Guizang Game Skill turns Claude Code into a 3D game builder
An open repository packages a step-by-step agent skill for making playable browser 3D action games, with a 33MB single-file demo attached.
On 2026-11-14, the repository op7418/guizang-game-skill listed 327 stars and 48 forks, three days after it was created on 2026-11-11. It packages a Claude Code 3D game skill: instructions and reusable scripts an AI coding agent follows to build a browser action game from an empty folder. Main language is JavaScript; the main branch documents and scripts are AGPL-3.0.
What the skill asks the agent to do
Installation is either handing the repository link to the agent or cloning it into ~/.claude/skills/guizang-game-skill; the source says Codex and other tools use their own skills directories.
The prerequisites are given as a table. Node.js 18 or newer is required for the local server and packaging. Tripo handles image-to-3D and rigging, an image tool such as Jimeng produces character sheets, icons, title lettering and voice, and Suno supplies music. Without them the skill falls back to free assets and procedural geometry, which the author says looks worse. Generating models straight from text gave poor results; going through a concept image first was a clear step up.
The workflow runs in five stages: research and planning, design and assets, animation, combat feel and effects, sound and packaging. Before writing code the agent is meant to ask about references, camera angle, mood, scope, and which asset tools and how much credit are available. It builds its own test tools early, including frame-by-frame stepping, URL parameters that jump to a level, and an F3 diagnostic panel. It screenshots and checks its own output, then separates what it verified from what merely produced no errors.
Combat feel gets a checklist rather than vague advice: wind-up and recovery frames, hit-stop, graded knockback, blocking and parrying, enemies attacking in turn instead of all at once, and dodges that interrupt an animation. A pitfalls file lists symptoms, causes and fixes, and the packaging references cover squeezing large assets into a small build.
The demo it ships with
《雾锁山门》 is a top-down 3D action game in a Chinese fantasy setting: one scene, one group of minor enemies, one boss fight with a second phase and an entrance cutscene, story and voice-over. The source counts 5 rigged 3D characters, 2 weapons, 12 scene props and 5 drop types, with three-hit combos, dodge, block, parry, executions, potions, sword flight and an ultimate move. Audio runs to 8 voice lines plus music, ambience and effects. The author says the AI made every creative decision, and that he specified only a Chinese fantasy setting.
Raw assets came to nearly 100MB; the packed single file is about 33MB of HTML with no network requests. Three builds come from the same source: an online version that renders the scene before loading the rest and adds touch controls on phones, the single file, and a container build of roughly 8MiB made for Xiaohongshu mini-tool limits, given as no network access, no WASM and a 10MiB package cap.
The author is explicit that this was not one prompt: roughly six hours and 4.5亿 Token with Fable 5.2, part of the spend going to tests run with Opus 5.5, then seven rounds of playtesting with five to ten complaints per round.
Limits and open questions
The skill has run end to end on one project, plus one comparison test at the planning stage. The author states that it reduces wrong turns but does not guarantee the quality of the output. Licensing splits in two: the main branch is AGPL-3.0, while the demo game's art, models, voice, music and text on the gh-pages branch are all rights reserved, so republishing it under another name or using it commercially needs permission. The source gives no comparison against an agent working without the skill, and no pricing or credit figures for Tripo, the image tool or Suno, which matter if you plan to follow the asset pipeline. The two model names it cites are not explained.
What it means for building apps with AI
The transferable part is the shape of the pack, not the game: interview the user first, fix the stages, build self-checking tools early, and write down the mistakes so they are not repeated. That pattern applies to any build you brief an agent on. Draftship, an AI app builder that writes, builds and hosts an app from a plain-language description, covers comparable ground for app projects in its guide to writing effective prompts. Draftship has no connection to this repository and does not claim support for 3D game pipelines.
Primary sources
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.