AI stealth game maker: the guard has to be readable

A stealth game is only fair if the player can model the guard. Where he is looking, how far he can see, what he heard, what he will do about it. When that model is accurate, being caught feels like a mistake and the player tries again. When it is not, being caught feels arbitrary and the player quits.

Everything else in stealth design follows from that. Vision cones exist so seeing is legible. Alert states are staged so escalation is visible. Sound has a radius so the player can reason about noise. These are not conventions for their own sake; they are the genre solving its information problem.

Building a stealth game therefore means building a system a player can read, and then building levels that make reading it worthwhile.

Perception with distance, angle and line of sight. Alert states with entry and exit conditions and a search behaviour between them. Patrol routes with waits and glances. Noise from footsteps, doors and objects you throw. Hiding, cover and the light level that decides whether cover works. Detection meters, if your design shows them.

Director 1 writes these into your Godot project as GDScript. You then play, and the useful feedback is exactly the kind a player gives: he saw me through a wall, he gave up too fast, I could not tell he had noticed me. Each of those is a change.

You are not required to read the code. Being non-technical does not make you a worse judge of whether a guard is behaving sensibly; if anything it makes you a better one, because you are watching the game rather than the implementation.

You build stealth levels by hand in the builder, and this is where the game exists. Patrol paths that cross at an inconvenient moment. A vent that offers a route at the cost of information. A courtyard with two guards whose cones leave a gap that closes every twelve seconds.

That timing is authored. It cannot be generated, because the whole point is that it was arranged by someone who knew what they were arranging. Terrain, buildings, cover and light all go in directly, under your hands.

Light matters enough here to plan around. So does verticality, and so does the sight line from the balcony you thought was decorative.

Stage Engine is a desktop app for Mac and Windows. It installs on your machine and bundles Godot, so the game runs locally. Your project is an ordinary Godot project.

An internet connection is needed while building, because Director 1, asset generation and world data come over the network. Playing what you built does not need one.

Free to download with a free plan. Indie is $10 a month, Pro is $50, for more usage rather than different software.

The oldest argument in stealth design is what happens when you are caught. Instant failure and a reload makes the game a series of restarts, and players start save-scumming rather than playing. Full combat makes stealth optional, and once it is optional most players stop bothering with it.

The answer most good stealth games arrive at is a middle state: being spotted is expensive but recoverable. You can break line of sight, guards search and eventually give up, and the level continues in a worse condition than before. That keeps the tension without turning the game into a checkpoint simulator.

Getting that middle state right is a tuning job. How long guards search, how far they spread, how much they remember, whether the alert decays fully or leaves them jumpy. Each is a number, each changes the whole game, and each is found by being caught on purpose and seeing what happens next.