Game Skill for Claude Code builds a 3D action game into a single 33 MB HTML

⚡ The gist in 5 seconds

• published a Skill for Claude Code and Codex: the agent takes an idea all the way to a playable 3D action game right in the browser.
• The demo runs on three.js and WebGL, opens on both PC and phone, and is packed into a single ~33 MB HTML with no network requests.
• For proper visuals you need Node.js 18+, Tripo, and an image generator. Without them, the Skill will assemble primitives instead of models.

🔍 What happened

An author going by published the repository op7418/guizang-game-skill under the AGPL-3.0 license. Inside there's a SKILL.md with a step-by-step process and a references folder: assets, characters and animation, combat feel, VFX and sound, a list of pitfalls, packaging. Plus scripts: input handling, a local server, and three build tools.

The demo is a top-down action game in the xianxia style. One location, one wave of mobs, a boss, a story, and voiceover. The author says the AI made all the creative decisions: he merely said "make a Chinese xianxia fantasy." The build took about 6 hours and 450 million tokens — part of the runs were on Opus 5.5. A lot of time went into rollbacks and rework.

What got generated: 5 characters with skeletons, 2 weapons, 12 environment props, 5 types of loot. Concepts were made in, models and rigging in Tripo, with the agent writing the prompts itself. Combat features a three-hit combo, dash, block, parry, finisher, potion, sword flight, and the " " ultimate. The boss has two phases and voice lines. Sound: 8 voice lines, music, ambience, effects.

Size: sources are nearly 100 MB, and after packing it's a single HTML of roughly 33 MB with all models, textures, music, and voice. Two more versions are built from the same code. The online one streams assets as you go, with a touch joystick and quality settings. The containerized one fits mini-app limits: no network, no WASM, packing up to 10 MiB — it came out to about 8 MiB.

Installation is a one-liner: git clone https://github.com/op7418/guizang-game-skill ~/.claude/skills/guizang-game-skill. For Codex — into its skills directory.

🧠 Our take

Agent skills have gone from demos to production in a year, and this release is valuable because the author honestly shared not a "wow result" but the meat: a list of pitfalls, verification criteria, packaging scripts. It's closer to a checklist from an experienced tech lead than a prompt template. Our assessment: the value here isn't in the model but in the codified sequence of actions.

What it means financially. 450 million tokens for one evening isn't "free" — it's a noticeable API bill, plus subscriptions to Tripo and an image generator, plus seven rounds of testing that the author openly mentions. The barrier to entry is higher than the headline suggests. We believe the real payoff here is in prototypes and vertical slices, not finished commercial games.

Who wins: solo developers and small studios that need a playable draft in an evening instead of a month of gray-box geometry. Who loses — or rather, gets disappointed — is anyone expecting "one line and a finished product." The author himself warns: he doesn't understand game dev, tells the agent "the movements feel stiff," and gets them translated into technical terms. Without that dialogue, nothing will come of it.

⚠️ What we don't know

All the numbers are one person's self-report. The README states outright: the Skill has fully passed one project and one plan-level test, and it doesn't guarantee quality. Judging how playable the demo is requires playing it yourself — there are no performance measurements or comparisons with manual assembly in the source. At the time of editing, the repository has 165 stars and 23 forks: it's too early to judge reliability. The licensing is subtle: the Skill itself is under AGPL-3.0, while the demo content is © 2026, all rights reserved, redistribution and commercial use prohibited.

Forecast: within a quarter, we'll see whether public games built with this Skill appear — the author asks people to send in their results. Fewer than three third-party projects will mean the barrier to entry (Tripo, tokens, testing rounds) is indeed high.

Disclaimer: a single source — the author's README; the figures and effort estimates have not been independently verified.

Source: GitHub — op7418/guizang-game-skill