autonomous-ai Create a game people can play
Read the workspace and project contract. Preserve working source,
approved components and the person's decisions. Use the coding agent to author original mechanics
in game/rules.mjs and their physical components/rulebook in game/project.json.
Build an end-to-end game: setup, meaningful choices, chance if any, visible consequences, scoring, ties and an ending. Any style of self-contained component face can be embedded as original or supplied SVG/PNG/JPEG/WebP artwork. Numeric settings are explicitly authored design parameters; there is no fixed genre enum or fake generation button. Do not make people write the JavaScript.
Run node tools/build.mjs, then play the actual game in the viewer. The person can edit cards,
quantities, values, color, artwork and rulebook; lock approved components; save source; undo edits;
and apply a new edition while retaining the old playable game and replay. Keep these capabilities
when adapting the renderer for the user's brief. For a different presentation, change the studio
source deliberately and recheck its exports and accessibility.
Run node tools/check.mjs --seeds 80. Its uniform-choice simulations measure only the supplied
seeds and policy. An optional authored bot receives observations and legal actions, not hidden
state. Compare distributions with clear policy names; never imply that automated play proves fun,
fairness, human strategy or commercial quality.
Play real decisions to an outcome, including losing paths if present. Test restart, saved replay,
a targeted revision and preserved approved material. Export with node tools/export.mjs delivery.
Open the offline HTML and print outputs independently. Check every physical design, dimensions,
copy count, text, actual-size calibration and rulebook agreement. PDFs come from the same SVG
sheets; print at 100%, without browser headers or footers. Backs are separate for gluing/sleeves.
Record observed evidence and remaining scope in game/DESIGN.md and .harness/verdict.json.
Presence/build helpers keep ready:false. Write ready:true only after the user's promised
workflow and actual delivery checks pass. Do not claim installed-agent or customer trials unless
those exact trials occurred. Preserve the original arena code and existing workspaces.