Skip to content

gamedev-skills/awesome-gamedev-agent-skills / skills/genres/card-game

card-game

Verified

Designs card games and deckbuilders in any engine: cards as data with one effect interpreter, deck zones, a trigger stack and seeded shuffles. Use for a Slay the Spire- or Balatro-like or a TCG, or cards that duplicate or streaky shuffles. Pair with the engine skill for card data and hand UI (`godot-resources`, `godot-ui-control`, `unity-scriptableobjects`). Not for grid crawlers (`roguelike`).

Card game

Owns the rules engine of a card game: zones, draw, costs, effect resolution, rewards and balance math, including the run structure of a roguelike deckbuilder. Engine skills supply the data assets and UI.

When to use

  • The core objects are cards moving between zones: deckbuilder, roguelike deckbuilder, TCG/CCG, solitaire, poker-hand scorer, card battler.
  • Designing costs, draw, rewards, triggers, enemy intents or a scoring formula, or fixing cards that duplicate, vanish or resolve in the wrong order.

When not to use: grid dungeon runs (roguelike); board games with matching or pushing rules (puzzle); an RPG with a small card minigame (start in rpg).

Workflow

  1. Name the reference game and fix the frame. Write the numbers down before any card: deck size, hand size, draw per turn, resource per turn, what happens to the hand at end of turn, copy limits. Reference frames are in the table below.
  2. Model zones and cards as data. Every card instance has a unique id and lives in exactly one zone. Card text maps to a list of ops for one interpreter (Pattern 1).
  3. Pick the resolution model: a stack (last in resolves first, for response-heavy TCGs) or a queue (first in first, simpler for PvE deckbuilders). Write the order for simultaneous triggers (active player first, then by timestamp).
  4. Write a power budget for one resource and price every card against it (below).
  5. Build the reward loop (deckbuilders): after each fight offer 3 cards, pick 1 or skip, with rarity weighted by act; give removal at shops or events.
  6. Check it. Write and run these as tests and report the results:
    • 10,000 random legal turns with a bot: after every action each card id appears in exactly one zone and the total count never changes;
    • shuffle uniformity: shuffle a 4-card deck 24,000 times; each of the 24 orders should land near 1,000 (a chi-square statistic above about 41 at 23 degrees of freedom means bias, p < 0.01);
    • same seed and same inputs replay to the same final state hash;
    • a trigger loop (A draws when B is drawn, B when A) stops at the depth cap with a log line;
    • a greedy bot plays the starter deck for 1,000 turns; print damage and block per turn and size the first enemies from it (about 3-4 turns of starter damage is a common start).

Reference frames

GameDeckHand and drawResourceEnd of turn
Slay the Spire10-card starter (Ironclad: 5 Strike, 4 Defend, Bash)draw 53 energydiscard hand
Balatro52 playing cardshand size 8, play up to 54 hands, 3 discards per blindredraw to 8
Hearthstone30 cardshand cap 10, extra draws are destroyedmana 1 to 10, +1 per turnkeep hand
Magic: the Gathering60+ constructed, 40+ limited, max 4 copies7-card opening hand, draw 1landsdiscard to 7

Balatro's run is 8 antes of three blinds (small 1x, big 1.5x, boss mostly 2x the ante's base score). Slay the Spire's enemies show their next action (intent), which is what makes a PvE card game a planning puzzle rather than a guess. Copy the idea, never the cards.

Power budget and deck math

  • Budget per resource. Pick a baseline and price everything against it. Slay the Spire's starters set 1 energy = 6 damage (Strike) or 5 block (Defend), and an upgrade adds about half again (Strike+ deals 9). Rares may exceed the budget; zero-cost cards need a drawback or a small effect.
  • Every added card dilutes the deck. The chance a specific single card is in a 5-card hand is 5/deck size: 50% in a 10-card deck, 33% at 15, 25% at 20. That is why skipping a reward or removing a Strike is often the strongest pick, and why removal belongs in the reward economy.
  • Copies in constructed. A 4-of in a 60-card deck is in the opening 7 about 40% of the time and in the first 10 cards about 53% (hypergeometric). Use the same formula to tune any draw consistency.
  • Cycle time is deck size / draw per turn: a 10-card deck drawing 5 sees every card every 2 turns. Draw and deck thinning shorten it; that is where infinite combos come from.
  • Scoring pipelines (Balatro-style): fix the order and write it down, for example base chips, then added chips, then added mult, then multiplying mult, applied left to right in slot order. Additive and multiplicative stages in the wrong order change scores by orders of magnitude.

Pattern 1: zones, draw and play

# Pseudocode. One zone per card instance; moving = remove from one list, append to another.
def draw(player, n):
    for _ in range(n):
        if not player.draw_pile:
            if not player.discard:
                return on_empty_draw(player)    # deck-out loss, fatigue, or simply stop
            player.draw_pile = player.discard; player.discard = []
            shuffle(player.draw_pile, player.rng)            # seeded Fisher-Yates (refs §6)
        card = player.draw_pile.pop()
        if len(player.hand) >= HAND_CAP:
            move(card, to=player.discard); log("burned", card)   # define overdraw explicitly
        else:
            player.hand.append(card)

def play(card, player, targets):
    if player.energy < cost(card) or not legal(card, targets):
        return False                                         # nothing changed yet
    player.energy -= cost(card)
    move(card, frm=player.hand, to=player.limbo)             # in play while resolving
    for op in card.ops:
        push_effect(op, Ctx(player, card, targets))
    resolve_stack()                                          # refs §2
    move(card, frm=player.limbo, to=destination(card))       # discard, exhaust, board
    return True

Holding the card in a "limbo" zone while it resolves keeps a draw effect from reshuffling the card that is being played back into the draw pile.

Gotchas

  • A card in two zones, or gone → move done as add-then-remove, or removing by value when two copies exist. Give every instance an id and assert the invariant after each action.
  • Draw effect shuffles the played card back in → the card went to discard before its effects resolved. Resolve from a limbo zone (Pattern 1).
  • Triggers loop forever → A triggers B triggers A. Cap resolution depth (for example 100 pushes per action), stop with a log line, and add a test for the known loop.
  • Order changes between runs → triggers collected from a hash map or unordered set. Sort by a fixed key (controller, then timestamp) before pushing.
  • Biased shuffle → sort with a random comparator, or swap with any index instead of 0..i. Use Fisher-Yates with an inclusive upper bound and check the engine's range convention (refs §6).
  • Replays and undo drift → shuffle uses the engine's global RNG, which particles and AI also consume. Give the deck its own seeded generator and save its state.
  • "The shuffle is streaky" → players misread true randomness. A shuffled deck is already a bag randomiser; resist hidden pity rules in deck draws, and add visible tools instead (scry, mulligan, retain).
  • Cancelled targeting leaves the board half-changed → paid before validating. Validate cost and targets first, then commit (Pattern 1).
  • Starter deck dies to the first fight or trivialises act 1 → enemy HP set by feel. Size it from the bot's measured damage per turn (Workflow step 6).

References

  • references/effect-resolution.md: read when writing the op interpreter, the stack or queue, triggers and keywords, targeting and fizzle rules, deckbuilder versus constructed structure, shuffle code and per-engine RNG calls, or cost curves.

Related: godot-resources or unity-scriptableobjects (card data assets), game-ui-ux with godot-ui-control (hand layout, drag and targeting UI), save-systems (run and collection saves), game-ai (opponent choice), game-feel (card hover, play and hit 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