 lines

For a 90-minute session, I'm pretty impressed with the result. A fully playable, polished game with special gems, animations, sound, music, and level progression. Choosing a simpler engine and taking a spec-first approach made a big difference.

Try It Yourself

The full source is on GitHub: github.com/dylanmtaylor/gemmatch

You'll need Godot 4.2+ to run it. If you're on a Linux distro with Flatpak:

flatpak install flathub org.godotengine.Godot

Every line was AI-generated, and I think that's worth being transparent about. We're at a point where AI can build a complete, polished game from scratch in the time it takes to watch a movie. That's wild to me.