Iterplay
Iterplay turns a design doc into a browser game (for now, a single-level 2D platformer), then improves it automatically.
Core trick
Gemini writes a browser platformer, bots record playtests, a critic reviews footage against the design doc, and a deterministic verifier checks claims against telemetry.
Description
Iterplay turns a design doc into a browser game (for now, a single-level 2D platformer), then improves it automatically. Gemini writes the game, bots play it while being recorded, Gemini critiques the footage against the design doc, a deterministic verifier checks every claim against telemetry, and the fixer patches the code. The loop repeats until the game meets its rubric.
README excerpt
# iterplay
Iterplay turns a design doc into a browser game (for now, a single-level 2D platformer), then improves it automatically. Gemini writes the game, bots play it while being recorded, Gemini critiques the footage against the design doc, a deterministic verifier checks every claim against telemetry, and the fixer patches the code. The loop repeats until the game meets its rubric.
Design: [docs/iterplay_high_level_design.md](docs/iterplay_high_level_design.md). Code comments cite it as "HLD", e.g. `HLD §4 [6]`.
> **Status: working MVP.** The whole loop of HLD §2 runs end to end on the 2D platformer template: the Builder, the smoke test with quick Fixer retries, three recorded bot playtests, media prep, the critic, the Verifier, the Scorer with keep-best and reverts, and the Fixer. Every artifact in HLD §6 is written as it is produced, and every model call is cached in its version's `llm/`. `play`, `replay` and `view` work on finished runs. Not built yet: the Managed Agents critic (`--critic managed` stops with a config error), a CLI switch for the direct critic's persona mode, 3D templates, and the stretch goals in HLD §13.
## Layout
```
src/iterplay/
contract/ §3 game contract: getState() schema, game templates, controls, URL params,
allowed files
models/ §4 §6 pydantic models for every artifact passed between stages or saved to runs/
agents/ §4 §5 Builder [1], critic backends [5], Fixer [8], guardrails, prompt