The Vibe Coding Gap
My feed has been full of "I built X with AI in one afternoon" posts for months. Chrome extensions, SaaS tools, mini-games — everyone seemed to be shipping.
I wanted to make an interactive fiction game. Branching story, multiple endings, character art, music. A visual novel you play in the browser.
So I tried the standard vibe coding approach:
"Build me a branching story engine in React" → AI wrote something. It ran. Then I spent 3 days debugging state management bugs in the branching logic.
"Generate character portraits" → Got images, but each one looked like a different art style. No consistency.
"Add background music" → AI suggested downloading royalty-free tracks. Now I'm managing audio files and dealing with licensing.
"Make choices meaningful" → The branching logic AI wrote had fake branches everywhere — choose A or B, see the same next scene.
After two weeks: a broken prototype with inconsistent art and no audio. The "vibe" was gone.
Realization: vibe coding is great for single-domain tasks. Multi-domain projects (story + logic + art + audio + frontend) fall apart because each prompt is isolated.
What novel-game Does Differently
Found novel-game in Alibaba Cloud's Model Studio Skills repo. It's not a "prompt AI to help you code" workflow. It's a pre-packaged pipeline that handles the entire interactive fiction production at once.
One input:
Use novel-game to create an interactive fiction game.
Genre: Sci-fi thriller. Programmer discovers company AI has become sentient.
Player decides what to do. 15 minutes. Image mode. Include BGM.
One output (25 minutes later): a complete, playable game in the browser.
What I Got vs What Vibe Coding Gave Me
| Aspect | My 2-week vibe coding attempt | novel-game (25 min) |
|---|---|---|
| Branching | Buggy, fake branches | Real forks (choice = different scenes) |
| Art | Inconsistent styles | Batch-generated, consistent |
| Audio | Gave up | Procedural Web Audio (zero files) |
| Save system | Didn't get to it | Auto-save + 3 manual slots |
| Mobile | Didn't get to it | Portrait-first, touch-ready |
| Status | Broken prototype | Playable game |
Why It Works Where Generic Vibe Coding Didn't
The difference: novel-game embeds game design expertise, not just code generation.
Things I didn't know I needed until I hit them:
- "Choice = fork" principle — every major option must lead to genuinely different scenes (no fake branches)
- When to merge branches back vs let them diverge to separate endings
- Genre → visual theme inference (sci-fi → minimalist modern, wuxia → ink wash)
- Procedural audio that matches scene mood without external files
- Asset pre-generation (zero API calls at runtime)
These are lessons indie game devs learn over months. novel-game packages them into the pipeline. You skip the learning tax.
Limitations (being honest)
- Art style is "digital painting" — you can't guarantee anime style or photorealism
- BGM is procedural — loops become noticeable in longer sessions
- Only works for interactive fiction / visual novels (not action games)
- Costs money — my 15-min game was ~$2. Free credits cover the first one.
Setup
npm install -g bailian-cli
bl auth login
npx skills add modelstudioai/skills --skill novel-game -g
Needs: Node.js >= 18 + Free API Key
The Takeaway
If you've been watching vibe coding posts and wondering why your own attempts don't produce finished products — it might not be you. Generic "tell AI to write code" breaks down on multi-domain projects.
The answer (at least for interactive fiction) isn't better prompting. It's domain-specific packaging — someone already figured out the hard parts and wrapped them into a pipeline.
25 minutes vs 2 weeks. Same person, same idea, different approach.
- GitHub: modelstudioai/skills
- Free API Key: Get one
- Bailian CLI: Install