Generating puzzles is easy. Generating good ones is not

The first is a tool that makes puzzles for a game you already have: a sudoku grid with a unique solution, a nonogram that reads as a picture, a maze with exactly one path, a word ladder of a given length. That is an algorithm, and writing it is a proper programming job.

The second is a tool that invents a puzzle game. That one is largely marketing. Inventing a mechanic worth building a game around is a design act, and no generator has ever done it well.

Flockbay helps enormously with the first and is honest about the second.

A useful generator has a solver attached to it. You generate a candidate, you solve it, and you keep it only if it has exactly one solution, is not trivial, and cannot be brute-forced by a player who is not thinking. That validation loop is the difference between a puzzle game and a random grid.

You describe those constraints in plain language: the board is nine by nine, every puzzle must have exactly one solution, no puzzle should be solvable without at least one deduction of this kind. Flockbay writes the GDScript in your project.

Then you play the output. If puzzle after puzzle is too easy, you say so, and the difficulty constraint changes. That is what tuning a generator looks like, and it does not require you to read the code.

This is the subtle part. A generator cannot know what is hard, only what satisfies the rules you gave it. Difficulty gets expressed as constraints: how many deduction steps are required, which techniques are needed, how deep the search has to go before a contradiction appears.

Working out which of those correlates with a puzzle feeling hard to a person is the design work. It comes from you playing generated puzzles and noticing that the ones you enjoyed all had something in common.

So the loop is: generate, play, notice, tighten the rule, generate again. Fast iteration is worth more here than clever generation.

The same place. Board rendering, input, hint systems, undo, a timer, progression, saves — code, described in plain language and written into your project by Flockbay. It builds the scenes and the board's presentation too, from the brief you approved. You play it and say what reads wrong, or move a piece yourself in the Godot editor.

Flockbay is a desktop app for Windows, macOS and Linux with Godot bundled inside it, and your game runs locally when you press play. Generating thousands of candidate puzzles to validate is exactly the kind of work that benefits from running on your own machine rather than someone else's server budget.

Building needs an internet connection, since the chat and asset generation come over the network. Playing what you built does not. Free to download, and the world tools cost nothing. Flockbay is free, AI included: free hosted AI every day, or your Claude or ChatGPT plan, your own agent, or an API key.

The strongest puzzle games often do both. A hand-designed campaign teaches the mechanic in a deliberate order, and a generated endless mode gives players somewhere to go once they have learned it. The generator is not carrying the teaching load, so it does not have to be clever, only fair.

That split also protects you from the classic failure, which is a game made entirely of generated puzzles where nothing ever escalates and every board feels like the last one.

Building both is a matter of describing both. The campaign is authored levels in a deliberate order, which Flockbay builds as scenes and you play, reorder and adjust; the endless mode is the generator and its solver. They share the same rules, so the second one costs far less than the first.

More on this: Roguelike game maker: the run is a system, and systems are code