Triple Triad — a match's flow

The card minigame from Final Fantasy VIII — a project started years before generative AI came along, rebuilt from scratch. The opponent is a dedicated AI that knows the game's rules, not a generalist model and not Claude. This is a first sketch, not the final design: capture rules (Same, Plus, elemental...) aren't mapped out yet.

At a glance

Type
The card minigame from Final Fantasy VIII, rebuilt from scratch
Opponent
A dedicated AI that knows this game's rules — not a generalist model, and not Claude
Origin
Started years before generative AI came along
Status
First sketch — the capture rules aren't mapped out yet

How a match runs

  1. New match — the board starts empty
  2. Player turn — the player places a card on the board
  3. Resolve captures — the ranks on the touching sides are compared, and the cards that lose flip over
  4. Board full? — if it isn't, play passes to the opponent; if it is, the match ends
  5. Opponent AI turn — the dedicated AI places a card and its captures are resolved the same way, then it's the player's turn again
  6. Tally and winner — the cards are counted and the winner is declared

Decisions and constraints

  • The opponent is a dedicated rules AI — a small AI that knows this game, not a generalist model and not Claude [1]
  • The project predates generative AI — it was started years earlier and is being rebuilt from scratch, not patched onto the old one
  • One capture resolution for both sides — the player's move and the AI's move go through the same step — one rule set, not one per side
  • The capture rules are still open — Same, Plus, elemental and the rest aren't mapped out yet. This is a first sketch, not the final design

References

  1. Karpathy, A. — ReinforceJS: WaterWorld (interactive demo). Stanford University. cs.stanford.edu/people/karpathy/reinforcejs/waterworld.html The demo the idea started from: an agent that learns by reinforcement to play, not a language model.