Skip to content

gamedev-skills/awesome-gamedev-agent-skills / skills/genres/roguelike

roguelike

Verified

Designs run-based roguelikes and roguelites in any engine: run structure, permadeath, meta-progression, seeded and daily runs, turn order and symmetric field of view. Use for NetHack- or Hades-style games, or runs that feel cheap. Pair with `procedural-gen` for dungeon generation and `save-systems` for run vs meta saves. Not for roguelike deckbuilders (`card-game`).

Roguelike

Owns the run: what a run contains, what survives death, how turns and seeds work, and what a generated floor must guarantee. The generators come from procedural-gen; this skill decides what they must produce and checks it.

When to use

  • A game where each run starts fresh, death ends it, and content is regenerated: a traditional turn-based grid roguelike or a real-time roguelite.
  • Deciding permadeath versus meta-progression, run length and act structure, seeds and daily runs, turn scheduling, or FOV.

When not to use: the algorithm that carves rooms or caves, noise, WFC or loot-table rolling (procedural-gen); deckbuilders such as Slay the Spire or Balatro, even roguelike ones (card-game); persistent-character RPGs (rpg). For a real-time roguelite also load the action genre skill (platformer, fps-shooter) for moment-to-moment control.

Workflow

  1. Settle the type with the user. Traditional (turn-based, grid, strict permadeath, runs of hours, knowledge is the progression) or roguelite (real-time or turn-based, runs of 20-60 minutes, some power or options persist). The Berlin Interpretation (RogueBasin, 2008) lists the traditional traits; use it as a vocabulary, not a checklist.
  2. Lay out the run. 3-4 acts or biomes, each ending in a boss or gate, with a choice between paths or rewards at least every few minutes. Plan a first death around 5-15 minutes in for a new player, and make the death screen show the seed, depth, cause and what was unlocked.
  3. Split state into run state (wiped on death) and profile state (kept). Decide strict permadeath (delete the run save on load) or checkpointed runs, and write the rule down.
  4. Seed everything from one run seed with per-floor, per-purpose sub-seeds (refs §2).
  5. Turn order (turn-based games): energy scheduler (Pattern 1) with a few discrete speeds.
  6. Generate floors with procedural-gen, then gate each floor with the acceptance checks (refs §1) and the spawn budget (refs §4).
  7. Check it. Write and run these as tests in the project's runner and report the numbers:
    • same seed twice gives the same hash for floors 1 to N, even after playing floor 1 differently between the two runs;
    • 500+ seeds per depth pass the acceptance checks, with the reject rate printed;
    • FOV symmetry on random maps (refs §3);
    • speeds 50/100/200 over 1,000 time units give turns in a 1:2:4 ratio;
    • dying deletes or invalidates the run save and leaves the profile file unchanged.

Design rules and numbers

DecisionStart hereWhy
Speeds50, 100, 200 (slow, normal, fast)odd ratios such as 110 give an unexplained extra enemy move every tenth turn
Action cost100 energy for move, attack and item usechange cost per action (heavy attack 150) rather than adding speeds
FOV radius6-9 tiles; lit rooms fully visiblesmall radius adds tension and cuts tactics
Out-of-depth spawnabout 5% of spawns, up to 3 levels deeperrare spikes stay memorable
Safe startno monster within 6-8 tiles of the arrival pointthe first turn must not decide the run
Meta-progressionnew options (characters, items in the pool, starting choices) before flat stat boostsflat boosts make early acts trivial and turn runs into a grind
After the first wina stacking difficulty ladder (Slay the Spire's Ascension, Hades' Heat)gives experts a reason to keep playing
Daily runone seed per UTC day, one attempt, shared leaderboardcomparable runs need identical seeds

Pattern 1: energy scheduler

# Pseudocode. Actors gain energy each time unit and act when they have enough.
TURN_COST = 100
def next_actor(actors):
    while True:
        for a in actors:                 # 1) serve anyone already ready, in a stable order
            if a.energy >= TURN_COST:
                a.energy -= TURN_COST    #    or the chosen action's cost
                return a
        for a in actors:                 # 2) nobody ready: advance time for everyone at once
            a.energy += a.speed          #    100 = normal, 200 = hasted, 50 = slowed

Grant energy in its own pass. Adding it inside the loop that returns the first ready actor starves everyone after it: with two speed-100 actors the first acts every time.

Gotchas

  • Same seed, different floor 3 → one RNG shared by generation and combat, so the number of attacks on floor 1 shifts later floors. Derive a sub-seed per floor and per purpose (refs §2).
  • Seeds differ between runs of the same build → seeding from a language hash() of a string (Python randomises it per process) or iterating an unordered set while generating. Hash with a fixed algorithm and iterate in sorted order.
  • One monster never acts → energy granted inside the "find who is ready" loop (Pattern 1).
  • Double moves or input races → a real-time loop pretending to be turn-based. Resolve one discrete player action, then let the scheduler run until the player is ready again.
  • You see the monster, it does not see you → classic or ray-cast FOV. Use symmetric shadowcasting (python-tcod: compute_fov(..., algorithm=tcod.constants.FOV_SYMMETRIC_SHADOWCAST)).
  • Stairs behind an unreachable wall, or the key locked behind its own door → generator output accepted without checks (refs §1).
  • Deaths feel unfair → damage the player could not see coming. Show enemy intent or telegraph big hits, and log every damage source in the message log.
  • Permadeath wipes unlocks → run and profile in one save. Separate them (refs §6).
  • Save-scumming → the run save survives a load. Delete or invalidate it on load if strict permadeath is the design; copying save files cannot be prevented, so do not try.
  • Identification knowledge carries between runs → fixed appearance names. Shuffle them per run (refs §5).

References

  • references/run-seeds-fov.md: read when writing floor acceptance checks, sub-seed and daily-seed code, symmetric shadowcasting, spawn budgets, per-run identification, or the run and profile save split.

Related: procedural-gen (generators, noise, loot tables), save-systems (atomic writes, versioning), game-ai (monster behaviour), godot-tilemap or unity-tilemap-2d (grid rendering), game-feel (impact feedback).

Version notes

Last verified: 2026-11-11.

brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall