An AI agent builds complete mobile games end to end — design, code, art, music, automated tests, store assets, monetisation and a signed release — with a human involved at only two gates.
This is not a prompt collection. It is the working machinery — procedures, guardrails, headless build and QA automation, store packaging — plus 61 numbered rules distilled from real failures, and a complete game as proof that the machinery produces something real.
An agent starts at AGENTS.md, reads which market it is building for from
factory.json, and goes.
concept → human approval → pure game logic + tests → screens → generated art & audio
→ feature baseline → QA gates → store package → milestone build → feedback → LESSONS.md
The human decides two things: is this concept worth building, and does this build feel right. Everything else — architecture, implementation, testing, asset generation, store compliance, packaging — is the agent's.
Two commands are the whole quality bar:
pipeline/check_game.sh <slug> # import, unit tests, smoke run, autoplay bot, exports
pipeline/build_stores.sh <slug> # one artifact per store, with cross-store leakage checks- All game rules live in a pure, scene-free class. That one decision is why 146 tests run with no display and no device — an agent can verify its own work without a human or a phone.
- A bot plays the real UI. Unit tests miss "the button does nothing"; a headless run drives the same handlers a finger does, to game over, three times.
- Guards for the bugs an agent cannot see: layout-overlap tests at four aspect ratios, font glyph coverage (no tofu boxes), scroll-blocking detection, and a screenshot composer that doubles as a layout oracle.
- Platform features hide behind
available()— billing, notifications, network. A missing plugin makes a feature invisible, never a crash. - Every failure becomes a numbered rule. LESSONS.md is why the second game costs less than the first.
Store rules, billing SDKs, listing limits, compliance, language and cultural design live in market modules, and nowhere else:
| Module | Stores | Language |
|---|---|---|
| markets/international.md | Google Play, itch.io, Steam | English (default) |
| markets/iran.md | Cafe Bazaar, Myket | Persian |
Say "target the Iran market" or "cover both" and the agent switches the field and reads a
different module. Adding a market is one file — copy _TEMPLATE.md and
fill eleven sections. No core document is ever forked for a market, and nothing locale-specific is
applied unless its module is selected.
games/mergedrop — an offline-first merge puzzle built through this factory for
the Iran module and packaged for two stores.
Built by the agent: the merge engine and its difficulty curve, a daily challenge with a provably identical board for every player, a companion character that reacts to your streak and rank, an economy, local reminders with a tested policy, an offline-first global leaderboard, a synthesised two-stem soundtrack, generated art, and store listings.
⬇ Download the APKs · four builds, all signed.
| Component | What it is |
|---|---|
addons/myket |
A Myket in-app billing plugin for Godot 4 (Java + AIDL + GDScript), written from scratch and MIT — none with a clean licence existed. Source included. |
scripts/iap.gd |
One billing API driving multiple stores, loading plugins dynamically so builds without them still run. |
pipeline/build_stores.sh |
One artifact per store; fails the build if a store's billing permission or classes leak into another's. |
server/ |
Dependency-free Python leaderboard; the client queues scores offline and uploads when a network appears. |
scripts/jalali.gd |
Jalali/Shamsi calendar in pure GDScript, with tests. |
scripts/notify.gd |
Local reminders with a tested policy: quiet hours, timezone-correct scheduling, back-off when ignored. |
pipeline/make_screenshots.py |
Renders store screenshots by mirroring the game's own layout constants — and catches layout bugs doing it. |
| File | Purpose |
|---|---|
| AGENTS.md | Start here. How the factory runs, and the market switch. |
| factory.json | The only switchboard: markets, language, engine. |
| markets/ | Store facts — the only place they live. |
| GAME_BLUEPRINT.md | What every game must ship, and why. Market-agnostic. |
| PLAYBOOK.md | How to build it, step by step. |
| LESSONS.md | 61 binding rules from real failures. |
| ENGAGEMENT.md | Retention design patterns. |
| FACTORY.md | Toolchain and machine setup. |
| THIRD_PARTY.md | Vendored components and their licences. |
Everything is headless — no editor, no display, no device.
cd games/mergedrop
godot --headless --path . -s addons/gut/gut_cmdln.gd -gdir=res://tests -ginclude_subdirs -gexit
godot --headless --path . -- --autoplay # a bot plays the real UI to game overAndroid builds need Godot's Gradle template, JDK 17 and the Android SDK. The leaderboard server is
cd server && PORT=3000 python3 scoreboard.py.
- The agent writes the code, tests, art, audio and store material. A human still creates store accounts, pays fees, holds the signing key and presses publish — every market module lists those explicitly as human-only.
- The included game is packaged and verified, and not yet published to a store.
- In-app purchases are integrated and verified inside the built package, but a real purchase can only be exercised on a device with the store app installed and products created.
- The CI workflow ships at
ci/and is not active — enabling it needs a token with theworkflowscope.
MIT — see LICENSE. Vendored fonts are SIL OFL and vendored addons keep their own MIT licences under their authors; see THIRD_PARTY.md.
© 2026 1xai Games Studio