The premise: a cat knocks over a shelf of board games. You shrink down and enter the fallen game boxes. Each box imposes different rules: poker in one, dice in another, tile puzzles in a third. You carry mechanics forward between boxes, creating hybrid gameplay.
The name is Patchwork Playroom. It has seven finalized PRD documents, a design philosophy manifesto, a six-phase roadmap, and zero lines of production code.
Design-First, Code-Never (So Far)
The .planning/ directory has more content than the source directory. By a lot:
.planning/
├── BRIEF.md
├── ROADMAP.md
├── best-practices.md
├── prd/
│ ├── PRD-01-primitives.md
│ ├── PRD-02-scoring.md
│ ├── PRD-03-turn-structure.md
│ ├── PRD-04-boxes-foundation.md
│ ├── PRD-05-boxes-advanced.md
│ ├── PRD-06-tokens.md
│ └── PRD-07-verbs.md
├── phases/
│ ├── 01-core-systems/ (COMPLETE)
│ └── 02-game-boxes/ (COMPLETE)
└── research/
Seven PRDs. Two completed design phases. All before a prototype exists. This is deliberate, and here's why:
Testing 10 Designs in the Time It Takes to Code 1
Game design with AI:
"What if the pressure mechanic resets between levels instead of accumulating?" → Claude reasons about downstream effects in 30 seconds: softer difficulty curve, less risk/reward tension, players can play safe. Game loses its push-your-luck identity. Rejected.
"What if players carry 3 active rules instead of 2 in Act 2?" → Claude evaluates the combinatorial space: 6 foundation boxes produce 20 three-box combinations vs. 15 two-box. More content, but cognitive load spikes. Needs playtesting for the threshold.
"What breaks if tiles have rotation as a verb?" → Claude traces through the system: rotation creates 4x more tile states, complicates the UI, but adds depth to the Placement box specifically. Worth it for one box, not as a universal verb.
Each of these questions would have taken hours to test in code: build the mechanic, implement the variation, playtest, observe, record. With design documents and Claude's ability to reason about interconnected systems, we evaluated 10 variations in the time it took to document the winning design.
The cheapest code is the code you don't write because you already proved the design was wrong.
The Insight Claude Surfaced
Five design principles emerged from this process. Four of them I would have reached eventually: depth over breadth, player perception matters more than math balance, push to breaking then pull back 20%, use familiar objects to reduce cognitive load.
The fifth one I wouldn't have found: "UI is gameplay." Claude pointed out that in most games, the UI is a window into the game state. But in a deckbuilder where you're combining mechanics from different boxes, the act of looking at your tokens and considering combinations is the game. The interface isn't showing you the game. The interface is the game.
I wouldn't have articulated that on my own. Claude surfaced it. I recognized it as right. That's the collaboration model, and it works the same way outside games. Pricing tiers, permission systems, workflow builders, anything with interdependent constraints. AI can reason about the downstream effects of a design change faster than you can build a prototype to test it.
The Takeaway
The most valuable output of an AI design session isn't the final document. It's the 15 variations you explored and rejected on the way there.
Whether Patchwork Playroom becomes a real game depends on whether I ever stop starting new projects long enough to finish one. The odds are not in its favor.