SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsprompt-translator

prompt-translator

Agents & models

View repository

Updated 10-15 17:32

About this skill

prompt-translator

Cross-model prompt converter.Key difference: rather than rewriting based on AI intuition,consult 110 comparison benchmarks for in-context learning。

When not to use this skill

  • The user is writing a new prompt from scratch (not converting) → use seedance-prompter / kling-prompter / happyhorse-prompter or model-selector
  • The user asks "which model is better" → use model-selector
  • An existing prompt has problems but the model is not being changed → use seedance-debugger

Core data asset

prompts/data/cross-model-matrix.json — this is the translator's "training data":

  • 10 core scenarios: product ad / emotional reunion / skateboarding action / image-to-video / multi-person meeting / horror suspense / nature time-lapse / abstract art / wuxia duel / viral cute pet
  • Each scenario × 11 models = 110 comparison prompts, each strictly following the corresponding model's official formula
  • This forms a "comparison of the best writing for the same scenario across 11 models,"which is the translator's lookup table

Workflow

Step 1: Receive input

At minimum, you need:

  • Source model (such as Sora 2 / Kling 3.0 / Wan 2.7)
  • Source prompt (the user's existing prompt)
  • Target model (which one the user wants to switch to)

Optional:

  • Conversion preference (more concise / more detailed / keep Chinese)

Step 2: Analyze the semantic content of the source prompt

Extractcore scene elements(unrelated to the specific writing style):

  • Subject: who/what
  • Scene: where/what environment
  • Motion: what temporal events happen
  • Mood: overall atmosphere
  • Camera: how it is shot
  • Audio: what sound is needed
  • Dialogue: whether there is speech
  • Style: visual anchor

This step isstrip away style, extract semantics. Take the actual content in Sora's Style: → Cinematography: → Actions: layered structure and abstract it into a "core scene description."

Step 3: Consult the 110-item benchmark comparison table to find the most similar scenario

Read prompts/data/cross-model-matrix.json, and among the 10 scenarios findthe 1-2 scenarios most similar to the user input:

User input looks like... Reference scenario
Product rotation / static subject close-up scene-1-perfume
Two-person dialogue / emotional narrative scene-2-reunion
Outdoor action / physical movement scene-3-kickflip
Image-to-video (with reference image) scene-4-i2v-cafe
Multi-person conversation / indoor meeting scene-5-meeting
Horror suspense / slow push-in atmosphere scene-6-horror-balloon
Nature time-lapse / unpopulated landscape scene-7-mountain-sunrise
Abstract art / fluid effects scene-8-liquid-metal
Wuxia / Chinese-style fight scene-9-wuxia-duel
Cute pet / viral short video scene-10-surfing-dog

Step 4: Convert based on the comparison pattern of the similar scenario

Give Claude a few-shot template like this in the prompt:

我要把这条 [源模型] 的 prompt 转换成 [目标模型] 的最佳写法。

【参考对照】下面是一个相似场景在两个模型上的对照写法:

[源模型 in scene-N]:
{基准数据中该场景在源模型上的 prompt}

[目标模型 in scene-N]:
{基准数据中该场景在目标模型上的 prompt}

注意观察:
- 字段标签的变化 (e.g. "Cinematography:" → "Camera:" → "镜头:")
- 段落结构的变化 (e.g. 分层 → 5 层 → Entity+Scene+Motion+Sound)
- 措辞密度的变化 (e.g. 100 词 → 30 词 → 中文短句)
- 音频处理的变化 (e.g. "Background Sound:" → "Audio:" → "Sound:")

【用户的源 prompt】
{源 prompt}

【请输出】基于上面对照模式,将用户 prompt 转换为目标模型最佳写法。保留所有语义内容,仅调整结构/标签/措辞。

Step 5: Output format

⚠️ Hard constraints on the target format (check every time)

Regardless of whether the source prompt is prose or tagged, the target prompt must strictly follow the target model's cross-model-matrix.json tag format. Do not output prose-ified.

Target model Must have tags / structure Counterexamples (LLMs easily make)
Kling 3.0 Scene: / Characters: / Action: / Camera: / Audio & Style: / Negative: ❌ A block of prose cramming 5 layers into natural language
Sora 2 Style: / Cinematography: / Actions:(- beats)/ Background Sound: / Dialogue: ❌ Merged into a single paragraph description
Veo 3.1 8 elements + Dialogue: / Audio: dual markers ❌ Missing explicit Dialogue/Audio tags
Wan 2.7 Entity: / Scene: / Motion: / Sound: four segments ❌ Squeezed into one paragraph
Seedance 2.0 8-element prose single paragraph (subject + action + scene + light + camera + style + audio + constraints) ✓ This one is prose
HappyHorse 1.0 Compact prose 30-55 words; when timing is explicit start with Ns duration. start with ❌ Burying the duration in the middle
Hailuo 02 Restrained concise prose, 1-3 sentences ❌ Overly long and piled up
Pika 2.5 Single-focus prose,Negative: must include no morphing ❌ Multiple subjects stacked
Runway Gen-4.5 prose + Aleph editing verbs (recolor / add / remove) ❌ Describing a static state instead of an editing action
Hunyuan / LTX / Mochi / CogVideoX Detailed prose, single paragraph —
Jimeng 8-dimension formula (same as Seedance) ❌ Missing dimensions

Self-check list(must pass before output):

  1. Does the target prompt have the tags/structure corresponding to the table above?
  2. database cross-model-matrix.json what does the prompt for this model and this scene look like? Is my output format consistent with it?
  3. If inconsistent,rewrite, don't settle.

Output template

## 转换结果

**源模型**: [Sora 2] · **目标模型**: [Kling 3.0]
**参考场景**: [scene-N-xxx · 用户输入最像哪个对照场景]

\`\`\`
[转换后的目标模型 prompt — 必须带上表对应标签]
\`\`\`

## 转换映射(供检查)

| 源结构 | 目标结构 | 注释 |
|---|---|---|
| Style: ... | Scene: + Audio & Style: | Sora 单一 Style 段拆成 Kling 的环境+风格混合 |
| Cinematography: ... | Camera: | 字段重命名 |
| Actions: - beat - beat | Action: 流畅描述 | 按 beats 列表合并成自然语言 |
| Background Sound: ... | Audio & Style: | 合并进风格段 |
| (无角色对白栏) | (保留对白嵌入 Action) | Kling 5 层不单独分对白 |

## 注意事项

- ⚠️ [目标模型] 与 [源模型] 在 [某能力] 上有差异:[具体说明,例如 "Kling 中文比 Sora 强,可考虑改成中文版本"]
- 💡 推荐参数:[根据目标模型给出宽高比/时长/分辨率建议]
- 🔄 如果效果不理想,试试:[备选转换方向,如再转 Veo 用多人对话能力]

Practical examples

Example 1: The user gives a Sora 2 prompt and wants to convert it to Kling 3.0

Input:

Source model: Sora 2 Source prompt:

Style: 1970s romantic drama, 35mm film with warm halation, slight handheld micro-shake.
A park bench at golden hour in late autumn. Two men in their early 30s.
Cinematography: Medium-wide, slow dolly-in. 40mm lens, shallow focus.
Actions:
- A sits down hesitantly
- A (low, hesitant): "It's been a long time."
- B turns, eyes widening
- B (warm): "Too long. I almost didn't recognize you."
- Both share a quiet laugh
Background Sound: Soft piano, ambient park, no music swell.

Target model: Kling 3.0

Steps 2-3: Identify that this is the scene-2-reunion (emotional dialogue) scene.

Step 4: Read the sora-2 and kling-3.0 comparison for scene-2-reunion in the baseline data, and extract the conversion pattern.

Output:

## 转换结果

**源模型**: Sora 2 · **目标模型**: Kling 3.0
**参考场景**: scene-2-reunion(情感双人对话)

\`\`\`
Scene: A park bench at golden hour in late autumn, amber sunlight filtering through bare branches, fall leaves drifting through the frame.
Characters: Character A (early 30s, hesitant) and Character B (early 30s, expressive).
Action: A walks up and sits down beside B with visible hesitation. A says: "It's been a long time." B turns slowly, eyes widening with recognition, replies warmly: "Too long. I almost didn't recognize you." They share a quiet laugh that breaks the tension.
Camera: Slow push-in on A as he speaks, over-the-shoulder cut to B's reaction, ending on a held wide two-shot.
Audio & Style: Soft solo piano underscore building gently, ambient park sounds (distant kids playing, leaves rustling), warm cinematic teal-and-amber grade, 1970s romantic drama with 35mm halation aesthetic, naturalistic dialogue delivery.
\`\`\`

## 转换映射

| Sora 2 | Kling 3.0 | 注释 |
|---|---|---|
| Style: ... | (合并到 Audio & Style:) | Kling 5 层把 Style 嵌在最后一层 |
| Cinematography: 段 | Camera: | 字段重命名,Kling 更简洁 |
| Actions: beats + dialogue 嵌入 | Action: 自然语言 + 嵌入对白 | Kling 5 层 Action 段直接含对白,不像 Sora 用 beats 列表 |
| Background Sound: | (合并到 Audio & Style:) | Kling 把音频与视觉风格放一段 |
| (无 Characters: 段) | Characters: 显式列角色 | Kling 5 层强制要求 Characters: 段 |

## 注意事项

- ✅ Kling 5 层结构能很好承接 Sora 的所有内容
- 💡 推荐参数:16:9 / 10s / 1080p
- 🌐 Kling 中文理解优于 Sora,如果场景偏中式可把对白和场景描述改成中文(Kling 中文是业界并列第一)
- 🔄 如果想要更强的音画对白同步,再转 Veo 3.1(原生音频最强)

Example 2: The user's Wan 2.7 prompt converted to Hailuo 02

(Same table lookup process, refer to the wan-2.5 ↔ hailuo-02 comparison row for scene-5-meeting or the most similar scene)

Example 3: The user's Kling Chinese prompt converted back to a concise English Pika prompt

(Refer to the kling-3.0 ↔ pika-2.5 comparison row for scene-1-perfume or the most similar scene,emphasize Pika's "single focus / no morphing" constraints)

Limitations — must be honestly disclosed

⚠️ This skill is a table-lookup conversion based on 110 formula comparison data entries, not end-to-end optimization based on actual tested video results。

Cannot do:

  • Cannot guarantee the actual video quality generated by the target model from the converted prompt
  • Cannot predict the target model's special quirks (models iterate frequently)
  • Cannot do the final A/B testing for you — that must actually be run

Can do:

  • Reorganize the source prompt into a structure that conforms to the target model's formula
  • Preserve all semantic content (subject/scene/action/audio/style)
  • Point out key field mappings and style differences
  • Give recommended parameters and directions for subsequent optimization

Best Practices: Use the conversion results as a "draft" and fine-tune after testing on the target platform. If you find that a certain conversion works consistently well, you are welcome to submit a PR to add it to cross-model-matrix.json as a reference sample for a new scenario (see CONTRIBUTING.md)。

Roadmap (to be done once the data is more solid)

Version Enhancement
v1 (current) 110 reference lookup entries + 5-step process
v2 Add 5 scenarios to expand to 165 reference entries
v3 Add a tested video link for each reference entry (contributed by the community)
v4 Add automated A/B evaluation (driven by ELO ratings after testing the same prompt across multiple models)

Resources