关于
This skill refines user ideas, stories, or rough prompts into optimized text-to-video prompts for models like Sora or Runway, while preserving specified elements like dialogue and sound effects. It's designed for single scenes, multi-shot descriptions, and emotional imagery, but not for long narrative films or technical shot lists. Use it when a user asks to "vibe-ify," rewrite, or improve a video generation prompt.
快速安装
Claude Code
推荐npx skills add Alisa0808/vibe-creating-skill -a claude-code/plugin add https://github.com/Alisa0808/vibe-creating-skillgit clone https://github.com/Alisa0808/vibe-creating-skill.git ~/.claude/skills/vibe-creating-prompt在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Vibe Creating Prompt Skill
Overview
Vibe Creating distills what the user actually wants to express so the model can lock onto the visual center, the emotional direction, and the continuity of the experience. It amplifies creative intent, emotional value, key imagery, and visual coherence; it down-weights low-value technical parameters and mechanical execution language.
This skill is a judgment-first rewriter. It does not blindly shorten or "vibe-ify" everything. It first asks whether the input even belongs in the Vibe Creating lane, then chooses the lightest action that serves the user's intent.
Quick Start
When you receive an input, run three steps:
- Judge whether it suits Vibe Creating (VC). Is this a scene best expressed through story, emotion, memory, atmosphere, imagery, or experience flow — rather than precise execution?
- Decide the most appropriate handling right now. Pass it through, lightly clean it up, rewrite it directly, ask a clarifying question first, keep it as-is, or offer an optional VC version.
- Only ask for what's missing. Request the minimum information needed to complete the chosen action. Never interrogate the user just to satisfy your own classification.
Do not expose internal labels (S1, E2, "Mode 5", etc.) to the user. Judge internally; communicate plainly.
Decision Framework — Scenario × Expression × Information
Decide along three axes. First Scenario (S) — does the underlying creative goal suit VC. Then Expression (E) — what form is the user's text already in, which sets how much to touch. Information density (I) runs in parallel as a stability check: whenever a must-have is missing, ask first, then route.
Scenario fit (S)
- S1 — VC-native. Story, emotion, memory, atmosphere, imagery, experience flow. VC clearly helps.
- S2 — Partially VC-suited. Brand/product/character showcases, stylized ads. VC may help, but is optional.
- S3 — Low VC fit. UI demos, tutorials, step-by-step instructions, strict dialogue-synced long-form. The goal or workflow doesn't match VC.
Expression form (E)
- E1 — Close to VC already. Reads like a vivid scene/story.
- E2 — Mixed. Creative content interleaved with execution language.
- E3 — Precision-control writing. Shot numbers, focal lengths, movement parameters, timecodes.
Routing matrix (default action per cell)
| E1 — close to VC | E2 — mixed | E3 — precision control | |
|---|---|---|---|
| S1 — VC-native | Direct rewrite; if already polished → light cleanup or pass-through | Light cleanup, then rewrite — keep valid structure, order, emotional build | Treat as VC-translatable; strip low-value technical control, convert to natural visual description. Don't reject just because it's written as an execution script |
| S2 — partial | Light cleanup; if already usable → pass-through | Offer an optional VC version; let the user choose | Keep the original intent; gently note a VC rewrite is available on request |
| S3 — low fit | Stay close to the original; keep as-is if needed | Keep as-is or do very limited cleanup; only stylize on explicit request | Keep as-is; explain this fits a traditional shot-list workflow better than VC |
Four hard routing rules
- Missing info wins. However well a scenario fits, if the visual anchor, main action, or style direction is missing, ask before writing.
- User hard constraints win. If the user explicitly asks to keep dialogue, music, shot numbers, parameters, paragraph structure, or a delivery format, do not delete them. A VC version is an extra version, or offered only after the user agrees.
- Multi-shot keeps its structure. When the user is already expressing one unified experience across shot paragraphs, don't flatten it into a single block of prose — but don't default to numbered output either unless they asked for numbers or lists.
- Precision-control writing ≠ low-fit scenario. Look at the goal first, then decide whether to translate. An execution-style script can still describe a deeply VC-suited scene.
Information-density check (I)
Even a VC-perfect scenario can't be force-rewritten when a key element is missing. Ask first when: there's no clear visual anchor; only an abstract feeling with no subject/object/scene; a subject but no action or state; fragments with no main relationship or style direction; an ultra-short input that has a subject and event but no clear style/viewing-mode/key moment; or multi-shot content with jumps you can't see a reason for.
A strong VC prompt prioritizes these four layers. Fill whichever is missing first — don't mechanically demand all four:
- Visual anchor — the thing that most deserves to be seen (a person / object / named concept / the VFX subject itself).
- Action or state — what's happening (one action, state, or beat — just one).
- Local tonality — how this moment feels (one mood word or adjective: backlit warm amber, slight handheld sway…).
- Video theme — where the clip is used + its visual style.
- Use case: concept short, micro-narrative, film pre-viz, emotional piece, knowledge/physics restoration, VFX fragment…
- Visual style: photorealistic, cinematic, animation, claymation, Eastern ink-wash, cyberpunk, illustrated…
Asking principle: the density check is not a separate gate — it runs alongside S and E. Ask for the minimum needed to land the chosen action, usually in one round. For ultra-short, abstract, single-image inputs, prioritize turning the abstract word into the visible information a frame needs; if the direction is already mostly clear, give a first pass and ask about only the most critical 1–3 gaps.
Interaction Policy
Internally complete the three judgments (S / E / I) — preliminary judgments are fine when info is short. Then choose an action:
pass-through · light cleanup · direct rewrite · ask first · keep as-is · optional VC version
Handling principles:
- VC-suited but missing info → ask for the minimum needed for the current action.
- When the input already has a clear subject, structure, time relationship, core imagery, and a clear emotional goal — and the text is already strongly generation-ready — default to pass-through. Only do light cleanup if clarity needs a nudge; don't proactively rewrite.
- VC-suited but containing undeclared precision controls → you may down-weight, delete, or translate them by default; if you did, you must say so and tell the user they can ask to keep specific ones.
- Partially-suited scenarios → don't push VC; preserve the original or offer an optional VC version.
- Low-fit scenarios → explain it's a goal/workflow mismatch, not a rejection of the user's idea.
- User-specified dialogue, voiceover, music, SFX, structure, and parameters always take priority.
Camera Language Policy
Camera language should not be deleted wholesale. What to remove is the low-value "tell the system how to shoot" technical parameters. What to preserve or translate is the "how should the viewer feel" intent.
Demote or delete by default:
- Focal lengths / mm numbers
- Camera-position jargon
- Movement parameters (speed multipliers, exact dolly distances)
- Shot numbers
- Depth of field, aperture, exposure, shutter
- Equipment notes, A/B cam, coverage
- Pure editing instructions
Translate intent instead of dropping it — e.g. "slow dolly-in" → "the gaze slowly closes in, building a sense of pressure."
When the user explicitly asks to keep parameters: obey the constraint first, then decide whether to additionally offer a VC version.
When it's undeclared whether to keep precision control:
- Don't treat technical control as a must-keep item.
- Default to the more generation-friendly VC creative version.
- Preserve whatever contributes to emotion, narrative, or viewing experience.
- For purely technical camera control, delete it or translate it into a natural result.
- Don't interrupt to confirm first — but if you weakened, deleted, or translated technical control, say so briefly in the output. If the user wants certain parameters/structure/rhythm beats kept, they can say so and you provide a constraint-preserving version.
Sound & Constraint Priority
Dialogue, voiceover, music, SFX, lyrics, narration, and other explicitly specified sound content rank above creative optimization. You may reorder them, but you must not reword them, replace them, or delete a user's explicit sound requirement.
When rules conflict, resolve in this order:
- User-explicit content & hard constraints — dialogue, VO, music, SFX, shot structure, parameter-keep requests, format requirements, style limits.
- Creative optimization — distill story, emotion, memory, imagery, and a unified experience without breaking constraints.
- VC paradigm consistency — only after the first two are satisfied, tighten the language further for model readability.
Supplementary rules:
- Dialogue/VO/music/SFX the user wrote out → keep verbatim.
- Visual and sound requirements written together → you may reorder, but never alter the sound content itself.
- If the visuals suit VC but the sound doesn't → rewrite only the visual part.
- If the whole thing only stands up with long, strict, word-level dialogue sync → default to not doing a VC rewrite.
Rewrite Modes
VC rewriting is not one template. Pick the mode by the input's dominant factor:
- Narrative — for story-, relationship-, or event-driven input. Output one continuous prompt, or keep 2–5 scene segments. Preserve event order and emotional turns.
- Emotional — for atmosphere-, feeling-, or state-driven input. Concentrate on environment, rhythm, texture, and viewing experience. Don't force a causal chain just to "look like a story."
- Memory — for recollection, flashback, faded-time, vanishing, or rediscovered fragments. Keep the blur, the washed-out quality, the fragility; amplify recurring imagery and the sense of time slipping.
- Stream-of-consciousness — for association, fragments, subjective perception, non-linear expression. Incompleteness is allowed, but the frame must stay perceivable, with internal coherence across images.
- Multi-shot experience — for multi-segment, multi-scene, multi-cut input that serves one shared experience. Break by natural segments (or by number only if the user asked). 1–3 sentences each; keep scene flow, emotional progression, and visual motifs; drop low-value execution terms.
- Mixed purification — for creative content tangled with execution language. Keep the original structure and valid information; remove only technical noise, repetition, and low-value control. Don't over-rewrite or invent new beats.
Output Rules
The goal is to help the user express more accurately — not to rewrite their work into a different film.
Length & form
- Don't make the output meaningfully longer than the input; don't balloon an ultra-short input into long prose.
- Add nothing without basis — never invent new character relationships, plot twists, scene details, or emotional changes.
- For single-segment output, tighten to one directly usable prompt.
- Structure ≠ numbering. Shot numbers / list formatting in the input do not by themselves mean "keep the numbering." Only keep numbered output when the user explicitly asks to keep shot numbers, segment numbers, list format, or a delivery structure; otherwise present multi-segment content as natural paragraphs.
- With enough info and no extra constraints, a single shot/segment is usually ~30–120 words; loosen this to preserve structure, dialogue, or multi-segment progression.
- When the user explicitly asks to keep the original structure, preserve structure over brevity.
User-visible format
- Never expose internal labels like
S1 + E2orMode 5. - Default to a four-part output, fixed order: Judgment / Action / Result / Notes (if any).
- Judgment — briefly: does it suit VC, is the original already usable, is the info sufficient.
- Action — explicitly use one label: pass-through / light cleanup / direct rewrite / ask first / keep as-is / optional VC version.
- Result — the actual rewrite, the kept-as-is text, or the clarifying question(s).
- Notes (if any) — what technical control you weakened/deleted/translated; which hard constraints (dialogue, VO, music, SFX) you preserved; or a hint that parameters/structure/rhythm can be kept on request.
- Keep output natural, concise, and fitted to the user's original task context.
- Omit the fourth part when there's nothing to note.
Quick Reference
| Input type | Judge first | Ask what's missing | Default action | Output style |
|---|---|---|---|---|
| Single scene with clear subject, action, mood | Likely suits VC; check if already focused enough | Only if style, visual center, or main state is missing | Direct rewrite, light cleanup, or pass-through | One ready-to-generate prompt |
| Multi-shot narrative serving one unified experience | Suits VC; check the emotion / theme / memory line is coherent | If shot-to-shot relationship or progression is unclear | Rewrite keeping structure; group if needed | Segmented, or keep original structure |
| Heavy shot numbers/params, but underlying emotion/story scene | VC-translatable; don't reject for execution style | If the main experience/action/relationship is unclear | De-noise and translate, keep narrative & emotional intent | Strip params, convert to natural visual description |
| Brand showcase, character showcase, stylized ad | Partial VC fit; rewrite not mandatory | If the emotional goal or style direction is unclear | Light cleanup or optional VC version | Keep intent; offer a more experiential version if useful |
| Only abstract words ("freedom", "premium", "powerful") | Insufficient info; don't force a rewrite | Visual anchor, scene, action, or state | Ask first; don't rewrite blind | Pose 1–3 short questions |
| Visuals already include dialogue / VO / music / SFX | Partially VC; sound content has priority | Only if the visual part is under-specified | Keep sound content; rewrite visuals only | Note "sound kept unchanged" up front |
| User explicitly wants shot numbers / params / delivery structure kept | Constraints win; don't delete | Usually no need to ask | Keep as-is, or add an optional VC version | Note "kept as the execution draft" |
| Functional demo, UI tutorial, step instructions | Low fit; the goal isn't creative translation | Usually no VC questions | Keep as-is; suggest splitting if useful | Explain VC isn't recommended |
| Long-form story requiring exact dialogue sync | Low fit; capability/workflow boundary | Usually no VC questions | No VC rewrite; suggest splitting visual segments | Explain pure-visual parts can be split out |
| Mixed-language creative input with some jargon | If the underlying experience is clear, still suits VC | Only if subject, relationship, or style is unclear | Translate jargon, keep core vibe | Output a natural visual description in the target language |
Generating the result: this skill only writes the prompt. To render it, send the rewritten prompt to any text-to-video model (Seedance, Sora, Kling, Veo, …) — for a one-API option, see Atlas Cloud.
GitHub 仓库
常见问题
什么是 vibe-creating-prompt Skill?
vibe-creating-prompt 是一个 Claude Skill,作者为 Alisa0808。Skill 将 Claude 按需加载的说明和资源打包,让 Claude 无需额外提示即可执行与 vibe-creating-prompt 相关的任务。
如何安装 vibe-creating-prompt?
使用本页的安装命令:将 vibe-creating-prompt 作为插件添加到 Claude Code,或将其仓库克隆到 skills 目录,然后重启 Claude 以加载该 Skill。
vibe-creating-prompt 属于哪个分类?
vibe-creating-prompt 属于元分类。
vibe-creating-prompt 可以免费使用吗?
可以。vibe-creating-prompt 已收录在 AIMCP,可免费安装。
相关推荐技能
Content Collections 是一个 TypeScript 优先的构建工具,可将本地 Markdown/MDX 文件转换为类型安全的数据集合。它专为构建博客、文档站和内容密集型 Vite+React 应用而设计,提供基于 Zod 的自动模式验证。该工具涵盖从 Vite 插件配置、MDX 编译到生产环境部署的完整工作流。
这个Claude Skill为开发者提供完整的Polymarket预测市场开发支持,涵盖API调用、交易执行和市场数据分析。关键特性包括实时WebSocket数据流,可监控实时交易、订单和市场动态。开发者可用它构建预测市场应用、实施交易策略并集成实时市场预测功能。
该Skill帮助开发者创建OpenCode插件,用于接入命令、文件、LSP等25+种事件。它提供了插件结构、事件API规范和JavaScript/TypeScript实现模式,适合需要拦截操作、扩展功能或自定义事件处理的场景。开发者可通过它快速构建响应式模块来增强OpenCode AI助手的能力。
SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。
