mcpbeat Sign in

Bevy Procedural Generation Agent Skill

Bevy ECS skill for procedural generation. Keywords: Bevy, procedural, seed, dungeon, terrain, ECS.

2.3k tokens
context cost
SKILL.md, read when the skill is used
2
files
plus reference files
0
copies elsewhere
how many repositories repackaged it
854
stars on the repo
on the repository, not the skill itself

Install

one command, takes just this skill from the repository
npx skills add https://github.com/thedivergentai/Game-Design-Agentic-Skills --skill bevy-procedural-generation

What comes with it

540 bytes besides the instruction
references/ecs-ownership.md

The instruction itself

16 sections, as written by the author

Bevy procedural generation

When to use

Load this when a level, loot table, or map should be reproducible from a seed, and you need the ECS shape. Spawning the finished layout uses Commands on Startup or a later schedule you already use. Saving the seed belongs on bevy-save-load.

ECS shape

  • Components: cell facts you define, spawned with Commands.
  • Resources: one RunSeed for the whole run. The generator reads it; it does not invent a Bevy RNG type. rand 0.9 supplies StdRng.
  • Systems: insert the seed on Startup, then spawn from that data. .chain() so the spawn system sees the seed.
  • Schedule: Startup for a one-shot map. A streaming map is a later Update system that still reads the same seed.
[dependencies]
bevy = "0.20"
rand = "0.9"

The seed is a resource. The walk is plain data. Entities are spawned only after the seed exists.

use rand::rngs::StdRng;
use rand::{Rng, SeedableRng};

#[derive(Resource)]
struct RunSeed(u64);

#[derive(Component, Clone, Copy)]
struct FloorCell {
    x: i32,
    y: i32,
}

fn insert_seed(mut commands: Commands) {
    commands.insert_resource(RunSeed(42));
}

fn spawn_walk(mut commands: Commands, seed: Res<RunSeed>) {
    let mut rng = StdRng::seed_from_u64(seed.0);
    let mut x = 0;
    let mut y = 0;
    for _ in 0..24 {
        commands.spawn((
            FloorCell { x, y },
            Transform::from_xyz(x as f32, y as f32, 0.0),
        ));
        match rng.random_range(0..4) {
            0 => x += 1,
            1 => x -= 1,
            2 => y += 1,
            _ => y -= 1,
        }
    }
}

pub struct DungeonPlugin;

impl Plugin for DungeonPlugin {
    fn build(&self, app: &mut App) {
        app.add_systems(Startup, (insert_seed, spawn_walk).chain());
    }
}

Height samples stay in a resource. The same seed always rebuilds the same grid. Sampling happens once, not every frame.

#[derive(Resource)]
struct HeightGrid {
    width: i32,
    samples: Vec<f32>,
}

#[derive(Resource)]
struct RunSeed(u64);

fn hash_unit(x: i32, y: i32, seed: u64) -> f32 {
    let mut n = seed
        .wrapping_mul(0x9E37_79B9_7F4A_7C15)
        ^ (x as u64).wrapping_mul(0xBF58_476D_1CE4_E5B9)
        ^ (y as u64).wrapping_mul(0x94D0_49BB_1331_11EB);
    n = (n ^ (n >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
    n = (n ^ (n >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
    n ^= n >> 31;
    (n >> 40) as f32 / (1u32 << 24) as f32
}

fn build_heights(mut commands: Commands, seed: Res<RunSeed>) {
    let width = 16;
    let mut samples = Vec::with_capacity((width * width) as usize);
    for y in 0..width {
        for x in 0..width {
            samples.push(hash_unit(x, y, seed.0));
        }
    }
    commands.insert_resource(HeightGrid { width, samples });
}

Rooms are rectangles carved into a grid you own. Spawn markers only for floor cells.

struct Room {
    x: i32,
    y: i32,
    w: i32,
    h: i32,
}

fn carve(width: i32, rooms: &[Room]) -> Vec<bool> {
    let mut floor = vec![false; (width * width) as usize];
    for room in rooms {
        for y in room.y..room.y + room.h {
            for x in room.x..room.x + room.w {
                if (0..width).contains(&x) && (0..width).contains(&y) {
                    floor[(y * width + x) as usize] = true;
                }
            }
        }
    }
    floor
}

#[derive(Component, Clone, Copy)]
struct FloorCell {
    x: i32,
    y: i32,
}

fn spawn_rooms(mut commands: Commands) {
    let rooms = [
        Room { x: 1, y: 1, w: 4, h: 3 },
        Room { x: 8, y: 6, w: 5, h: 4 },
    ];
    let width = 16;
    for (i, open) in carve(width, &rooms).into_iter().enumerate() {
        if open {
            let x = (i as i32) % width;
            let y = (i as i32) / width;
            commands.spawn(FloorCell { x, y });
        }
    }
}

Choose the owner

| Need | Owner |

| --- | --- |

| Seeded run | A Resource holding a seed you define |

| Chunk or dungeon grid | Your components, spawned with Commands |

| Hash, rooms, or walk | Your functions over a RunSeed. rand::rngs::StdRng is the seeded generator. |

| Floor cells | Entities spawned from the grid. Tilemaps belong on the tile skill. |

| Nav bake after generation | Nav bake after generation belongs on bevy-navigation. |

Scope

  • Seeded walks and room carves above use rand 0.9 and your own grid. Bevy does not ship a dungeon solver.
  • Build the grid on the main schedule, then spawn. Commands stay on the system that runs there.
  • Navigation rebake after a generated map belongs on bevy-navigation.

Expert Knowledge Delta

Target: Bevy 0.20+.

Core ECS types: Component, Resource, Query, Commands, Startup, Update, Fixed / FixedMain (bevy::time), .chain(), Plugin, DefaultPlugins, MinimalPlugins. Messages and observers are an ownership slot. See bevy-events-messaging for their APIs. Systems run in parallel unless .chain() orders the ones that conflict.

Who owns state

| Piece | Owner | Who may mutate | Schedule |

| --- | --- | --- | --- |

| Data on one entity | Component | One system via Query | Update or FixedMain |

| One value for the whole app | Resource | One system | Startup to insert, then the schedule that reads it |

| Spawn | Commands | The spawning system | Startup for the initial world; later schedules for runtime spawn |

| Order that must not race | .chain() | The chained systems only | Same schedule |

| Plugin boundary | Plugin | build registers systems and resources | Before the app runs |

Borrow rule: two systems that both mutably query the same component must be ordered (.chain()) or split so only one writes. Unrelated systems stay parallel. That is the ECS page rule, not a guess at a system-set API.

Decision

| If you need | Do this | Do not |

| --- | --- | --- |

| Per-entity data | Component + Query | A value that is not a component |

| App-wide config or clock | Resource | Copy the value onto every entity |

| Stable simulation step | FixedMain and the Fixed clock | Tie simulation to raw frame Update |

| Startup spawn | Commands in Startup | Spawn from a render system |

| Cross-system notice that is not shared state | Message or observer slot in bevy-events-messaging | A new event type invented in this file |

| An API outside the documented set | Check the official docs first | A made-up crate or API |

Load references/ecs-ownership.md for the ownership table.

NEVER

  • Do not pin /0.20.0/ in docs URLs. WHY: links stay on latest; the 0.20 target is stated once in prose.
  • Do not mutably query the same component from two unordered systems. WHY: the schedule runs systems in parallel unless .chain() orders them.
  • Do not invent a physics, navigation, networking, or XR crate here. WHY: this skill only names documented Bevy crates and modules.
  • No Godot noise, tile, or thread-pool types. WHY: Bevy code stays idiomatic ECS and uses only documented Bevy APIs.
  • No invented noise, wave-function, or chunk APIs. WHY: Bevy code stays idiomatic ECS and uses only documented Bevy APIs.
  • No named third-party generation crate as the chosen tool. WHY: Bevy code stays idiomatic ECS and uses only documented Bevy APIs.

Reference

The library pin on bevy-master stays 0.20.0.

Official Documentation

  • Assets - Official asset catalog.
  • Examples - Example tree on the latest branch.
  • ECS - Component, Commands, Startup, and .chain().
  • Resources - App-wide seed state. API limits are on bevy-data-resources.
Prerequisites
Complements
Downstream / consumers
  • None yet.
Master

How to use it

Copy the folder

Take thedivergentai/bevy-procedural-generation from the repository into ~/.claude/skills for personal use, or into .claude/skills inside a project.

Check the name does not clash

The agent identifies a skill by the name field in its header. Two skills with the same name cannot sit side by side — one of them will be ignored.