A useful Seedance 2.5 prompt is not a collection of magic words. It is an executable, reviewable shot brief: what is in frame, what action happens, how the camera moves, which media must serve as references, and where the clip should end.
This guide is for creators and teams using the current Seedance 2.5 web workspace. It is a workflow guide, not an independent benchmark, API reference, licensing guide, or commercial-rights statement. It is based on traceable public creator posts and the current Seedance 2.5 video workspace request contract. To avoid presenting inference as fact, the article uses three labels:
- Public case facts: only what an author or official source publicly stated, linked to the original post;
- Editorial interpretation: a reusable workflow lesson drawn from the case;
- Practice templates: new prompts written for this guide. They are not an author's original prompt and do not promise to reproduce a particular work.
Quick answer: write a shot brief, then run a controlled iteration
For each run, name one primary action, its timing, one camera behavior, the reference role of each uploaded file, and the required end state. Keep the first attempt deliberately narrow. After reviewing it, change one high-leverage variable, such as action direction or camera path, instead of rewriting everything at once.
For the mode, duration, frame, file, and resolution constraints that determine whether a request is valid, start with the Seedance 2.5 Guide. This article focuses on turning a valid request into a more testable creative brief.
The short answer: real cases look more like workflows than magic phrases
The public cases reviewed here do not provide a collection of complete, reproducible original Seedance 2.5 prompts. Instead, they show three more reliable things: creators place the model inside multi-tool workflows, plan motion as its own task, and still generate and select from many attempts.
That means a prompt guide should not promise that copying one line will recreate someone else's clip. Its job is to help you turn one generation into a controlled, iterative shot task.
Public sources and cases: what is known and what is not
| Case | What the source explicitly states | What this guide does not infer |
|---|---|---|
| Lansen's wuxia duel | The creator says the work was made with Midjourney and Seedance 2.5, and says the prompt is in a comment. | The full comment prompt, settings, reference images, seed value, and generation configuration are not retained in this research record. |
| Yu Shimura / GOKKO production note | The creator says 991 cuts were made, 152 appeared on screen, and 2,288 takes were produced, with 160 takes selected; to improve the selection rate, they tried making the motion first and then generating. The base video model is identified as Seedance 2.5. | The prompt, reference media, failure reasons, and model settings for each take are not public in the source material. |
| Xanderwow's music video | The creator attributes lyrics for “Safety & First Place” to GPT-6-Astra, music to Suno V6, and video to Seedance 2.5. | The video prompt, shot list, whether audio was passed to the model, and the post-production workflow are not disclosed. |
| ByteDance Seed's announcement | The official announcement describes Seedance 2.5 in terms of long-form storytelling, multimodal reference, and editing. | It is not a public prompt library and does not provide reproducible settings for a specific case. |
These are public post links, and their authors may edit, delete, or restrict them later. This guide does not treat a screenshot, output clip, or short endorsement as proof of model quality, speed, or success rate.
Prompt architecture: four layers you can diagnose
Use the same four layers in every first draft. They make a disappointing output easier to diagnose without claiming that any particular wording guarantees a result.
| Layer | Decision to make | Example of a reviewable instruction |
|---|---|---|
| Shot intent | What single event must happen? | “The dancer turns once to the right.” |
| Time and camera | When does it happen, and what does the camera do? | “The turn begins after the first beat while the camera tracks laterally.” |
| Reference role | What is each uploaded file allowed to control? | “@image1 defines the product shape; @audio1 supplies the rhythm cue.” |
| Boundary and ending | What must not be introduced, and where does the shot stop? | “No text or additional people; end on the completed turn.” |
This is an editorial writing framework, not a hidden model syntax. Its value is that each layer can be reviewed separately before you spend credits on another run.
Case one: turn a wuxia scene into a shot task
Lansen attributes an imagined wuxia duel to a Midjourney and Seedance 2.5 combination, and directs readers to the comments for the prompt. Since the full comment prompt is not part of the verifiable source material here, the responsible approach is not to reconstruct an “original prompt.” Instead, identify the shot variables this kind of scene needs.
From the published theme and image, an editorial visual reading of the task can be separated into the following variables. These are not details publicly stated by the creator:
- Subject relationship: two opponents, their distance, orientation, and weapons;
- Primary action: a preparation, crossing, dodge, or recovery, with one main action chain;
- Spatial cues: bridge, architecture, distance, or weather that support the direction of action;
- Camera path: lock-off, pan, low-angle push, or tracking, with one dominant movement;
- End state: a standoff, contact, or departure.
Practice template, not the source prompt:
Two wuxia fighters face each other from opposite ends of a stone bridge. The foreground fighter steps forward and swings a long blade once; the distant fighter turns sideways to evade, completing only one crossing action. A low-angle camera makes a slow push along the bridge, with railings passing at both edges. Keep distant buildings and storm clouds as background layers. Do not add a third character or text. End with the two fighters facing off after they pass each other.
The value of this template is that it reduces a large spectacle to one inspectable action. It does not claim to come from Lansen or to reproduce the source clip.
Case two: turn GOKKO's “motion first, then generate” into a prompt workflow
Yu Shimura / GOKKO's public post gives a specific production approach: make the movement first, then generate, as part of an effort to improve selection rate. The post says that 991 cuts were made, 152 appeared on screen, and 2,288 takes were produced, with 160 takes selected. It also describes a large cut and take selection process. It does not prove that one writing style will always improve results, but it does support a more careful operating principle: write the action before adding style and decoration.
Write a prompt in this order:
- Action sentence: What state does the subject move from, and what state does it move to?
- Timing sentence: Does the change happen in the first half, second half, or final two seconds?
- Camera sentence: Is the camera static or moving? What are its direction and speed?
- Image sentence: How do setting, light, wardrobe, and materials support the action without taking focus from it?
- Boundary sentence: Explicitly exclude unneeded people, text, extra actions, or cuts.
Start with a draft that validates only the action:
A person begins standing still, raises their right hand to point toward the left side of frame, pauses, then lowers the arm. Medium static shot. Complete the action within the first six seconds and keep the person centered throughout.
After the action direction is correct, add scene and visual information:
A person begins standing still, raises their right hand to point toward the left side of frame, pauses, then lowers the arm. Medium static shot. Complete the action within the first six seconds and keep the person centered throughout. A train platform at dusk, warm back-side light, and a dark long coat; do not cut away or add other people or subtitles.
Both examples are practice templates. They take inspiration from the public “motion first, then generate” workflow, not from GOKKO's original text or project assets.
Case three: a music video is not one universal prompt
Xanderwow's public post assigns the lyrics, music, and video for “Safety & First Place” to GPT-6-Astra, Suno V6, and Seedance 2.5 respectively. The useful lesson is not a hidden prompt; it is task boundary. Writing, music, and video can happen in different tools, so one video-generation prompt does not need to carry every creative job.
When music or a rhythm reference already exists, decide its job in the video task:
- Is it a reference for editing pace or for on-screen movement?
- Does the image need to follow a distinct beat, or only preserve an overall feeling?
- Are you generating one shot, or gathering multiple editable shots for a music video?
The current workspace supports audio reference, with up to 10 clips and 30 seconds in total per task. When using audio, state what @audio1 should do instead of merely writing “match the music.”
Practice template, not the music video's original prompt:
Use the rhythm of @audio1 for an eight-second continuous shot of one dancer. The person holds still for half a beat, then takes one turning step to the right after the first clear beat. The camera follows with a slow lateral move. Dark empty stage, one overhead light, and a clean background; do not cut or add lyrics or text.
This template only shows how to turn “there is music” into an actionable movement and camera instruction. It cannot be used to claim reproduction of the music video.
How to write reference media into the current workspace prompt
The current Seedance 2.5 workspace lets you refer to uploaded media as @image1, @video1, and @audio1. Assign a job to each file before writing the prompt. That prevents the common situation of uploading many files without knowing which one is meant to constrain the result.
| Media | A suitable job | What to specify in the prompt |
|---|---|---|
@image1 | Subject appearance, product shape, or scene composition | Whether it is a person/product/scene reference or a first/last frame |
@video1 | Existing movement, camera rhythm, or source clip | Whether it should be edited, extended, or only used as movement reference |
@audio1 | Rhythm, action cue, or mood transition | Which beat triggers an action, or whether only the overall rhythm matters |
The current contract permits up to 30 images, 10 videos, and 10 audio clips. Reference video total duration cannot exceed 30 seconds, and reference audio has the same 30-second total limit. First- or last-frame references cannot be combined with video or audio references, and first/last frame work requires the adaptive aspect ratio. Check the Seedance 2.5 video workspace for the current available settings and limits.
Reusable prompt checklist
Before submitting, check your prompt against these questions:
- Does this clip have only one primary action chain?
- Does each of subject, action, and camera have one clear description?
- Does every reference file have a named job?
- Does the camera movement agree with the subject's direction of motion?
- Have you explicitly excluded unwanted text, people, cuts, or extra actions?
- Do duration, resolution, aspect ratio, and task type meet the current workspace limits?
- Does this iteration change only one key variable, so you can judge the difference?
Keep an iteration log, not a collection of overwritten prompts
The GOKKO source describes a large selection process, not a single perfect generation. You do not need that scale to borrow the discipline. Record each run so a useful discovery does not disappear when you rewrite the prompt.
| Run | One variable changed | What you observed | Next decision |
|---|---|---|---|
| 01 | Baseline action and static camera | Does the action complete in the intended direction? | Keep or simplify the action before adding style |
| 02 | Camera path only | Does camera movement support, rather than obscure, the action? | Keep the path or return to static framing |
| 03 | One reference role only | Does the supplied image, video, or audio produce the intended constraint? | Retain, replace, or remove that reference |
This table is a working method, not evidence that a certain number of runs produces a better result. It simply makes each decision traceable.
A general shot template
This structure is neither model syntax nor a verified original prompt. It is a writing checklist. Replace the bracketed terms with your specific task:
[Subject] is in [setting]. Starting from [initial state], it completes [one primary action] during [time window]. The [shot size] camera [locks/pushes/pans/tracks] in [direction] at [speed]. Lighting and environment: [description]. Use [@image1/@video1/@audio1] as [specific job]; preserve [elements that must stay], and do not introduce [unwanted elements]. End on [end state].
When the draft is complete, remove adjectives that cannot be checked. Compared with broad phrases such as “cinematic,” “stunning,” or “epic,” a testable action, camera direction, and end state usually make the next iteration easier to judge.
Sources and research boundary
Every case in this guide links to the original public post:
- Lansen: a Midjourney + Seedance 2.5 wuxia duel
- Yu Shimura / GOKKO: a motion-first production note
- Xanderwow: the multi-model workflow behind “Safety & First Place”
- ByteDance Seed: Seedance 2.5 official announcement
The source material for this research does not include complete reproducible prompts, seed values, all reference media, or controlled tests. This guide therefore does not offer “official look” or “guaranteed reproduction” claims. Every item labeled “practice template” is an editorial example that should be tested with your own media, task mode, and current workspace constraints.
Next step: turn one idea into one shot
The most practical lesson from the cases is simple: complete a short shot with a clear action, camera, and reference role first, then expand into more media and longer storytelling. Review the Seedance 2.5 Guide when you need to validate task mode and input constraints. When your prompt and files are ready, open the Seedance 2.5 video workspace to check current settings, media limits, and the credit cost before submitting.
