Make a puzzle game with AI: the rule is small, the levels are the work

Take any good puzzle game and write down its mechanic. It fits in a sentence. Blocks slide until they hit something. You can swap places with your clone. Colours combine when adjacent.

Now write down what makes the game good. That takes a hundred pages, because it is the order of the levels, the moment the rule reveals a second implication, the level that teaches without a tutorial, the one that is only hard because you assumed something.

Traditionally the sentence takes you two weeks to program and the hundred pages never get written because you ran out of energy. Reversing that is what this is for.

You describe it to Flockbay and it writes GDScript in your project. Grid movement, collision rules, win conditions, undo, reset, move counters.

Undo is worth asking for on day one. Puzzle games without undo are punishing in a way that has nothing to do with difficulty, and building it in later is much more awkward than having it from the start.

Then you play the mechanic in an empty room, doing nothing but pushing the pieces around, until the interaction feels clean. Ambiguity in the rule shows up here and is much cheaper to fix now than after thirty levels exist.

You build them in the builder, by hand, arranging the pieces. This is not a limitation. A generated puzzle level is either trivially solvable or accidentally impossible, and the interesting middle — the level whose solution feels like a discovery — comes from a person who knows what the player has understood so far.

The first four or five levels are the ones that matter most and they take the longest. They are teaching. If the player learns the rule without being told it, the rest of the game works. If they do not, nothing later saves it.

Play your own levels a week after making them and you will find out which ones actually teach.

More on this: Generating puzzles is easy. Generating good ones is not

Flockbay is a desktop app for Windows, with open-source Godot bundled inside. Your game runs locally on your machine. Not a browser tool.

Puzzle games are the genre least likely to strain any machine, so local performance is not the reason to be here. Owning a real Godot project on your own disk is.

Building needs an internet connection because the chat and asset generation are network services. Playing what you built does not. Flockbay is free on your own AI: your Claude or ChatGPT plan, your own agent, or an API key.

Later than you want to. The urge to add a second rule usually arrives when you have run out of ideas for the first one, which is normally about twenty levels before the first one is actually exhausted.

When you do add it, the good levels are the ones where the two rules interact rather than alternate. Those are hard to find and they are the reason to make a puzzle game at all.

More on this: Make the puzzle game you keep describing to people

You cannot test your own puzzles for difficulty. You know the solution and you cannot unknow it, so your sense of how hard a level is becomes useless the moment you finish making it.

What you can test is clarity. Watch someone else play, say nothing, and note the exact moment they get stuck and what they try instead. Their wrong attempts tell you what your level appears to be about, which is often not what you intended.

One person playing silently is worth more than a page of your own notes. This is the cheapest and most reliable quality step in the genre.