Godogen Asset Generator
Autonomous game development for Godot, Bevy, and Babylon.js — an agent skill that generates images, GLB models, rigged characters, and animated sprites.
Evidence shows API keys read from env vars only, explicit user confirmation required before every paid generation, transparent cost tables, and a sidecar mechanism that avoids double-charging and enables free resume. Deducted for: confirmation existing only as documentation instruction without code enforcement, no rollback mechanism, unverified publisher identity, and no sandboxing/isolation disclosure for external paid calls and file writes.
Scripts are self-consistent: uniform JSON stdout/error paths, argument validation, early biped-only abort with alternative guidance, and a complete timeout-resume pipeline. Deducted for: no test coverage, unverifiable module signatures (tripo3d, xai_sdk, google.genai; suspicious model version 'gemini-3.1-flash-image-preview'), hardcoded costs prone to drift; static review cannot exceed 10.
Scope is clear (game asset generation), with declared boundaries (no rembg for textures, glb for quadrupeds), trigger conditions, and cost ranges. Deducted for: core function entirely depends on overseas paid APIs (Gemini, xAI, Tripo3D), a real reachability concern for mainland-China users; allow_implicit_invocation=true risks false triggering of paid calls on broad 'generate art' requests.
Well-layered docs (SKILL.md → rembg.md → tool scripts), thorough cost tables, examples, and known-limitation disclosure; MIT license and CHANGELOG reference exist at repo level. Deducted for: no skill-level version/changelog, missing requirements file for asset_gen.py's Python dependencies, maintenance responsibility only via a personal Twitter handle, publisher unverified.
Cost quantification, dual-model selection guidance, double-charge prevention, loop-frame detection, and asset-manifest conventions show real marginal value over manual work, with directly usable PNG/GLB outputs. Deducted for: static review cannot verify generation quality or key-path reproducibility; capped at 7 under anchors, and output quality depends on external APIs.
Code and docs corroborate each other; costs and parameters are consistent between SKILL.md and asset_gen.py — auditable primary material. Deducted for: no CI workflows or committed test suites covering key paths, demos only as a YouTube link unverifiable statically, no third-party execution evidence; capped at 5.
- Core function depends on paid overseas APIs (Gemini, xAI, Tripo3D) that may be unreachable from mainland-China networks; every call costs real money — provision API keys and confirm billing expectations first.
- allow_implicit_invocation is true; broad 'generate art' requests may unintentionally trigger paid calls — require explicit confirmation in sensitive environments.
- asset_gen.py's dependencies (xai_sdk, google.genai, tripo3d modules) and cost figures are hardcoded with no requirements file and no tests; version drift may cause silent failures — verify before use.
- Publisher identity is unverified by the FollowSkills registry; maintenance and update paths rely on personal channels.
What does this skill do, and when should you use it?
Godogen is a repository for autonomous game development whose core artifact is a cross-engine asset-generation skill (asset-gen). It calls Gemini and xAI Grok for PNG images, Tripo3D for image-to-GLB models, biped rigging, and animation retargeting, and produces animated sprites via video frame extraction plus background removal. Every call is a paid API, and the skill embeds cost references, failure recovery (Tripo3D resume to avoid double-charging), and an asset-manifest convention. The repo itself is not a game but a publisher: pick an engine and a host agent (Claude or Codex), publish into a fresh game repo, and the agent builds the actual game inside it.
Runs Python tool scripts from the project root to generate PNGs from text prompts (Gemini: 512/1K/2K/4K; Grok: 2¢, high quality but imprecise); supports image-to-image with --image references, kit slicing via grid_slice.py, and rembg matting; builds animated sprites through 'reference → pose → video → extract frames → loop-trim → rembg' (video 5¢/s, 720p, 1–15s); calls Tripo3D for GLB (30¢, HD 60¢), biped-only rigging (+25¢), and retargeting ~100 animation presets (10¢ per clip); recovers timed-out Tripo3D jobs for free via the resume command and the saved task id sidecar; prints JSON to stdout with cost_cents, and requires an asset manifest in README.md with an in-game Size column.
- A solo developer who wants Claude Code to scaffold a Godot 4 (C#/.NET) project, generate assets, and prove the result from a running game, starting from one game description.
- A Rust/Bevy developer needing code-first ECS scenes with offscreen capture plus auto-generated textures and simple objects.
- A web game developer publishing a Babylon.js/TypeScript project the agent builds and verifies via a live URL or recording.
- A team producing multiple character actions from one reference image: pose variants, video-to-sprite animation, with action chaining for positional continuity.
- A budget-conscious creator who wants explicit per-step costs (texture 2¢, reference 7¢, full 3D asset 37¢, rigged character with animations ≈ 92¢) before committing.
What are this skill's strengths and limitations?
- End-to-end pipeline covering images, 3D, rigging, animation retargeting, and sprite animation in one skill.
- Transparent costs: documented per-operation prices with an explicit rule to confirm with the user before generating.
- Strong engineering details: Tripo3D timeout resume to avoid double-charging, loop-frame detection, direction-flip savings, manifest with in-game Size column.
- Dual host support (Claude and Codex) and three engines (Godot, Bevy, Babylon.js) chosen at publish time.
- Every call is a real paid API (from 2¢ to ≈ 92¢ per rigged character workflow), so experimentation costs accumulate.
- Heavy setup: three API keys plus many system packages (ffmpeg, ImageMagick, vulkan-tools, etc.).
- Generators have weak spatial sense: unreliable direction, muddy small sprites — every PNG must be reviewed before GLB conversion.
- Only tested on Ubuntu, Debian, and macOS; no test suite is mentioned and Windows is unverified.
- Full runs can take hours, typically needing a server or GPU instance.
How do you install this skill?
Prerequisites: Godot 4 (.NET build) on PATH, Rust/Cargo, Node.js 22.12+, Chrome/Chromium with hardware WebGL2, Python 3 + pip, ffmpeg, ImageMagick, vulkan-tools, xvfb, and env vars GOOGLE_API_KEY, XAI_API_KEY, TRIPO3D_API_KEY. Tested on Ubuntu, Debian, and macOS. Then publish a game repo: ./publish.sh --engine godot --agent claude --out ~/my-game (engines: godot/babylon/bevy; agents: claude/codex; --force wipes the target first). Claude gets CLAUDE.md + .claude/skills/; Codex gets AGENTS.md + .agents/skills/.
How do you use this skill?
Open the published game repo with Claude Code or Codex. The skill's default prompt is: 'Use ${ASSET_SKILL_COMMAND} to generate images, 3D models, or animated sprites for this game.' with implicit invocation allowed. Example image call: python3 ${ASSET_GEN_SKILL_DIR}/tools/asset_gen.py image --prompt "..." -o ${RUNTIME_ASSET_DIR}/img/car.png; 3D uses the glb/rig/retarget subcommands. Runs can take hours, so the README suggests offloading to a GPU server with tmux/screen and remote-control interfaces.
How does this skill compare with similar options?
Unlike ad-hoc AI asset workflows, Godogen chains image, video, and 3D generation into an automated pipeline with cost control and failure recovery, and it insists on 'proof over claims' — judging results from the running game rather than a clean compile. It explicitly targets both Claude Code and Codex as host agents, which single-file skills typically do not.