Brainstorm a game's design with one AI; hand the result to another AI to build it — backed by deterministic checkers that won't let a broken handoff through.
🌐 中文版 / Chinese version: ProdaZhang/aigd-zh
Who it's for — and what each role gets:
- Product / design / solo devs — brainstorm the feature with an AI (feed it a design doc if you have one, else it interviews you) → a clickable prototype to try → iterate → a finalized package another AI can build from.
- Engineers — a typed interface contract (client = server) + config spec + acceptance cases → another AI (or you) implements it without ambiguity.
- QA — acceptance cases + a visual Excel checklist → another AI runs the tests, or do black-box testing straight from the Excel.
You only do two things: brainstorm the design with an AI, and set the numbers. The flow handles the rest — it turns the discussion into structured output (rules carry IDs · numbers live in config · UI-DSL · interface contract · acceptance cases), gates consistency with deterministic checkers, and packages a platform-agnostic handoff that another AI can implement directly — verified by a real run in this repo (below).
Discussion-driven · doesn't decide your numbers for you · doesn't bind to an engine.
This is the repo landing page. Get a feel by running it first →
aigd/examples/potion-crafting/· full methodology →aigd/README.md.
The most common way game-design handoff goes wrong isn't too little documentation, it's docs and config quietly drifting out of sync ("doc fixed first, table changed later without writing it back"), with downstream each reading its own → forked implementations. AIGD blocks this three ways: structured output (rules tagged with numbers / numbers living in config / prose only referencing table[primary key].field), explicit ledgering of the undecided ([to confirm] handed to a person to decide), deterministic machine checks (config_check/value_check/manifest_check, 0 major counts as handoffable).
↑ a real run — a from-scratch implementation built only from the example's handoff package passes all 5 acceptance scenarios (5 passed, 0 failed).
↑ a real run of config_check — it catches a config↔doc drift (UNDOC_COL), then passes once it's fixed.
Copy these 7 folders as a whole into the host's skills directory, keeping them at the same level:
aigd/ aigd-concept/ aigd-system/ aigd-iterate/ aigd-handoff/ aigd-sync/ aigd-ui-capture/
| harness | install to |
|---|---|
| Claude Code | .claude/skills/ |
| ZCode (Claude family) | ~/.zcode/skills/ |
| Gemini CLI | ~/.gemini/skills/ (or gemini skills install https://github.com/<owner>/<repo> to install from the repo in one step) |
| Codex | ~/.codex/skills/<name> (or use its built-in skill-installer to install from the repo; restart after installing) |
| Copilot CLI 1.0.63 | ❌ no skills mechanism, goes through AGENTS.md/MCP/plugin, needs adapting |
The package structure (SKILL.md + name/description frontmatter) is common across Claude Code / ZCode / Gemini / Codex (all tested in practice); Copilot 1.0.63 doesn't support it. Where to install / how to invoke / tool-name mapping: see aigd/references/harness-adapt.md. Running the checkers needs Python (mostly pure standard library; some need openpyxl/Pillow, see aigd/references/scripts/requirements.txt).
- Install the 7 folders.
- Read
aigd/README.mdto understand the 6-piece set + the flow. - Run the three check commands in
examples/potion-crafting/to see "machine-check gating" in action. - New project: call
aigd(let it route if you don't know which step) or directlyaigd-conceptto set the concept →aigd-systemsystem by system →aigd-handoffto finalize.
Honestly drawing the scope, to avoid misuse:
- Manages structure and consistency, not balance: the checkers look for broken links / coverage / monotonicity / schema drift, they don't judge whether the numbers are fun — balance is a job for people / dedicated tools.
- The html prototype verifies information architecture and flow, can't verify feel / timing / networking: enough for UI-dense systems (inventory / shop / progression); for the "feel" of real-time combat / physics / multiplayer interaction, leave it to an engineering prototype or dedicated verification, don't treat a clickable wireframe as having verified feel.
- Doesn't decide your numbers / conventions for you: anything undecided is uniformly tagged
[to confirm]and handed to a person, the AI doesn't make things up. - Scope of evidence for "another AI can develop from the handoff package": cross-validated consumer-side dual implementations on a real system within the same model family (Claude), run through; cross-vendor models (GPT/Gemini) not verified. It's strong evidence, not a universal proof.
| harness | install | discovery | routing | execution |
|---|---|---|---|---|
| Claude Code (native · real project) | ✅ | ✅ | ✅ | ✅ |
| ZCode 3.1.3 (Claude family) | ✅ | ✅ | ✅ | ✅ |
| Gemini CLI 0.47 (Google · cross-vendor) | ✅ | ✅ | ✅ | ✅ |
| Codex 0.140 (OpenAI · cross-vendor) | ✅ | ✅ | ✅ | ✅ |
| Copilot CLI 1.0.63 (GitHub) | ❌ no skills mechanism | — | — | — |
Four harnesses tested working in practice (discovery + routing + execution), including the two cross-vendor ones Gemini and Codex; Gemini installs from this repo in one step with gemini skills install <repo>, Codex with its built-in skill-installer. Copilot CLI 1.0.63 was tested and does not support the SKILL.md skills mechanism (goes through AGENTS.md/MCP/plugin), aigd needs adapting to be usable. Where to install / how to invoke: see aigd/references/harness-adapt.md.
MIT © 2026 ProdaZhang. Free to use / modify / redistribute, just retain the copyright and license notice.
v0 (pre-release). patterns/ is a starter pack that will grow (currently: 5 core loops / a combat-unit progression paradigm / 10 number-tuning traps). Checker tests are in aigd/references/scripts/tests/ (pure stdlib runner). For contributing, see CONTRIBUTING.md.