Seedance 2.5 Skill

Language route

• For an English request, follow this file and the *.md references.
• For a Chinese request, read the Chinese workflow first, then use the matching *.zh-CN.md reference files.
• Keep model IDs, JSON keys, commands, media placeholders, and native-audio symbols exactly as code. Do not translate them.

Do not force every request through one image-grid pipeline. Choose the video route first, then enable only the preparation modules the job needs.

1. Choose the creative route

Use Seedance 2.0 as the executable default. Offer a Seedance 2.5 whole-short route only when the selected provider exposes that model and its actual limits.

The last four routes depend on capabilities that differ per model and per provider. Verify availability before offering one; see capabilities for the 2.0/2.5 comparison and for which published capabilities are platform features rather than API parameters.

Storyboard or individual keyframes

• Use the whole storyboard image in one R2V request by default whenever individual panels remain visually readable. Do not crop it first. Seedance can interpret the ordered panels as one continuous multi-shot video.
• Use I2V shot pairs only when independent reshoots or precise start/end states matter more than the transition quality of one R2V generation.
• Do not upload every storyboard cell as a default R2V asset pack. Use several R2V images only when each has a distinct role, such as subject, product, setting, style, or motion reference.

2. Enable only needed preparation

Subject brief

Use this only for a person, product, prop, hand, vehicle, or scene that must recur. Record 3–5 invariants: silhouette or proportion, signature material or wardrobe, key colour, and any must-preserve marking. Skip it for a one-off atmosphere shot.

For a recurring person, make a clean face close-up and a separate full-body reference. Do not use a front/side/back composite as the identity input; it can be interpreted as multiple people. Multi-angle product references remain useful when the object itself must be shown from several sides.

Keyframes

Create a start keyframe for every I2V shot. Add an end keyframe only when the shot must land on a specific action, composition, product pose, or hand position. Use one clean scene per keyframe.

Storyboard

Use an existing storyboard directly as the R2V reference. Seedance normally understands panel order and does not require panel numbers, dividers, arrows, or notes to be removed in advance. Make a clean copy only after a test generation actually renders an unwanted divider, number, caption, or multi-panel layout.

When a multi-shot request has no storyboard, use the Seedream template in prompt templates to create one board. Display that board in the host UI, inspect it yourself, then continue to Seedance R2V when it passes review. Showing the board is a progress update, not a user-approval gate.

For a supplied storyboard, display it unless it is already visible in the conversation. Check planned order, readable key beats, recurring-subject consistency, and content-specific constraints such as anatomy, product form, or critical text. Refine a visibly failed board before video generation. Ask the user only when a creative choice cannot be inferred.

If the route deliberately changes to I2V shot pairs, crop panels only then. Inspect the layout first: automatic crops can verify position, not whether the image model drew the intended panel layout.

3. Design continuity and cuts

First-and-last frames control one shot; they do not mean every shot must inherit the prior clip's tail.

Put important cuts, matches, and occlusions in the storyboard and prompt. Do not rely on a cross-dissolve to repair unrelated shots.

4. Write the video prompt

Scope: put each instruction where it applies

Before writing blocks, sort what you know by what it governs. Instructions in the wrong place are the most common cause of drift — a global rule written inside beat 1 stops applying at beat 4.

Restate the two or three most expensive locks at the physical end of the prompt; recency helps. That is a placement convention, not a fourth scope — the content still belongs to Locks and appears there first.

This mirrors the model-agnostic spec format in the Universal Video Prompt Skill. Use that skill when one brief has to run on more than one model; use this file for Seedance-specific writing. This is a required companion for a complete Seedance 2.5 Skill setup. If the link does not resolve, help the user install universal-video-prompt-skill before continuing; do not invent the missing shared specification.

Blocks

Use only the blocks that affect the shot:

[subject/reference binding]
+ [one observable action]
+ [space and important object relationships]
+ [one primary camera move, coherent composite move, or a cut]
+ [light/style when it matters]
+ [audio or dialogue when enabled]
+ [end state, for any stage that must land somewhere specific]
+ [must-preserve constraints]

End states carry multi-event work

For anything with more than one event, state what is visibly true when each stage ends. This is the highest-leverage single addition to a multi-stage prompt: it converts "keep it consistent" into something the model can target and you can check.

weak:   the two of them keep working on the bouquet
strong: end state: the florist holds the bouquet in the left hand;
        the scissors are back on the right side of the bench

An end state must be visible. "She feels relieved" is not one; "her shoulders drop and the frown clears" is. Read long video for the staged structure in full.

Time granularity: decide before writing beats

Granularity is a prior decision. Writing beats at second precision and then downgrading means rewriting them.

Infer it when the input settles it — a supplied music or voiceover track means second-level, a stated mood piece means none, an explicit fixed beat means second-level. When the request is a multi-event narrative with no external constraint, ask, and recommend with a reason rather than presenting a bare menu.

Timestamps allocate a time budget; they are not frame-accurate edit points, and actions may land slightly before or after a boundary. Do not demand impossible density such as three distinct actions inside one second.

• Name reference roles explicitly, for example Image 1: person, Image 2: product, Image 3: kitchen setting.
• For multi-shot R2V, list Shot 1, Shot 2, and Shot 3 in event order. Set duration in provider controls rather than forcing exact seconds in text.
• Prefer one primary movement per shot. A composite movement is valid when its direction, relation to the subject, and speed express one synchronized intent.
• When native audio is enabled, use （） for music, <> for sound effects, {} for dialogue, and 【】 for on-screen captions.
• State only constraints that are costly to redo.

Read the reference matching the job:

5. Generate, review, and finish

1. For a generated storyboard, create only the still first, display it in the conversation, and inspect it before any video request.
2. For a supplied storyboard, display the input unless it is already visible, then verify it fits the chosen route.
3. If the board passes review, generate one representative video pass without waiting for approval. Otherwise refine or regenerate the board first.
4. Review identity, locks, stage end states, composition, motion, seam, and audio in that order, and stop at the first failure — later checks are wasted effort on a wrong identity. Regenerate only the failed shot or segment.
5. Generate chains in order because the next segment needs the real prior tail. Generate cut-based clips independently and edit the planned transition.

The bundled script is a draft assembler, not a colour-grading or music-mixing system.

Atlas execution layer

Keep creative route selection independent from how a job is submitted. The default models are Seedream 5.0 Pro for stills and Seedance 2.0 for video; users may override them only after verifying route support.

In an agent conversation, use the Atlas Cloud Skill as the default direct generation route. It can discover a model, upload local media, submit an image or video request, poll, and retrieve outputs. Report Execution: atlas-skill only when it actually submitted the generation.

Use atlas-mcp only when the user explicitly selects MCP and its generation tools are exposed. Use atlas-cli only when the user explicitly selects a terminal, script, CI, or batch run. If the Atlas Cloud Skill is missing, help install AtlasCloudAI/atlas-cloud-skills before selecting a fallback.

Before reporting that an Atlas Cloud API key is missing, check the credentials in the selected execution process. For the REST runner, check ATLASCLOUD_API_KEY first and ATLAS_CLOUD_API_KEY as a compatibility alias. Do not infer credential availability from a different provider, plugin, or process; each execution channel can have an independent credential scope.

If neither key exists, direct the user to https://www.atlascloud.ai/console/api-keys?utm_source=github&utm_campaign=awesome-seedance-2.5-prompts-skills. Never ask them to paste the key in chat. Tell them to set ATLASCLOUD_API_KEY in the submitting process or the host's secure environment settings, then refresh or restart the execution session if needed. If the key exists in a parent or host configuration but is absent from the submitting process, report an environment-scope mismatch instead of saying the user has no key.

Billable task state machine

Apply these rules to every image and video generation:

1. After submission, record the prediction ID and logical stage immediately.
2. Treat starting, queued, pending, and processing as active. Poll the same ID every 2 seconds; never submit another task for that stage.
3. Treat completed and succeeded as successful terminal states. Download and inspect the output before starting a dependent stage.
4. Treat failed, timeout, and canceled as terminal failures. A new task requires an explicit retry decision; report the old ID and possible extra cost first.
5. A zero or missing processing-time field, delayed output, a local polling timeout, a stopped turn, or a temporary status-query error is not proof of failure. Preserve the ID and resume polling.
6. Interpret continue as “resume the existing task,” never as permission to retry. Do not submit video while its required storyboard is still active.

The 2-second interval applies to every Atlas execution route in this workflow. For atlas-skill, repeat its prediction-result step with the same ID. For atlas-mcp, call atlas_get_prediction with the same ID every 2 seconds. The MCP server performs a single status lookup per tool call; the agent owns the loop. The bundled REST and CLI adapters enforce the interval in code. A status lookup is read-only and must never be replaced with another generation call.

When the runner must resume, set execution.resumePredictionIds.<stage> to the existing ID. Supported stage keys include grid, ref1, ref2, seg1, shot1, and clip1. Never create a replacement merely because a prior polling process ended.

scripts/generate.mjs cannot invoke an agent Skill or MCP server. It is a separate batch runner: it defaults to atlas-rest, and can use atlas-cli only when explicitly set in execution.adapter. It never selects CLI automatically. See execution adapters.

# Generate a storyboard only. The runner prints [storyboard-preview] with an
# absolute path; display and inspect that image before starting video work.
GRID_ONLY=1 node scripts/generate.mjs scripts/myjob.json

# Run only the first independent clip or segment as a quality gate.
CLIPS_MAX=1 node scripts/generate.mjs scripts/myjob.json
SEGS_MAX=1 node scripts/generate.mjs scripts/myjob.json

{ "execution": { "adapter": "atlas-cli" } }

Available runner modes:

• grid: crop a model storyboard and make one independent I2V clip per panel; suitable for a hard-cut montage.
• shot-pairs: make independent I2V shots from segments[], with a first and optional last keyframe index.
• reference: send a small role-specific set of R2V references. A reference may be generated from a prompt or read from a local file.
• chain: continue one action by feeding the generated tail frame to the next segment. It is an alignment aid, not a guarantee of an invisible seam.

For fault-specific fixes, read troubleshooting.