AI logic puzzle maker: the rule is small, the consequences are not

A logic puzzle game is one rule and a hundred consequences. The rule takes an afternoon to implement and the consequences take the rest of the project, which is why implementation should not be your bottleneck.

The shape of the genre

The great logic puzzle games have a mechanic you could explain in one sentence. Boxes can be pushed but not pulled. Time rewinds and you meet yourself. Colour determines what you can stand on. That is the entire rule set.

What follows is the design, and the design is finding the situations where that rule does something surprising. That is a search problem carried out by a person, and it is slow, and it does not compress.

Which means the implementation should be nearly free, and traditionally it is not. Grid movement with correct edge cases, undo that actually restores everything, state that rewinds cleanly — these are fiddly and they consume the exact energy you needed for puzzle design.

Get the rule running the same evening

You describe the mechanic to Director 1 and it writes the GDScript into your Godot project. "The player moves on a grid, can push one block at a time but never pull, and pushing into another block does nothing."

Then the parts you would have forgotten: undo that steps back through every change, a reset key, a move counter, a solved state. Say you want them and they exist.

Undo in particular is worth calling out. In a puzzle game it is not a convenience, it is the primary verb — players undo far more than they move — and a subtly wrong undo makes a good puzzle feel broken.

Designing the puzzles is the actual project

You lay out levels in the builder, by hand. In a flat project that is a grid in the most literal sense: ground is a tile layer you drag on and drag off, and the tiles, the edges and the collision are all derived from which cells you covered, so the room is exactly the shape you painted and a rebuild never reshuffles its edges. The blocks, switches and doors on top are sprites placed with the mouse at an exact position, turn and size.

Nothing generates a good puzzle, because a good puzzle is a piece of misdirection aimed at a person, and the aiming is the craft.

The reliable method is boring and works: build a situation, solve it yourself, then find out whether it teaches. Every puzzle should introduce one idea, and the puzzle after it should assume you learned it. When you find yourself explaining a level in words, the level is not doing its job.

Order matters as much as content. The same twenty puzzles in a different sequence are a different game, and often the difference between people finishing and people quitting at level six.

Testing on other people, early

You cannot judge the difficulty of a puzzle you designed. You know the answer, permanently, and there is no way back to not knowing it.

So the only instrument is someone else playing while you say nothing, which is uncomfortable and necessary. Watch where they stall. Half of what you thought was elegant is invisible.

The game runs locally on the bundled Godot, so handing it to someone on a laptop needs no connection and no setup.

Practical facts

Flockbay is a desktop app for Mac and Windows with Godot inside it. It asks 2D or 3D before the project exists — two pictures, no default — and for this genre both are real answers. The grid version is the flat world's natural shape and gets the flat world's own tools. The spatial version, where the rule plays out in three dimensions, is less crowded ground and is one choice away rather than the thing the app is built around.

Building needs an internet connection. The download and the builder are free; Indie at $10 a month unlocks the AI assistance.

Presentation does more than you expect

Two puzzle games with identical mechanics can feel completely different depending on what the pieces are. Blocks and grids read as an exercise. A room, a light and a door read as a place.

Since the world is built by hand here, dressing a puzzle as somewhere real is cheap in either kind of project. In a flat one you paint the floor of a room and stand a lamp and a door in it from a Catalogue where 55,073 of the 60,648 objects are sprites, searchable by name with a picture of each. In a 3D one you have terrain, water and a time of day to do it with. Either way it changes how long people are willing to sit with a hard puzzle.

It also affects how a rule is understood. A mechanic explained by what the objects obviously are needs no tutorial text, and tutorial text in a puzzle game is nearly always a sign the level failed.

Questions

Will it design puzzles for me?
No. It implements the rule and the surrounding systems. Designing a puzzle is aiming misdirection at a person and that stays with you.
Can I get proper undo?
Yes, and ask for it early. Undo is the most-used action in a puzzle game and a subtly wrong one makes good puzzles feel broken.
Does it have to be 3D?
No, and 2D is not the fallback. You pick one when you create the project. A grid puzzle in a flat world is painted tile ground with sprites for the pieces; a spatial puzzle in 3D is less crowded ground. Same app, same Director 1, either way.
Can I change the core rule later?
Yes, by describing the change. Doing that cheaply is useful, because the first version of a mechanic is rarely the interesting one.
What does it cost?
Free download, free builder. Indie at $10 a month unlocks the AI assistance, which the free plan lets you try first.