vibe-creating-skill

Purpose

Vibe Creating is an open-source prompt-writing paradigm and skill for AI video generation. It rewrites a rough idea, story, feeling, or over-specified shot script into a clean, model-friendly text-to-video prompt. Its core idea is: as text-to-video models get smarter, prompts should become simpler—no need to over-specify focal length, shot numbers, and frame-by-frame scripts; instead, focus on telling the story and let the model find the appropriate shots, lighting, and pacing on its own.

Main Capabilities

• Judgment first: Before rewriting, first judge whether the input is suitable for the Vibe Creating style. It scores along three dimensions—scene × expression × information—and chooses the lightest action, avoiding flattening the precise shot list the user actually wants.
• Four-layer information check: visual anchor (subject), action/state (what is happening), local tone (one emotional word), video theme (purpose + visual style). The skill automatically asks about missing layers.
• Fixed four-part reply format:
  1. Judgment—whether it is suitable for Vibe Creating, whether it is already directly usable, whether the information is sufficient;
  2. Action—one of direct pass-through / light cleanup / direct rewrite / ask first / keep as is / optional VC version;
  3. Result—the rewritten prompt, the retained original text, or a clarifying question;
  4. Notes (if any)—which technical controls were stripped away and which constraints were retained.
• Model-agnostic output: The output is pure natural-language description, not model-specific syntax, and applies to any text-to-video model.
• Bilingual: English and Chinese versions are provided.

How to Use

This skill is a portable SKILL.md folder, and the installation method can be chosen according to the environment:

• Method one (recommended) one-line command: automatically detects installed agents and copies the skill to their respective directories. You can add --all to cover all known agents, or explicitly specify agent names (such as … claude codex), or use --dir <path> Specify any folder.
• Method 2: Clone and copy: The same file, placed in different locations for different agents.
• Method 3: Cursor or any LLM: Paste the body of the skill file (or its Chinese version) as rules (Cursor), or as a system prompt/custom instructions (GPT, Gemini, local models, etc.).

After installation, restart the agent and describe what you want to shoot, for example, "a child playing basketball hitting a buzzer-beating three-pointer."

Use cases

• Writing video prompts for atmosphere, emotion, narrative, and visual exploration.
• Use with any text-to-video model; the original methodology was distilled from Seedance 2.0, and the same prompts also apply to Sora, Kling, Veo, Runway, Pika, Hailuo, etc.
• Can run in any agent that supports the open Agent Skills (SKILL.md) standard: Claude Code, Codex, OpenClaw, Hermes, etc.
• Send the rewritten prompt to any text-to-video model to generate a clip; you can also have the agent do it end-to-end: let it rewrite first and then generate.

Limitations

• Not suitable for scenarios requiring precise word-for-word dialogue synchronization, strict shot-by-shot execution, UI demonstrations, or step-by-step tutorials; traditional precise prompts are more suitable for these scenarios—the skill itself will also indicate this, rather than forcing a rewrite.
• It is not a tool for "making everything shorter," but rather for providing "the right information": removing ineffective technical noise while preserving story, emotion, and key imagery.
• More details about pricing, fees, platforms, and specific effects are not yet provided.

Source and license

The Vibe Creating paradigm, the original skill draft, and the example prompts originate from ByteDance/Volcano Engine (created using Seedance 2.0). This repository is an independent, faithful Chinese-English bilingual port of this publicly shared methodology, and is not an official product. The repository code and documentation are released under the MIT License; the underlying paradigm and related trademarks belong to their original owners. See the NOTICE file for more attribution information.