Seedance 2.5 Prompt Skill: Every Reference Needs a Job

sd25-pe sounds like it might give Seedance 2.5 a secret upgrade. Here is what actually happens: an agent reads the Skill, then organises the story and references into a prompt. Seedance receives that prompt and the files. The Skill never enters the video model; it disciplines what we send.

In our paper-boat project, its value was painfully practical. It kept one cat as one cat, stopped Act 1 from leaking into P08, and left resolution settings out of the plot. AI video projects have a talent for tripping over small details. Every trip still costs a generation.

This guide follows sd25-pe v0.1.1 as retrieved on 2 September 2026. Check the source again after an update.

Install it, then record the version

The Skill source provides this installation command:

npx --yes skills@latest add \
  "https://arkdocs-en.tos-ap-southeast-1.volces.com/skills/" \
  --skill sd25-pe \
  --yes

Version 0.1.1 also tells the agent to attempt an update the first time the Skill is triggered in each session:

npx --yes skills@latest update sd25-pe -y

Work can continue from the local copy if the update fails. Keep the version number whenever you document why a prompt was written in a particular way. An automatic update can attach new rules to the same sd25-pe name.

Start with intent. Pretty sentences can wait

The first principle in v0.1.1 is Intent First. The agent builds an internal story contract from the source request and locks:

• subjects and their counts;
• actions, event order and causality;
• scene, time, weather and spatial relationships;
• prop ownership, transfers and final states;
• visual treatment, camera, sound, dialogue and subtitle requirements;
• whether the job is generation, editing or extension.

Reference images contribute visible appearance, material, composition and lighting. Character names, relationships, events and outcomes still come from the user's text. A uniform does not spontaneously give a character a profession. Two similar portraits do not come with a complimentary clone.

Abstract intent needs filmable behaviour. “The cat becomes curious” can become a forward movement of the head, widened eyes and a shift of weight. “The boat plays with the cat” can become a boat that retreats when the cat approaches and drifts closer when the cat stops. The writing may carry philosophy. The shot still needs an action.

Generation, editing or extension: pick a lane first

The Skill routes the job to one primary task:

Primary task	What the prompt must control
Video generation	New subjects, events and scenes, plus the contribution of each reference asset
Video editing	One source video as the sole master, with only named objects, regions or sounds changed
Video extension	A new segment before or after the source, preserving the boundary state

Multiple references, longer duration, boundary frames, storyboard grids, blockouts and sound are modules attached to that task. A Blender blockout can supply blocking, paths and camera work while the job remains video generation.

If one request changes a source clip and then extends it, v0.1.1 creates two sequential prompts. The first output becomes the master for step two, so the operations do not wrestle for the steering wheel inside one generation.

Six sections keep a crowded prompt under control

A text-only prompt for one event can stay short: who does what, where it happens, then any essential visual, camera and sound direction. Once references and events multiply, v0.1.1 uses this working structure:

【Generation Goal】

【Reference Asset Roles】

【Unused Assets】

【Subjects and Relationships】

【Event Script】

【Maintain Consistency】

Keep only the sections the job needs. The generation goal names the video type, subject and main event. The event script controls the opening state, sequence and ending. Maintain consistency holds the identities, props and spatial relationships that must survive across shots.

The structure also catches collisions. Put the same action under asset roles, event script and consistency, and three versions of its timing can quietly enter the prompt. Keep timing in the event script. Let each asset-role line explain what its file defines. The prompt becomes much easier to audit.

Every asset needs a job title

The Skill uses this mapping priority:

User's explicit specification > Prompt description > Asset content > Filename and metadata > Upload order

Every active asset gets one clear assignment. A portrait defines a face and clothing. A scene image defines layout and lighting. An action video defines a hand path. “Use all references” is where accountability goes to disappear.

Bind different characters on separate lines. When several views define one character, say that they jointly describe one entity. Leave that relationship vague and the model may treat the views as separate people.

When the agent can inspect the full inventory, it lists each inactive file under 【Unused Assets】. That stops later prompt enhancement from inviting the file back as a person, scene, prop, action, camera reference or sound source.

If an asset cannot be inspected, the Skill preserves the user's existing @ImageN, @VideoN and @AudioN relationships. It does not fake an inspection or reshuffle established references.

Multi-panel storyboard reference: composition yes, exact timing no

BytePlus calls the capability Multi-panel storyboard reference in its Seedance 2.5 Prompt Guide. In sd25-pe v0.1.1, the storyboard grid is a composable module. The prompt should state:

• the panel count and reading order;
• the shot assigned to each panel;
• the framing, subject position and approximate composition to adopt;
• the sketch style, grid lines, labels or placeholder characters to exclude.

The Skill recommends a clean grid with little annotation and roughly 15 panels at most. That is stability guidance; platform hard input limits are a separate question.

Act 1 of our paper-boat film activated P01–P03 from an eight-panel grid. The prompt excluded the other five panels, along with the grey sketch treatment, grid and placeholder character designs. Separate assets defined the final cat, convenience store, cardboard box and paper boat.

The grid works as a semantic composition anchor. The model may redistribute hold times and adjust camera positions; pixel-perfect copying was never on offer. Frame extraction still gets the last word. Our Seedance 2.5 storyboard test places P01–P08 beside the corresponding landscape and portrait film frames.

Name the first frame or frame zero has no address

An ordinary image reference tells the model how a subject or scene looks. A first-frame instruction also gives frame zero a starting address. In an English prompt, v0.1.1 requires this exact role statement on its own:

Use @Image1 as the first frame.

The next sentence describes the opening composition, subject positions, poses, prop states, scene and camera direction defined by the image. Other references may contribute appearance, props or scenery only when they preserve that composition. The event script then continues naturally from the frame defined by @Image1.

The preserved paper-boat submission used a Chinese first-frame role sentence. That wording belongs to the historical submission. A new prompt written under v0.1.1 should retain the English sentence above.

The first image locks the output aspect ratio. When first and last frames are supplied together, they should share the same ratio. Treat those as submission checks and keep them out of the visual description. A first frame is still a semantic boundary. Pixel-identical continuity with the previous clip is a cheque the model has never signed.

Seconds are an event budget. Extract frames anyway

When the user supplies time segments or explicitly asks for timing, the Skill preserves continuous, non-overlapping whole-number ranges. Each range carries one major state change and ends with an observable state for the subjects and props.

When total duration exists only in the page or API settings, the prompt uses unnumbered stages. It does not reverse-engineer new 0–5s and 5–10s ranges. If the configured duration runs longer than the event timeline, the agent can extend movement, reactions, pauses and transitions. It cannot fill the gap with extra characters, a new main event or a surprise ending.

A time range allocates narrative room. It does not appoint an exact edit frame. Both paper-boat films contain all eight storyboard compositions, while events in Acts 2 and 3 landed at different times from the written ranges. The storyboard test records the extracted timings and final joins.

Keep ratio, duration, resolution and frame rate in settings

For regular video generation, v0.1.1 places aspect ratio and total duration on the page or in the API request. Resolution, frame rate and the sound toggle stay in the settings layer too. A prompt can describe rain, dialogue or the sound of a rocking paper boat. The interface controls whether sound is enabled.

Task	Ratio and duration handling
Regular video generation	Set ratio and total duration on the page or through the API
First-frame / first-and-last-frame generation	The first image locks the ratio; duration remains configurable; first and last images should match
Video editing	Inherits the source ratio and approximate duration; input-frame handling may move output duration by about 0.3 seconds
Video extension	Inherits the source ratio; extension duration remains configurable

If a reference video and target ratio differ, or the target duration differs from the reference by more than about 0.3 seconds, the Skill routes the task to regular generation. Explicit continuation before or after the source routes it to extension. Run this compatibility check before submission.

Once a confirmed ratio or duration conflict sends the job to regular generation, v0.1.1 permits exactly one fixed sentence under 【Generation Goal】: Please note that this is not video editing. Leave it out when the compatibility condition never fired.

One template line nearly welded three shots into a oner

The first paper-boat prompt said the selected panels should unfold “in one uninterrupted continuous shot.” The lines below it specified three camera setups and an edit. Great drama, broken camera logic. We removed the template sentence and moved future P07 and P08 events out of an Act 1 submission responsible for P01–P03.

The story stayed intact. Each submission got one smaller job, and the model no longer had to peek at the ending twelve seconds early.

Translation note for this case: our operations team is based in Taiwan, and the Act prompts actually submitted to Seedance were written in Chinese. The English wording in this series is an editorial translation. We did not submit or test the translated wording and do not guarantee equivalent results. Validate it with your own references and settings before relying on it.

The case evidence lives in two companion articles. From our first prompt to the prompts we actually submitted follows how the grids and reference images helped the prompt-writing LLM find problems in the text. The P01–P08 finished-film test checks every storyboard panel against an extracted film frame. The rest of this guide stays with the rules we can reuse.

Seven checks before you spend a generation

1. Is this prompt's primary task generation, editing or extension?
2. Do subject counts, prop counts, causality and the outcome still match the source intent?
3. Does every active asset have one clear role, with each inactive asset excluded when the full inventory is known?
4. Does the storyboard specify reading order, shot roles and visual annotations to exclude?
5. Does each first or last frame have its own role statement, with other references subordinate to the boundary composition?
6. Did the time ranges come from the user's request, and are they treated as event budgets?
7. Are ratio, total duration, resolution, frame rate and the sound toggle still in the page or API settings?

Seven basic questions. They become philosophical the moment paid credits are on the line.

For the underlying subject, action, scene, camera and style vocabulary, start with our Seedance 2.0 prompt guide. The 2.5 Prompt Skill carries that foundation into task routing, multi-asset mapping, boundary frames, storyboard grids and blockout references.

Version and sources

Last checked: 2 September 2026, against sd25-pe v0.1.1. Skills and platform constraints can change. Use the Skill and Seedance interface available at submission time as the current authority.

• Official sd25-pe v0.1.1 package (ZIP)
• Seedance 2.5 Prompt Guide: Multi-panel storyboard reference (BytePlus)
• Seedance 2.0 series prompt guide (BytePlus)