mcpField
P
procedural-generation

procedural-generation

github.com/jame581Development889167 installsUpdated 4d ago
Repository

About procedural-generation

A procedural-generation guide for Godot 4.3+ developers, covering reproducible randomness, noise terrain, dungeons, caves, and rule-based tile generation.

Key features

  1. 01Reproduce output with seeded randomness
  2. 02FastNoiseLite noise terrain
  3. 03BSP rooms-and-corridors dungeons
  4. 04Cellular automata cave generation
  5. 05WFC tile-constraint generation

Use cases

  1. 01Generate levels shareable by seed
  2. 02Create noise terrain or tile maps
  3. 03Build dungeon and cave layouts

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/jame581/godotprompter --skill procedural-generation

After installing, type/procedural-generationin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

references/
bsp-dungeons.md
cellular-automata.md
noise-generation.md
wave-function-collapse.md

Details

From the project README

Procedural Generation in Godot 4.3+

All examples target Godot 4.3+ with no deprecated APIs. GDScript is shown first, then C#.

Related skills: 2d-essentials for TileMapLayer usage, 3d-essentials for 3D terrain meshes, math-essentials for vectors and transforms, godot-optimization for chunk loading and performance.


1. Seeded Randomness

Always use seeds for reproducible generation. This enables shareable seeds, replay, and deterministic testing.

GDScript
# RandomNumberGenerator — per-instance, seedable
var rng := RandomNumberGenerator.new()

func generate_level(level_seed: int) -> void:
    rng.seed = level_seed

    var width: int = rng.randi_range(20, 40)
    var height: int = rng.randi_range(15, 30)
    var enemy_count: int = rng.randi_range(3, 8)
    var treasure_chance: float = rng.randf_range(0.05, 0.15)

# AVOID: Global randf()/randi() — one shared generator, randomized at startup
# and advanced by every other caller, so a sequence is hard to reproduce
# USE: rng.randf(), rng.randi(), rng.randf_range(), rng.randi_range()
C#
private RandomNumberGenerator _rng = new();

public void GenerateLevel(ulong levelSeed)
{
    _rng.Seed = levelSeed;

    int width = _rng.RandiRange(20, 40);
    int height = _rng.RandiRange(15, 30);
    int enemyCount = _rng.RandiRange(3, 8);
    float treasureChance = _rng.RandfRange(0.05f, 0.15f);
}

Tip: Generate a seed from a string for shareable level codes: var level_seed: int = "MyLevel".hash() (C#: GD.Hash("MyLevel"))


2. Noise-Based Generation (FastNoiseLite)

FastNoiseLite for height maps, biome distribution, 2D terrain. Key params: noise_type (Perlin / Simplex / Cellular / Value), frequency (lower = larger features), seed. For terrain, sample noise at each tile coord, threshold the value to pick a tile.

See references/noise-generation.md for the basic noise-map recipe, noise-type reference table, and 2D terrain + TileMapLayer walkthrough.


3. BSP Dungeon Generation

Binary Space Partitioning recursively splits a rectangle into smaller rectangles, carves a room inside each leaf, connects siblings with corridors. Produces grid-aligned room-based dungeons (think roguelike).

See references/bsp-dungeons.md for the full recursive partition + room placement + corridor connection algorithm in GDScript + C#.


4. Cellular Automata (Cave Generation)

Fill a grid with random walls/floors at ~45% density, then iterate "a cell becomes a wall if ≥ 5 of 8 neighbors are walls" 4-5 times. The result is organic cave shapes — no straight corridors.

See references/cellular-automata.md for the full GDScript + C# implementation with TileMapLayer integration.


5. Wave Function Collapse (WFC)

WFC is a constraint solver: given a tile set with adjacency rules, pick the lowest-entropy cell, collapse it to a valid tile, propagate constraints, repeat. Produces tile-rule-respecting output but is non-trivial to implement.

See references/wave-function-collapse.md for concept overview and a simplified GDScript + C# implementation.


6. Common Pitfalls

SymptomCauseFix
Same level every timeNot seeding the RNGSet rng.seed before generation
Same seed gives a different levelGlobal randf() / randi() — randomized at startup and shared with any other code that calls themUse a dedicated RandomNumberGenerator instance
Noise looks blockyFrequency too highLower frequency (try 0.01–0.05)
Caves are all wall or all floorfill_chance too extreme or too few iterationsUse fill_chance 0.40–0.50 and 4–6 iterations
BSP rooms overlapSplit position too close to edgeEnsure min_room_size buffer in split calculation
WFC contradiction (no valid tile)Adjacency rules too restrictiveAdd more allowed neighbors or implement backtracking
Generation takes too longProcessing entire map in one frameUse await get_tree().process_frame to spread across frames, or use a thread

7. Implementation Checklist

  • All generation uses a seedable RandomNumberGenerator, never global randf()/randi()
  • Seeds are stored with save data so levels can be reproduced
  • FastNoiseLite frequency and octaves are tuned for the game's tile/world scale
  • Large generation is spread across frames or run on a thread to avoid freezing
  • Generated TileMapLayer content uses terrain autotiling when possible (not hardcoded tile coords)
  • BSP dungeons verify all rooms are connected before finalizing
  • Cave generation runs a flood-fill to ensure reachability between key points
  • Player spawn point is validated to be on a floor tile, not inside a wall

Comments

Markdown supported