# Prompt Port Desk > A prompt written for Veo and the same prompt written for Seedance are not one text with > different labels. They are different shapes over different capabilities, so a port is > a lossy transformation rather than a rewrite, and moving between them costs > something. Five lanes over one prompt, plus a free browser-side transpiler that > emits the target grammar and computes the delta. URL: https://prompt-port-desk.skillsafe.ai/ API: https://prompt-port-desk.skillsafe.ai/api.html Derived from cclank/lanshu-awesome-ai-video-kit (https://github.com/cclank/lanshu-awesome-ai-video-kit), MIT: its prompt-translator label-grammar table, its cross-model comparison and its per-model formula sheets. Not affiliated with or endorsed by cclank, ByteDance, Kuaishou, Alibaba, OpenAI, Google, MiniMax, Tencent, Pika or Runway. Model capabilities change - check the vendor docs before committing a budget. ## What the free transpiler does It runs entirely in the browser, on the prompt you paste, with no account and no credits. 1. **Reads** the prompt into slots using the source model grammar. A labelled prompt gives them up directly; a prose one has exactly one slot, and saying so beats guessing which sentence was meant to be the camera line. 2. **Emits** it in the target grammar - its labels, in its order, in its form. Labels with nothing behind them come out empty rather than invented. 3. **Drops** what the target cannot render, and says so. A Dialogue line carried into a model that does not speak competes for attention with the picture description the model can act on - and it makes the loss report a liar. 4. **Computes the delta** between the two capability rows: seconds, native audio, multi-speaker dialogue, Chinese strength, image-to-video. Losses AND gains, because a move is not only a loss. 5. **Checks the shape** - the documented number-one failure is emitting prose for a model that requires labels. ## The eleven model grammars ### `seedance` - Seedance 2.0 (ByteDance) - Form: prose - Formula: eight elements in one prose paragraph: subject, action, scene, light, camera, style, audio, constraints - Words: 60-140. Max duration: 15s. - Audio: separate. Dialogue: no. Chinese: good. Image-to-video: yes. - Strongest at: complex multi-shot narrative - Anti-pattern: none - this one really is prose, so do not add labels to it ### `happyhorse` - HappyHorse 1.0 (Alibaba) - Form: prose - Formula: compact prose, 30-55 words; when the timing matters, open with `Ns duration.` - Words: 30-55. Max duration: 15s. - Audio: native. Dialogue: no. Chinese: good. Image-to-video: yes. - Strongest at: compact shorts with ambient sound - Anti-pattern: burying the duration in the middle instead of opening with it ### `kling` - Kling 3.0 (Kuaishou) - Form: labelled. Required labels: `Scene`, `Characters`, `Action`, `Camera`, `Audio & Style`, `Negative` - Formula: five labelled layers plus a negative - Words: 60-160. Max duration: 15s. - Audio: native. Dialogue: yes. Chinese: best. Image-to-video: best in class. - Strongest at: Chinese dialogue, physical feel, image-to-video - Anti-pattern: folding the five layers back into one paragraph of prose ### `sora` - Sora 2 (OpenAI) - Form: labelled. Required labels: `Style`, `Cinematography`, `Actions`, `Background Sound`, `Dialogue` - Formula: layered shot list; Actions is a list of beats - Words: 80-200. Max duration: 25s. - Audio: native. Dialogue: yes. Chinese: fair. Image-to-video: yes. - Strongest at: cinematic quality and extreme physics - Anti-pattern: merging the layers into a single description ### `veo` - Veo 3.1 (Google DeepMind) - Form: labelled. Required labels: `Subject`, `Action`, `Scene`, `Camera`, `Lighting`, `Style`, `Dialogue`, `Audio` - Formula: eight elements with Dialogue and Audio as explicit separate labels - Words: 70-180. Max duration: 8s. - Audio: native-multi. Dialogue: yes. Chinese: fair. Image-to-video: yes. - Strongest at: the strongest native audio, and multi-speaker dialogue - Anti-pattern: dropping the explicit Dialogue and Audio labels ### `wan` - Wan 2.7 (Alibaba) - Form: labelled. Required labels: `Entity`, `Scene`, `Motion`, `Sound` - Formula: four sections: Entity, Scene, Motion, Sound - Words: 40-120. Max duration: 15s. - Audio: native. Dialogue: yes. Chinese: good. Image-to-video: yes. - Strongest at: native audio and the most accurate digital-human lip-sync - Anti-pattern: folding the four sections into one ### `hailuo` - Hailuo 02 (MiniMax) - Form: prose - Formula: restrained prose, one to three sentences - Words: 15-60. Max duration: 10s. - Audio: separate. Dialogue: no. Chinese: good. Image-to-video: yes. - Strongest at: the best physical simulation - water, fire, smoke, cloth, gravity - Anti-pattern: piling on length; this model is directed by restraint ### `pika` - Pika 2.5 (Pika Labs) - Form: prose. Required labels: `Negative` - Formula: single-focus prose, and the Negative must carry `no morphing` - Words: 25-80. Max duration: 10s. - Audio: separate. Dialogue: no. Chinese: fair. Image-to-video: yes. - Strongest at: fifteen-plus creative effects and keyframe extension - Anti-pattern: stacking several subjects into one shot ### `runway` - Runway Gen-4.5 / Aleph (Runway) - Form: prose - Formula: prose, and for Aleph an editing verb rather than a static description - Words: 30-110. Max duration: 10s. - Audio: separate. Dialogue: no. Chinese: fair. Image-to-video: yes. - Strongest at: video editing with Aleph verbs - add, remove, change, re-light, re-style - Anti-pattern: describing a static scene where an editing verb belongs ### `hunyuan` - Hunyuan Video 1.5 (Tencent) - Form: prose - Formula: detailed prose in a single paragraph - Words: 60-180. Max duration: 10s. - Audio: separate. Dialogue: no. Chinese: good. Image-to-video: yes. - Strongest at: the strongest open-source model, with LoRA training - Anti-pattern: under-describing; this one rewards detail ### `jimeng` - Jimeng 3.0 (ByteDance CapCut) - Form: prose - Formula: the same eight-dimension formula as Seedance, written for Chinese - Words: 60-140. Max duration: 15s. - Audio: separate. Dialogue: no. Chinese: best. Image-to-video: yes. - Strongest at: the strongest Chinese, and CapCut integration - Anti-pattern: dropping dimensions from the eight ## The slots - **Subject** - who or what is in the shot. Read from: `subject`, `entity`, `characters`, `character`. - **Scene** - where, and in what conditions. Read from: `scene`, `setting`, `environment`, `background`. - **Motion** - what happens, in order. Read from: `action`, `actions`, `motion`, `movement`. - **Camera** - how it is shot. Read from: `camera`, `cinematography`, `shot`, `framing`. - **Lighting** - the source and its direction. Read from: `lighting`, `light`, `illumination`. - **Style** - the visual anchor. Read from: `style`, `audio & style`, `aesthetic`, `look`, `mood`. - **Audio** - what is heard. Read from: `audio`, `sound`, `background sound`, `audio & style`. - **Dialogue** - spoken lines, if any. Read from: `dialogue`, `speech`, `lines`. - **Negative** - what must not appear. Read from: `negative`, `constraints`, `avoid`. ## Worked deltas **Veo 3.1 -> Seedance 2.0** - loses native multi-speaker audio: Veo 3.1 generates audio with the picture, including multiple speakers. Seedance 2.0 does not - audio is a separate job there. - loses spoken dialogue: Veo 3.1 renders spoken lines; Seedance 2.0 does not. - gains: Seedance 2.0 runs to 15s, against 8s. - gains: Seedance 2.0's Chinese is good, against fair. - gains: Seedance 2.0 is strongest at complex multi-shot narrative. **Kling 3.0 -> Veo 3.1** - loses up to 7 seconds: Kling 3.0 runs to 15s and Veo 3.1 stops at 8s. - gains: Veo 3.1 generates audio with the picture, including multiple speakers. - gains: Veo 3.1 is strongest at the strongest native audio, and multi-speaker dialogue. **Sora 2 -> Hailuo 02** - loses up to 15 seconds: Sora 2 runs to 25s and Hailuo 02 stops at 10s. - loses native audio: Sora 2 generates audio with the picture. Hailuo 02 does not - audio is a separate job there. - loses spoken dialogue: Sora 2 renders spoken lines; Hailuo 02 does not. - gains: Hailuo 02's Chinese is good, against fair. - gains: Hailuo 02 is strongest at the best physical simulation - water, fire, smoke, cloth, gravity. **Seedance 2.0 -> Kling 3.0** - loses nothing, but the shape has to be invented: Seedance 2.0 is written as prose and Kling 3.0 needs 6 labelled sections. The content for each label has to be decided, and anything the prose left implicit will be guessed. - gains: Kling 3.0 generates audio with the picture. - gains: Kling 3.0 renders spoken lines. - gains: Kling 3.0's Chinese is best, against good. - gains: Kling 3.0 has the strongest image-to-video control. - gains: Kling 3.0 is strongest at Chinese dialogue, physical feel, image-to-video. ## The five lanes ### `read` - Pull it apart Strip the model's style off and keep the meaning: who, where, what happens, how it is shot, what is heard. A labelled prompt gives these up directly; a prose one has exactly one slot, and saying so beats guessing which sentence was the camera line. Sections: `Summary`, `Slots`, `What Is Implicit`, `Shape Note`, `Next Step`. Fields: `prompt`*, `from`*, `notes`. ### `port` - Port the prompt Re-emit it in the target's own grammar - its labels, its form, its word budget. Labels with nothing behind them come out empty rather than invented, because a transpiler can only move what was there. Sections: `Summary`, `Ported Prompt`, `Slot Map`, `Shape Check`, `Next Step`. Fields: `prompt`*, `from`*, `to`*, `seconds`, `keep`. ### `loss` - What does not survive Two capability rows, one delta. Seconds, native audio, multi-speaker dialogue, Chinese strength, image-to-video - and for each one, what to do instead. A move is not only a loss, so what the target does better is reported too. Sections: `Summary`, `Losses`, `Gains`, `Mitigations`, `Next Step`. Fields: `from`*, `to`*, `seconds`, `dialogue`, `audio`, `chinese`. ### `check` - Verify the shape The documented number-one failure when porting is quietly emitting prose for a model that requires labels. This checks the result is actually in the target's grammar, inside its word budget, and carrying whatever that model insists on. Sections: `Summary`, `Verdict`, `Findings`, `Repaired Prompt`, `Next Step`. Fields: `prompt`*, `to`*, `focus`. ### `batch` - Port to several at once The same prompt in three or four models' grammars, side by side, with each one's loss listed under it. That is what a cross-model A/B test actually needs, and doing it one model at a time hides the comparison. Sections: `Summary`, `Ported Set`, `Loss Table`, `Recommendation`, `Next Step`. Fields: `prompt`*, `from`*, `targets`*, `seconds`. ## What this app does not do It moves prompts between formats. It does not generate video, it calls no generation API on your behalf, and the capability rows it computes deltas from are a snapshot of published vendor behaviour rather than anything measured here.