Interactive session to design and implement a new GridPonder game pack from scratch. Proceeds in three phases: discovery questions, design approval, then full implementation (entity kinds, systems, rules, tiles, solver, levels). Works for concepts fitting the GridPonder DSL — deterministic, turn-based, single-player, perfect-information, 2D grid puzzle. Invoke with a concept ("create a sokoban-style game") or blank ("create game").
Create a new GridPonder level from scratch. Works for any game pack. Reads the game definition to understand entity kinds and mechanics, designs a level meeting the stated constraints, writes the JSON file, registers it in game.json, and runs the integration test to verify. Use a solver if one exists for the game. Arguments: <pack-id> <new-level-id> [design notes] e.g. "rotate_flip rf_004 4x4 grid, 8 moves, use all 4 colours".
Analyse, redesign, or complete a GridPonder level. Reads the level JSON, evaluates design quality (aha moment, difficulty, gold path validity, cheap alternatives), then proposes and writes a revised version. Use when a level is incomplete, too easy, too hard, lacks an aha moment, or has mechanical errors. Arguments: <level-id> [notes]
Run an integration test for a specific GridPonder level. Navigates to the level, executes gold-path moves (all or N), takes a screenshot after every move, then analyses the sequence. Use when debugging a level's visual layout, animations, equipment mechanics, or win conditions. Arguments: <level-id> [num-moves] e.g. "fw_001 3" or "fw_005" (all moves).
Generate game tile or sprite images from a natural language description using local AI (SDXL-Turbo + pixel-art LoRA). Runs entirely on-device via MPS (Apple Silicon). Single tiles use text2img; animation sequences use img2img conditioning so all frames share the same character and palette. Arguments: <prompt> [--size N] [--count N] [--style STYLE] [--strength F] [--output DIR] [--name NAME]