This skill splits every agent into layers that each pass one small value to the next: perceive, decide, path, steer and animate. Most weak-AI bugs come from a layer doing another layer’s job, such as a state that raycasts or a pathfinder called every frame.
It gives a table for choosing the decision model, a workflow that writes perception once, and runnable checks for A* correctness, state flicker and stuck agents.
What You Get
- A decision-model table: FSM, hierarchical FSM, behavior tree, utility scoring or GOAP, by situation
- A perception pattern with one sensor writing facts to a Blackboard at 5 to 10 Hz
- Gotcha lists for FSMs, perception, steering and pathfinding, including Godot AStarGrid2D defaults
- Three runnable checks: A* cost against Dijkstra on random grids, transitions per second, and progress toward a waypoint
Common Use Cases
- Choosing between an FSM and a behavior tree for a new enemy
- Fixing an agent that flickers between chase and patrol at a distance threshold
- Stopping A* from cutting diagonal corners through walls
- Making enemies chase the last known position, then search, after they lose sight of the player
What Makes It Unique
It supplies numeric checks and thresholds a reviewer can run: A* must match Dijkstra at h = 0, no more than about 2 transitions per second per agent, and at least 0.5 units of progress in 2 s. It names engine facts such as the Godot AStarGrid2D default diagonal mode and sends BT code, Unreal assets and Unity NavMesh to their own skills.
Under The Hood
The skill is instructions and two reference files, one on pathfinding and one on FSMs, perception and steering. It has no scripts. It tells the agent to read the project’s existing AI and navigation code first, then to write and run its own check code, such as an A* comparison on random grids and a headless run of agents, using the project’s tooling. It does not fetch web pages, call APIs or install anything.