A new ModelStudio Skill ships today — feed it any novel or short story and it scaffolds a full React SPA with branching scenes, AI-generated portraits, cinematic cutscenes, optional TTS narration, procedural Web Audio music, and save slots. Then you open
localhostand play.Repo: github.com/modelstudioai/skills
Skill path:skills/novel-game/
By ModelStudioAI
What it actually does
In one sentence: hand your Agent a novel, get back a playable browser game you can share with a friend.
This is not "generate a plot synopsis and call it interactive". It is not "here is a prompt, go paste it into some web UI". The novel-game skill literally:
- scaffolds a real React project with full scene graph, character roster, branching choices, endings, and a save system;
- generates every asset offline via the
blCLI — character portraits as 5-second looping videos or "breathing" still images, cutscenes as 1080p video or Ken-Burns-style stills, narration via TTS, BGM and SFX procedurally synthesized through Web Audio (no external audio files); - downloads everything to
public/assets/so the running game makes zero API calls; - spins up the dev server. Open the browser, play it, on desktop or mobile.
A typical run produces 8–10 scenes, 3–5 branching endings, 6–8 AI portraits, 5–8 AI cutscenes, a procedural soundtrack, and three manual save slots plus autosave.
Why this skill matters
It is the first true end-to-end demo of the bl multimodal stack. Earlier first-party skills lean on a single modality — docs lookup, prompt studio, financial agent, single-shot short-form video. novel-game is the first that wires video + image + speech + procedural audio + frontend engineering into a single pipeline whose output you can click on.
It pushes "Agent Skill" from tool-call to product delivery. Most skills today expose an API. This one delegates a creative pipeline and hands back a runnable web app. That is a different abstraction altogether.
It is the cleanest "show, don't tell" demo we have. Walking a stakeholder through a live run beats ten slides — the Agent literally turned a novel into a game while we watched.
The pipeline in six steps
The full version lives in SKILL.md. Condensed:
1 · Requirements. The Agent fires a single AskUserQuestion with seven decisions: source material (EPUB / TXT / freeform prompt), game type (visual novel / text adventure / text RPG), UI style (pixel / cyberpunk / ink-wash / minimal), narrative POV, asset mode (video / image / hybrid — hybrid recommended), audio mode (BGM only / +SFX / +TTS narration), target duration (15 / 30 / 60+ minutes).
2 · Story design. Extract 1–3 main lines, 3–5 branching choice points, 3–5 endings, 6–8 characters, 5–8 cutscene moments, plus unlockable codex entries.
3 · Project scaffold. npx create-react-app with the canonical layout: components/, data/, hooks/, styles/, scripts/, public/assets/.
4 · Data model. A single story.js describes the full scene graph — branching choices, flag mutations, cutscene triggers, codex unlocks, ending conditions. A sibling generated-assets.json indexes every AI-generated asset with its local path and type (video / image / mp3).
5 · Implementation patterns. Typewriter text via timed setInterval (40–50ms per char), choice panel with hover affordances, hash routing so any chapter is a deep link, localStorage-backed autosave plus three manual slots, portrait component that auto-detects video vs image, cutscene component with Ken Burns for stills, mobile-aware touch targets (≥44px) and touch-action: manipulation, lazy video loading with explicit memory release on unmount.
6 · Asset generation. scripts/generate-assets.sh drives the bl CLI:
bl video generate --download # portrait or cutscene as 5s 720p video
bl image generate # portrait or background, 768x1024 / 1920x1080
bl speech synthesize --voice longxiaochun # TTS narration
bl video ref # multi-reference video for character consistency
Video jobs take 2–5 minutes each, so the script submits in --async mode with 3–5 concurrent jobs, then batch-downloads. Roughly a 4× wall-time speedup vs sequential.
What does one run cost
Assuming a 30-minute / 15–18 scene mid-tier run in hybrid asset mode:
| Item | Count | Unit | Subtotal |
|---|---|---|---|
| Character portraits (image) | 8 × 768×1024 | bl image generate |
cents range |
| Key cutscenes (720p video) | 5 × 5s | ¥0.9/s × 5s × 5 | ≈ ¥22.5 |
| Scene backgrounds (image) | 8 × 1920×1080 | bl image generate |
cents range |
| TTS narration (optional) | 15 × ~30s | bl speech synthesize |
single-digit ¥ |
Authoritative pricing is on the ModelStudio console. New accounts get free credits — plenty to walk the demo end-to-end.
The headline: roughly a cup of coffee, and you have your own playable visual novel.
Try it in three commands
1. Install the skill:
npx skills add modelstudioai/skills
Pick novel-game from the prompt (or --all to grab the whole bundle).
2. Wire up bl and an API key:
npm i -g bailian-cli
bl auth login
Grab a key at bailian.console.aliyun.com — new accounts get free credits.
3. Ask your favorite Agent (Claude Code / Qoder / Cursor / Cline / …) in natural language:
Adapt the Ye Wenjie arc from "The Three-Body Problem" into a visual novel,
ink-wash style, 30-minute playtime, hybrid assets.
The Agent owns the rest: ask the seven decisions, design the branches, scaffold the React project, write the code, generate every asset, and npm start.
Four engineering details worth calling out
Prompt safety. Video prompts containing weapons, smoking, or explicit violence get rejected. The skill ships a built-in rewrite table — content lessons distilled from dozens of real runs.
Procedural BGM. Music is synthesized live via Web Audio, not pre-baked MP3s. Fixed MIDI pitch arrays, multi-voice layering, convolution reverb, ADSR envelopes, detuned pad sustains. Sounds composed; ships at zero file size.
Mobile gotchas. iOS Safari auto-zooms inputs under 16px. touch-action: manipulation to kill the 300ms click delay. env(safe-area-inset-bottom) for the home indicator. Dual-bind click + touchend to recover responsiveness on edge browsers. All wired in by default.
Video memory release. When leaving a scene, video.pause(); video.removeAttribute('src'); video.load(); is required — without it the WebView leaks frame buffers and a 30-minute play session ends in jank.
Closing
Our bar for first-party skills is one line: someone shipped real output with it, and the design is worth copying.
novel-game clears both bars. Author @lishengzxc used it to produce a complete novel-to-game adaptation — not a screenshot demo, an actual React project people have played. The pitfall guide in SKILL.md is paid-for-in-real-time wisdom.
If you build content, games, or interactive narratives, this is a low-cost weekend to spend.
Share what you build over at Issues, or open a PR if you spot a missing tip in the SKILL.
Repo · github.com/modelstudioai/skills
Skill · skills/novel-game/
Try the model · ModelStudio HappyHorse 1.1 playground
— ModelStudioAI on GitHub
Top comments (0)