# Plyquarry

> A rules sandbox for inventing two-player abstract strategy games (perfect information, no luck). Build a ruleset from
> controls, play it on the board, auto-play it engine against engine, then ask a model to critique the balance with
> concrete rule changes, or to write a printable rules card. https://plyquarry.skillsafe.ai/

Built on the skill @jwynia/abstract-strategy (https://skillsafe.ai/skill/@jwynia/abstract-strategy) by jwynia:
designing abstract strategy games with perfect information, no randomness and strategic depth, and evaluating balance.

## Who it is for

Hobby game designers, teachers inventing classroom games, families making up a game for the holidays.

## What runs free in the browser

- Rule controls: square or hex (rhombus) board 4x4 to 10x10; up to three piece types drawn on a 5x5 move stencil
  (inner ring = step/slide directions with a range, outer ring = leaps over anything); setup (empty, home rows,
  alternate cells); phase (move only, place only, place then move, with a reserve); capture (replacement, jump-over
  chains of up to 8 jumps, custodial trapping, forced capture); win conditions (connect N in a row, reach the far row, capture all,
  majority at a full board, immobilise); actions per turn, first-turn actions, extra reserve for the second player;
  a ply limit ending in a draw or a majority count. Lint warnings for impossible or dead rules.
- A live board: legal moves shown when a piece is selected, hand play for both sides with undo, or play against a
  random, greedy (one ply) or minimax (two ply, beam-limited) engine.
- Auto-play: seeded, bounded (400 games, 60 with minimax, 400 plies each), run in short slices with progress. Reports
  first-player (ink) win rate, second-player win rate, draw rate, the first-player edge, average and median length,
  a length histogram, which win condition decided each game, captures per game and moves per position.
  After a rule change, each headline number shows what it was on the previous auto-play.
- Six presets rebuilt from classic public-domain abstracts and named generically: Five Stones, Far Row Race,
  Diagonal Jumpers, Sandwich, Place and Shift, Hex Lines. Three carry free example replies for both asks.
- Exports: ruleset JSON (opens again), rules card as Markdown and a print stylesheet, simulation CSV, board SVG,
  and a link that carries the ruleset in its #fragment so anyone who opens it gets the same rules on the board.
- History: one record per ruleset with its last simulation, in the signed-in user's account (collection
  `rulesets`), mirrored in the browser.

## The two metered asks (model gpt-terra)

Input (all strings): `task` (`critique` | `sheet`), `ruleset` (JSON), `summary` (JSON simulation summary; an empty string on a `sheet` ask made before any auto-play), `lint`,
`fields` (the only fields a change may target, with legal values), `previous` (earlier summary and accepted changes,
for a re-critique), `designer_note`, `clipped`. The model never sees moves or game trees.

- `critique` returns a verdict (promising, lopsided, drawish, degenerate, too-short, too-long, broken), a headline,
  findings each tied to a summary statistic with its exact value, rule changes as field -> value with a one-line
  reason, what to watch after re-simulating, and on a re-critique what moved since the last round.
- `sheet` returns a rules card: components, setup, turn, each piece's moves, capture, winning (one entry per enabled
  win condition), ending, tips and a stats line.

The browser reconciles every reply: cited rates must match the summary within 0.03, lengths within 5%, every change
must name a catalogue field with a legal, different value, and a rules card must cover exactly the ruleset's pieces
and win conditions. Accepted changes apply to the controls; one click re-simulates; the next critique compares.

## Links

- App: https://plyquarry.skillsafe.ai/
- API tutorial: https://plyquarry.skillsafe.ai/api.html