Vibe Creating — AI Video Prompt Skill
Rewrites rough ideas, stories, or over-specified shot scripts into model-ready text-to-video prompts — and first judges whether your input even suits the style.
The SKILL.md requests no permissions, has no external side effects or data flows, and mandates preserving user hard constraints with disclosure when technical parameters are weakened — good safety posture. Deductions: repo license metadata is NOASSERTION, publisher unverified, disclosed commercial affiliation of maintainers with Atlas Cloud, and paradigm provenance rests on statically unverifiable Lark doc links.
Instructions are self-consistent: routing matrix, quick reference and four-part output format align; as a pure prompt skill there are no runtime dependencies, and low-fit inputs have clear handling. Deductions: no test cases or reproducible boundary-input validation; rewrite quality depends wholly on the host model, and static review cannot confirm key-path reproduction (capped at 10).
Fully bilingual EN/ZH docs; trigger conditions, fit boundaries (S1–S3) and non-fit cases (UI demos, word-for-word dialogue sync) are clearly stated, supporting precise semantic invocation. Deductions: no versioning/changelog on how boundaries evolve, example videos are GitHub-hosted attachments, and the full value chain depends on overseas video models / Atlas Cloud API, limiting mainland-China reachability even though local rewriting works offline.
Well-layered docs (SKILL.md, README, FAQ, NOTICE), clear MIT license with NOTICE separating paradigm attribution from code license, honest non-official ByteDance disclaimer. Deductions: metadata license is NOASSERTION, no version number/changelog/maintenance ownership or update path, and the npx installer is unverified in static scope.
The core task — turning rough ideas into directly usable video prompts — is fully specified with representative before/after cases and a fixed output format; marginal value is clear. Deductions: static review cannot execute; actual rewrite quality rests on author-submitted videos only (capped at 7).
README offers multiple concrete before/after cases, comparison videos and attribution — more auditable than pure marketing. Deductions: videos are self-generated and not independently reproducible; Lark source docs cannot be statically verified; no CI/tests or third-party execution evidence (capped at 5; thin coverage → 4).
- Repo license metadata is NOASSERTION, inconsistent with the in-file MIT declaration; verify licensing metadata.
- Generation depends on overseas video models and the Atlas Cloud API (requires reachability and a paid API key); mainland-China users can rewrite locally but the full chain may be limited.
- Maintainers have a disclosed commercial affiliation with Atlas Cloud; treat API recommendations with that in mind.
- No automated tests back rewrite quality; results vary by host model — trial with small samples first.
- Attribution to ByteDance's original methodology rests on Lark doc links that cannot be statically verified.
What does this skill do, and when should you use it?
Vibe Creating is an open-source, bilingual prompt-engineering skill that turns a rough idea, story, feeling, or over-specified shot script into a clean, model-friendly text-to-video prompt. It ports the methodology ByteDance/Volcengine introduced alongside Seedance 2.0 into a reusable Agent Skill. Unlike tools that blindly shorten prompts, it is judgment-first: it scores input along Scenario × Expression × Information density, then picks the lightest action — pass-through, light cleanup, direct rewrite, or asking first. It strips low-value technical parameters, translates shot intent into natural visual description, and always preserves user-specified dialogue, voiceover, music, and SFX verbatim. The skill ships as a single standards-compliant SKILL.md that runs in Claude Code, Codex, OpenClaw, Hermes, or as a rule/system prompt in any LLM, and its output works with any text-to-video model including Seedance 2.0, Sora, Kling, and Veo.
Reads the user's idea, scene, or shot-script input; internally scores it on three axes (Scenario / Expression / Information density) to decide fit and information sufficiency; asks 1–3 clarifying questions when a key element is missing; on rewrite, demotes focal lengths, apertures, shot numbers and similar parameters while translating camera intent into natural visual results; keeps user-written dialogue, VO, music, and SFX verbatim; preserves multi-shot structure as natural paragraphs rather than defaulting to numbered lists; returns one generation-ready prompt in a fixed four-part format (Judgment / Action / Result / Notes). The skill only writes prompts — it does not call any video-generation API.
- A short-video creator has an abstract concept like 'freedom' or 'premium' and needs the visual anchor and scene clarified before anything can be generated
- A filmmaker has a precise execution-style shot list with timecodes and lens data, but the underlying scene is an emotional arc they want preserved without the parameter noise
- A brand or product marketer wants an optional atmosphere-driven version of an ad prompt without overwriting the original
- A user whose prompt already contains dialogue, music, and SFX wants only the visual part rewritten, sound content untouched
- Anyone using Seedance, Sora, Kling, or Veo who wants their agent to handle 'rewrite the prompt, then generate the video' end-to-end
What are this skill's strengths and limitations?
- Judgment-first: inputs that don't suit the style (UI demos, tutorials, dialogue-synced long-form) are told so honestly rather than force-rewritten
- Hard constraints win: user-specified dialogue, VO, music, and SFX are kept verbatim, and any parameter weakening must be disclosed in Notes
- Bilingual docs plus a Chinese skill edition, with mixed-language input handled
- Single file, zero runtime dependencies, follows the open Agent Skills standard — portable across agents and plain LLMs
- Includes real test cases with generated video comparisons (8 cases plus an example gallery), so the effect is verifiable
- Writes prompts only — video generation requires a separate API call (e.g. Atlas Cloud), adding cost and setup
- GitHub license field reads NOASSERTION; README states MIT with a NOTICE file, but licensing status deserves verification before reuse
- Worked examples with generated clips are Seedance 2.0 only; no systematic evidence of testing on Sora, Kling, Veo, etc.
- Not an official ByteDance project — an independent port that may drift from the source methodology
- No automated test suite; quality depends on how well each model interprets natural-language prompts
How do you install this skill?
Option 1 (recommended): run npx github:Alisa0808/vibe-creating-skill in a terminal — the installer auto-detects installed agents and copies the skill; add --all for every known agent, name agents explicitly (… claude codex), or target a folder with --dir <path>. Option 2: git clone https://github.com/Alisa0808/vibe-creating-skill.git, then copy skills/vibe-creating-prompt into your agent's skills directory — Claude Code: ~/.claude/skills/, Codex CLI: ~/.codex/skills/, OpenClaw: ~/.openclaw/skills/, Hermes: ~/.hermes/skills/. Option 3 (Cursor or any LLM): paste the body of SKILL.md (or its Chinese edition SKILL.zh.md) as a rule or system prompt.
How do you use this skill?
Restart your agent and describe what you want to film — e.g. 'a basketball kid hits a buzzer-beater three.' The skill judges the input, asks clarifying questions if needed, and returns a rewritten prompt in the four-part Judgment/Action/Result/Notes format. To render the result, send the prompt to any text-to-video model. The quickest path is Atlas Cloud's unified API (Seedance 2.0, Kling, Veo, Hailuo, Wan, Vidu, Gemini Omni, Grok Imagine), or install the official skill (npx skills add AtlasCloudAI/atlas-cloud-skills) or MCP server (claude mcp add atlascloud -- npx -y atlascloud-mcp) to have the agent rewrite and generate in one flow.
How does this skill compare with similar options?
Versus simply writing a longer, more detailed prompt, the methodology targets 'the right information' rather than length: strip ineffective technical noise, keep story, emotion, and key imagery. Versus tool-specific rule files (e.g. Cursor rules), it follows the open Agent Skills (SKILL.md) standard, so Claude Code, Codex, OpenClaw, and Hermes load it natively.