gamedev-skills/awesome-gamedev-agent-skills已通過檢查
SKILL DETAIL
roguelike
gamedev-skills/awesome-gamedev-agent-skills/roguelike
Build a roguelike: turn-based grid movement, procedural dungeons, permadeath, field-of-view, and loot tables. Use for a roguelike/roguelite or turn-based grid dungeon crawler with procedural levels.
安裝量 · 230查看來源
Installation
npx skills add https://github.com/gamedev-skills/awesome-gamedev-agent-skills --skill roguelike
技能檔案
SKILL.md
最近同步 · 2026年10月24日
references/generation-fov-loot.md›
# Roguelike generation, FOV, and loot (depth)
Algorithms behind the systems in `SKILL.md`. Engine-neutral pseudocode. For the noise/RNG
primitives and seeding, defer to `procedural-gen`; this file is the roguelike-specific glue.
## 1. Dungeon generation approaches
Pick by the *texture* you want. All must end with a **connectivity pass** (§2).
### Rooms-and-corridors (classic, readable)
```python
# Place non-overlapping rooms, then connect their centers with L-shaped tunnels.
rooms = []
for _ in range(MAX_ROOMS):
r = random_rect(rng, min=4, max=10)
if any(r.intersects(other.expand(1)) for other in rooms):
continue # reject overlaps (expand by 1 to keep walls between)
carve_room(map, r)
if rooms: # connect to the previous room
carve_h_then_v(map, rooms[-1].center, r.center, rng)
rooms.append(r)
# First room = player spawn; last/farthest room = stairs down.
```
### BSP (binary space partition — tidy, building-like)
Recursively split the map rectangle into sub-rectangles; place one room per leaf; connect
sibling rooms as you unwind the recursion. Produces evenly distributed, non-overlapping rooms.
### Cellular automata (organic caves)
```python
# Random fill, then smooth: a cell becomes wall if most neighbors are walls.
fill_random(map, wall_chance=0.45, rng=rng)
for _ in range(4, 6): # 4–6 smoothing passes
for cell in map:
walls = count_wall_neighbors(map, cell) # 8-neighborhood
map[cell] = WALL if walls >= 5 else FLOOR
# Then keep only the largest connected floor region (discard isolated pockets).
```
### Drunkard's walk (cheap, winding)
Start at a point and random-walk, carving floor, until a target floor percentage is reached.
Guaranteed connected (one continuous path) but can be too snaky without tuning.
## 2. Connectivity (never skip this)
```python
# Flood-fill from the player's start; any walkable cell not reached must be connected or removed.
reached = flood_fill(map, player_start)
regions = find_disconnected_floor_regions(map, reached)
for region in regions:
carve_corridor(map, nearest_cell(reached), nearest_cell(region), rng)
# Re-flood and assert: every floor cell is now reachable.
```
For caves, the common shortcut is to keep only the single largest region and discard the rest.
## 3. Field of view — symmetric shadowcasting
Properties you want: **symmetric** (if A sees B, B sees A), no blind spots, no flicker as the
player moves. Recursive shadowcasting scans the eight octants around the origin, tracking a
"shadow" of slopes blocked by walls:
```
for each of 8 octants:
scan rows outward from the origin up to radius R
track the visible slope range [start_slope, end_slope]
when a wall is hit, recurse into the narrower slope range beyond it,
and continue the current row with the reduced range
mark a cell visible if it lies within the current slope range and within R
```
Implementation notes:
- Transform each octant into a common coordinate frame so one routine handles all eight.
- Use a circular radius check (`dx*dx + dy*dy <= R*R`) for a round light, not a square.
- Cache `explored` separately; render explored-but-not-visible cells dimmed.
A symmetric variant ("symmetric shadowcasting") is simpler to reason about and avoids
single-cell asymmetries; prefer it if you implement from scratch.
## 4. Weighted, depth-scaled loot and spawn tables
```python
# A drop table is a list of (entry, weight). Weight is relative, not a probability.
goblin_loot = [
("nothing", 60),
("gold_small", 25),
("healing_potion", 10),
("rare_scroll", 5),
]
def roll(table, rng):
total = sum(w for _, w in table)
pick = rng.range(0, total) # [0, total)
upto = 0
for entry, w in table:
upto += w
if pick < upto:
return entry
# Depth scaling: shift weights toward stronger entries as depth increases,
# e.g. add depth-gated entries or multiply rare weights by a depth factor.
```
Design guidance:
- Keep "nothing"/common as the bulk of weight so rares stay exciting.
- Gate powerful items behind a minimum depth so early runs can't snowball.
- Separate **monster** spawn tables from **item** drop tables; both scale with depth.
- For "exactly one of" results use weighted pick; for "each may drop" iterate independent rolls.
## 5. Run vs. profile state (permadeath + meta-progression)
- **Run state** (current dungeon, HP, inventory, seed): lives only during the run. For *true*
permadeath, delete or invalidate this save the moment it is loaded so it cannot be reloaded.
- **Profile state** (unlocked items/classes, high scores, achievements, currency): persists
across deaths. This is what a roguelite spends on meta-progression.
- Keep them in separate files/sections so wiping a run never touches the profile (see
`save-systems` for slot/versioning details).
SKILL.md›
---
name: roguelike
description: >
Build a roguelike: turn-based grid movement, procedural dungeons, permadeath, field-of-view,
and loot tables. Use for a roguelike/roguelite or turn-based grid dungeon crawler with procedural levels.
---
# Roguelike
A playbook for roguelikes — the turn engine, procedural dungeons, field-of-view, permadeath,
and run economy. This is a **compositional** skill: it orchestrates procedural generation,
tilemaps, save handling, and AI into a run-based game. It does not re-teach noise/RNG or
tilemap APIs; it defines the loop and the systems that make a *run* compelling.
## When to use
- Use when building a **turn-based, grid-based** dungeon crawler where each death ends the
run and the world is regenerated — a roguelike or "roguelite" (with meta-progression).
- Use when designing procedural dungeons, FOV/fog-of-war, permadeath stakes, or loot tables.
**When *not* to use:** real-time action with roguelike *dressing* → build the action genre
(`platformer`/`fps-shooter`) and layer `procedural-gen`. Deep stats/quests/dialogue without
permadeath → `rpg`. Open-world needs/crafting → `survival-crafting`.
## What makes it a roguelike (design anchor)
The community reference is the **Berlin Interpretation** (RogueBasin, IRDC 2008): a set of
"roguelikeness" factors, not a checklist. The high-value ones are the design targets:
**random environment generation, permadeath, turn-based, grid-based, non-modal** (all actions
in one mode), **complexity** (many item/monster interactions), **resource management**,
**hack-and-slash**, and **exploration & discovery**. Lean into these to feel roguelike; pick
which to keep deliberately. "Roguelite" usually relaxes permadeath with meta-progression.
## Core loop
**Take a turn (move / fight / use) → the world resolves its turn → see new state → descend /
loot / survive → die and restart with a fresh dungeon.** Replayability comes from the *world*
changing each run, not the player memorizing a fixed layout.
## Must-have systems
1. **Turn scheduler** — energy/initiative system so fast actors act more often (Pattern 2).
2. **Grid map + movement** — tile coordinates; bump-to-attack; blocked/walkable queries.
3. **Procedural dungeon generator** — rooms + corridors, or BSP/cellular; guarantee connectivity.
4. **Field of view + explored memory** — what's visible now vs. seen-before (Pattern 3 / refs).
5. **Combat + entities** — HP, attack/defense, status; monsters obey the same rules as the player.
6. **Loot + drop tables** — weighted, depth-scaled item/monster spawning (refs).
7. **Permadeath + (optional) meta-progression** — wipe the run; persist only unlocks/score.
8. **Message log + clear UI** — the player reasons from text/state; surface numbers and events.
## Design knobs
| Knob | Effect | Notes |
|------|--------|-------|
| Dungeon size / room count | run length, density | Scale with depth. |
| Connectivity guarantee | no unreachable rooms | Always verify reachability after generation. |
| Monster density / depth curve | difficulty ramp | Spawn by depth-weighted table. |
| Loot rarity weights | power variance | Rarer = bigger swing; identify adds discovery. |
| FOV radius / lighting | tension, information | Smaller radius = scarier, slower. |
| Resource scarcity (food/HP/ammo) | pressure to descend | Core tension lever in classic RLs. |
| Permadeath vs meta-progression | run stakes vs. retention | Roguelite softens the wall. |
| Identification / unknowns | exploration value | Unidentified items reward experimentation. |
| Seedable RNG | daily runs, debugging | Always allow a fixed seed (see `procedural-gen`). |
## Patterns
### 1. Deterministic, seedable run RNG
```python
# Pseudocode. One seeded RNG per run makes dungeons reproducible (daily runs, bug repro).
run_seed = chosen_seed or random_seed()
rng = Rng(run_seed) # use your engine's seedable RNG, not global random
dungeon = generate_dungeon(rng, depth) # same seed + depth => same dungeon
# Persist run_seed in the save so a crash can resume the same world (see save-systems).
```
### 2. Energy-based turn scheduler (speeds differ)
```python
# Pseudocode. Each actor gains energy each tick and acts when it has enough.
# Faster actors gain more per tick, so they act more often — no fixed "player then enemies".
TURN_COST = 100
def next_actor(actors):
while True:
for a in actors: # stable order avoids ties favoring one side
a.energy += a.speed # e.g. speed 100 = normal, 150 = hasted
if a.energy >= TURN_COST:
a.energy -= TURN_COST
return a # this actor takes exactly one action now
```
### 3. Field of view + explored memory
```python
# Pseudocode. Recompute visibility from the player each time they move.
visible = compute_fov(map, player.pos, radius=8) # symmetric shadowcasting (see refs)
for cell in visible:
explored.add(cell) # remember it forever (dim "fog of war")
# Render: visible -> lit; explored-but-not-visible -> dim; never-seen -> hidden.
```
Use a proven FOV algorithm (recursive shadowcasting or symmetric shadowcasting). Do not
roll a naive raycast-per-cell — it produces asymmetric, "blinking" vision. See refs.
## Pitfalls / failure modes
- **Disconnected dungeons** → rooms the player can't reach. Always run a connectivity/flood-fill
pass and carve corridors until every walkable cell is reachable.
- **Naive FOV** → vision that flickers or is asymmetric (you see them, they don't see you).
Use shadowcasting; test symmetry.
- **Real-time loop pretending to be turn-based** → input races and double-moves. Resolve one
discrete turn at a time; queue input.
- **Flat difficulty** → no descent pressure. Scale monsters/loot by depth and keep resources scarce.
- **Permadeath that wipes meta-unlocks** → frustration. Separate *run* state (wiped) from
*profile* state (unlocks, scores) when saving (see `save-systems`).
- **Save-scumming a "permadeath" game** → delete or invalidate the run save on load if you want
true permadeath; keep only the profile.
- **Unreadable state** → the player can't plan. Show HP, turn results, and a message log.
## Composition (build it from these skills)
- **Generation:** `procedural-gen` (noise, seeded RNG, dungeon/room algorithms) — the engine of replayability.
- **Map rendering:** `godot-tilemap` / `unity-tilemap-2d` for the grid; `level-design` for set-piece rooms/vaults.
- **Enemies:** `game-ai` for monster decision-making (often simple on a grid: seek/flee/patrol).
- **Persistence:** `save-systems` for run-resume, profile/meta-progression, and true permadeath wipes.
- **Scripting/data:** `godot-resources` / `unity-scriptableobjects` to define items, monsters, and drop tables as data.
- **UI:** `godot-ui-control` for the message log, inventory, and HUD.
- **Feel:** `game-feel` for hit/death juice — screen shake and hit-stop that sell impacts on a turn-based grid.
## References
- For dungeon-generation algorithms (rooms-and-corridors, BSP, cellular automata, connectivity),
FOV (shadowcasting), and weighted loot/spawn tables, read `references/generation-fov-loot.md`.