歸藏 Game Skill + Claude Code (Fable 5.2)Partly tested🇨🇳 CN

Guizang built a xianxia 3D action game, Mist-Locked Mountain Gate, in one afternoon with Claude Fable 5.2, then turned the lessons into an open-source skill

Guizang (@op7418) only asked for "Chinese xianxia style" and had Claude Fable 5.2 build a top-down 3D action game, Mist-Locked Mountain Gate, in one afternoon: one map, a mob fight and a two-phase boss, with a three-hit combo, dodge, block and parry, an ultimate and 8 voiced lines, packed into one ~33 MB HTML file that plays in desktop and mobile browsers. Characters started as concept art in Jimeng, became rigged 3D models in Tripo, and the music came from Suno. He then wrote the workflow and his list of pitfalls into guizang-game-skill: install it in Claude Code or Codex, describe your game, and it asks questions first, connects your asset tools, checks its own screenshots and packages the result. We inspected the repo, ran its local server script and scanned it with SkillSpector; we did not build a game.

Visit site 歸藏 Game Skill + Claude Code (Fable 5.2) ↗
Pricing
Free
Region
Global
Where it's from
Chinese community
Added
Nov 14, 2026
Tested on
Nov 14, 2026
Version we used
op7418/guizang-game-skill 6001e56; author used Claude Fable 5.2 (some Opus 5.5 test runs)
Categories
CodingDesign
Original author
歸藏 (guizang.ai) @op7418 ↗
Our test
Partly tested
Cost / time
The skill is free (AGPL-3.0). Building a game uses a lot of Claude quota (he used Fable 5.2; the README says about 450M tokens). Tripo, Jimeng and Suno bill by your plan; without them it falls back to free assets and procedural content, with weaker visuals and sound.
Source
X @op7418 ↗GitHub op7418/guizang-game-skill ↗

The task

Without game-dev experience, get AI to build a browser 3D action game with real character models, hit feel and music instead of grey boxes.

Steps

  1. Install Node.js 18+ for local preview and packaging. Send the repo link to your agent to install it, or git clone https://github.com/op7418/guizang-game-skill ~/.claude/skills/guizang-game-skill; for Codex and others, clone into their skills folder.
  2. Describe the game and say which asset tools you have. It first asks about references, camera, mood, scope and your credits.
  3. Give it reference videos and insist on research before code. The flow: research and plan → concept art and assets (Jimeng concept art, Tripo to rigged 3D) → animation → hit feel and effects → audio and packaging.
  4. On day one, have it set up self-test tools: frame stepping, URL parameters that jump to a level, and an F3 debug panel. It checks its own screenshots and tells you what it looked at versus what merely didn't error.
  5. Play it and give feedback in plain words like "the moves feel stiff" or "hits feel weak"; the AI turns that into concrete fixes. He played 7 rounds with 5–10 issues each.
  6. Pick a package: one HTML file, a static site that shows the scene first and loads the rest, or an offline bundle for restricted containers such as Xiaohongshu mini tools.

Prompt

Make a top-down 3D action game for the browser with a wuxia theme: one map, one wave of enemies and one boss.
I have Tripo and Jimeng available.

What the author says

Guizang @op7418 (X, 12 Nov 2026): Fable 5.2 made this 3D action game "in one afternoon with 4.5 billion tokens", handling models, level design, animation and optimization, all fused into one 30 MB HTML file. The repo README says about 6 hours and 450 million tokens (part of it Opus 5.5 test runs), much of it spent on rework; raw assets near 100 MB packed to about 33 MB; 5 rigged characters, 2 weapons, 12 props; every creative decision was the AI's. The post and README differ tenfold on tokens, and we can't verify which is right. He also says it is not one-prompt-to-finished-game, the skill has been through one full project plus one plan-level comparison, and quality isn't guaranteed. Generating models straight from text looked bad at first; going concept art first, then 3D, was a big step up.

Our test

On 14 Nov 2026 on Linux we cloned op7418/guizang-game-skill (6001e56): a 99-line SKILL.md, 6 reference docs (assets, character animation, combat feel, effects/UI/audio, pitfalls, packaging) and 4 scripts. The two packaging scripts, build_single_html.mjs and build_minitool.mjs, pass node --check. serve.mjs run from a temp folder returned 200 for the index and 403 for a ../ path. The online demo at https://op7418.github.io/guizang-game-skill/ returned 200 and finished loading in headless Chromium in about 8 s, but with software-only rendering on our server we couldn't capture a frame within 30 s, so we didn't see the game itself. SkillSpector static scan: 56, HIGH, 6 findings; the 3 high ones all say the packaging scripts couldn't be fully parsed, and 2 low ones come from the LICENSE text. We did not build a game with Claude or Fable 5.2 and did not test Tripo, Jimeng or Suno.

What we found when we tried it:The repo matches its README, the pitfall list is specific (lost key-up events, preset animations drifting characters off their spot, bloom blowing out), and the local server script works. Game quality is something you'll only know by building one with your agent.

How to get in

The original post is in Chinese on X. The repo has an English README (README.en.md), and the demo game plays in any browser: https://op7418.github.io/guizang-game-skill/ . Jimeng is a ByteDance image tool; outside China use any concept-art image generator instead.

Who it suits

People with generous Claude Code or Codex quota who want a browser game or playable prototype, ideally with image-to-3D and image tools at hand.

Pitfalls

  • It is not one prompt to a finished game; you play and give feedback repeatedly.
  • The skill has been validated on one project; other genres are untested.
  • Asset tools matter a lot; text-only model generation looks poor.
  • AGPL-3.0: redistributing a modified skill requires open-sourcing it; the demo game's art, voice and text are the author's and not for reuse.
  • A SkillSpector static scan gives it 56 and DO NOT INSTALL because its packaging scripts couldn't be fully parsed; we read them and saw nothing malicious, but read the scripts yourself before installing.

More in this category