vibe-creating-skill

An open-source AI video-prompt skill — for Seedance, Sora, Kling, Veo & more

Turn a rough idea into a model-ready text-to-video prompt. Let creation return to expression.

English · 中文

Vibe Creating is an open-source, bilingual prompt-engineering skill that rewrites a rough idea, story, feeling, or over-specified shot script into a clean, model-friendly text-to-video prompt — and first judges whether your input even suits this style. It follows the open Agent Skills standard (a single SKILL.md), so it runs in Claude Code, Codex, OpenClaw, Hermes, and any agent on the same standard — or as a rule (Cursor) or system prompt (GPT, Gemini, any LLM). It works with any text-to-video model, and you can run many of them through one API on Atlas Cloud — Seedance 2.0, Kling, Veo, Hailuo, Wan, Vidu, Gemini Omni, and Grok Imagine, among others.

✨ What is Vibe Creating?

As text-to-video models get smarter, prompting gets simpler. Instead of over-specifying focal lengths, shot numbers, and frame-by-frame scripts, you focus on telling the story and trust the model to find the right shots, light, and rhythm.

Vibe Creating is that paradigm — introduced by ByteDance / Volcengine alongside the Seedance 2.0 video-generation model. This repo turns its methodology into a reusable AI video prompt skill:

• 🎯 Focus on Story — describe the situation, the texture of the air, the flow of emotion; let the model interpret.
• 🤝 Trust the Model — drop low-value technical parameters; keep and translate the shot intent.
• 🧭 Judgment-first — it decides whether your input even suits this style before rewriting, so it never flattens a precise shot list you actually wanted.

It is not a "make everything shorter" tool. Read the full philosophy for the idea behind it.

🧠 How it works

The skill is judgment-first: it scores your input on three axes — Scenario × Expression × Information — picks the lightest action that fits, and always replies in the same four-part format.

Input  (idea · scene · script)
  │
  ├─ Suits Vibe Creating? ───── no ─────▶ keep as-is  (explain the mismatch)
  │     │ yes
  ├─ Enough info? (4 layers) ── no ─────▶ ask first   (1–3 short questions)
  │     │ yes
  └─ How polished already?
        ├─ strong      ─▶ pass-through
        ├─ near / messy ─▶ light cleanup
        └─ execution    ─▶ direct rewrite  (strip params, translate intent)
                                │
                                ▼
        Four-part output:  Judgment · Action · Result · Notes

The four layers it checks for: visual anchor (the main subject), action / state (what's happening), local tonality (one mood word), and video theme (use case + visual style).

And every response comes back in that same fixed four-part format:

Judgment — does it suit Vibe Creating, is it already usable, is the info enough Action — one of pass-through / light cleanup / direct rewrite / ask first / keep as-is / optional VC version Result — the rewritten prompt, the kept text, or a clarifying question Notes (if any) — what technical control was stripped, what constraints were kept

For example — You: Express the feeling of freedom. Skill: Judgment: suits VC, but too abstract to render — no visual anchor. Action: ask first — "Who is free, and where? A child in open fields, a diver, a city rooftop at dawn? Any style — photoreal, anime, ink-wash?"

📦 Install

The skill is one portable SKILL.md folder. Pick whichever option fits your setup.

Option 1 — One-liner (recommended). Auto-detects your installed agents and copies the skill into each:

npx github:Alisa0808/vibe-creating-skill

Add --all for every known agent, name agents explicitly (… claude codex), or target any folder with --dir <path>.

Option 2 — Clone and copy. Same file, different home per agent:

git clone https://github.com/Alisa0808/vibe-creating-skill.git
cp -r vibe-creating-skill/skills/vibe-creating-prompt <your-agent-skills-dir>/

Option 3 — Cursor or any other LLM. Paste the body of the skill file (or its Chinese edition) as a rule (Cursor) or as the system prompt / custom instructions (GPT, Gemini, a local model, etc.).

Then restart your agent and describe what you want to film — e.g. "a basketball kid hits a buzzer-beater three."

🎬 Before & after

Real test cases from the original handbook — the same scene, a regular prompt vs. a Vibe Creating prompt, with the clip each one generated (videos from Volcengine's Seedance 2.0). See all eight in the worked test cases, and browse the example gallery for dozens more prompts.

Case 2 · near-VC input → light cleanup

▶ Press play to watch, with sound.

Case 3 · execution shot-script → direct rewrite