Describe a game. Godogen plans it, writes the code, generates assets, runs the engine, checks screenshots, and fixes what looks wrong.
This repo is not a game. It is the source for a generator that produces games: godogen -> game repo -> game. You publish the skills into a fresh game repo, choosing the engine and host-agent flavor, then the agent runs inside that repo to build the actual game.
Source layout
The source is organized along the engine axis:
shared/ — engine-agnostic godogen stages, asset-generation tooling, shared stop hook, and common game-repo instructions
godot/ — Godot-specific godogen stages, Godot capture helpers, and the godot-api skill
bevy/ — Bevy-specific godogen stages, Bevy capture helpers, and the bevy-help skill
Claude Code vs Codex is a publish-time render choice, not a separate source tree. The root publish.sh renders the right runtime layout for the chosen engine and host agent.
What skills do
Godot 4 output — real C#/.NET projects with proper scene trees, scene builders, scripts, and asset organization.
Godot Android export — debug APK export remains available when the user requests an Android app.
Bevy output — Rust/Bevy projects with code-first scenes, local Bevy docs lookup, deterministic capture guidance, and final proof bundles.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
Extracted system prompts from ChatGPT (GPT-5.5 Thinking), Claude (Opus 4.7, Opus 4.6, Sonnet 4.6, Claude Code), Gemini (3.1 Pro, 3 Flash, Gemini CLI), Grok (4.3 beta), Perplexity, and more. Updated regularly.
Asset generation — Gemini creates precise references and characters; xAI Grok handles textures and simple objects; Tripo3D converts images to 3D models. Animated sprites use Grok video generation with loop detection.
Frame-grounded self-repair — the agent is carefully prompted to judge progress from captured screenshots, not from code that compiles, so visible defects (clipping, wrong scale, frozen motion, missing assets) drive the next iteration instead of being rationalized away.
Telegram proof push — published repos install a stop hook that pushes the latest screenshots/result/{N}/video.mp4 to Telegram when tg-push and the TG_* env vars are configured. No-op otherwise.
Runs on commodity hardware — any machine with the relevant engine toolchain, Python, and the required API keys can run the pipeline.
The setup script links bevy/skills/bevy-help/docs/ to that folder, clones the Bevy docs sources, and builds local rustdoc for the current stable release. No default path is assumed. See setup.md for the full workstation setup.
Running on a server
A full generation run can take hours, so it's convenient to offload it to a server, ideally a GPU instance, since engine rendering and video capture are much faster with hardware acceleration.
Keep the session alive across SSH drops with tmux or screen.
Install tg-push: the stop hook auto-sends the final proof video to Telegram on completion.
Enable remote control so you can check in and steer the run from any device — both Claude Code and Codex have official remote-control interfaces.
Improving the skills
After a full generation session, ask the agent you used to review how the pipeline performed:
Analyze this session. Were the instructions optimal? Flag anything that was too obvious, missing, or misleading. Did any tools pollute context with noise? Did the capture loop catch the real problems? Any tool failures or workarounds?
A Claude Code plugin that automatically captures everything Claude does during your coding sessions, compresses it with AI (using Claude's agent-sdk), and injects relevant context back into future sessions.
Wrap Gemini CLI, Antigravity, ChatGPT Codex, Claude Code as an OpenAI/Gemini/Claude/Codex compatible API service, allowing you to enjoy the free Gemini 3.1 Pro, GPT 5.5, Claude model through API