Seedance 2.5 Prompt
Check the official documentation and capability boundaries to verify the target model and execution endpoint before writing the main body. For a complete production package or example, read Generation Plan and Examples. This is a custom workflow, not an officially released Skill.
Use the official sd25-pe to organize generation targets, actual reference responsibilities, event scripts, and maintaining relationships; third-party 2.0 dialects are not involved in new 2.5 tasks. The third-party 2.5 experience that every segment must have a keyframe and the second consecutive segment must use extend are not official hard limits. A single generation can contain multiple camera shots; normal same-scene continuation can explicitly use the real last frame. Project conventions take precedence over third-party templates.
Lock Facts and Routing First
Read the original script, episode number, character definitions, and current appearance, voice bindings, accessible assets, and endpoint capabilities. Default to preserving characters, dialogue, and story facts; when the user authorizes script review, first correct any problematic facts, dialogue, and information order in the source text, then update the prompt, attachment responsibilities, voice mapping, and later text layers based on the current source text. Do not freeze plot holes with old dialogue. When assets cannot be viewed, retain existing references, do not guess appearance from filenames, and do not claim to have reviewed.
Distinguish between text-to-video, strict first frame, and image/audio joint reference. Only an image that is truly uploaded and specified as first_frame is a strict first frame; reference_image serves as an opening state reference but does not guarantee pixel-level locking. Seedance 2.5 supports pure audio reference. Strict first frame / start-end frame is mutually exclusive with full-modal reference; when character voice is needed, use full-modal reference and explicitly state the opening state in the main body. Check the current application routing before execution, handling reference generation, start-end frames, editing, and extending separately.
Assign responsibilities to each asset one by one: opening state, a person's identity and clothing, environment, or props. Use image numbers according to the actual attachment order; audio is numbered separately; use consistent names for the same subject, and for groups explicitly state which person they correspond to. Attachment library IDs cannot replace image/video numbers in the main body. The first-frame asset list and video asset list are two separate lists; do not automatically copy numbers. Unused assets are not forced into the frame; missing assets are only listed in production notes.
After review or rearrangement, execute Visible Character and Asset Coverage Review. Whether a character speaks does not determine whether they need a visual reference; check separately for visible at opening, visible later, off-screen voice, and merely mentioned, along with appearance version, real assets, attachment numbers, and capacity, then read back for verification.
When the editor uses named references, execute Named Reference to Actual Attachment Closure. Displaying as a reference tag, having a binding record, and actually being submitted with the request are three different things; verify with the current version and current selection compiled preview; do not use old production documents to overwrite bindings that have been fixed.
Segmentation and Cinematography
When the user requests a full review or rearrangement, first use ai-video-directing to redetermine boundaries from the event chain, then write the main body of this skill. A third-party prompt skill stating "do not change shot boundaries" is its stage responsibility, not a reason to freeze the user's old plan; authorized merging, splitting, and script revisions are completed at the directing stage. Do not use the old segment count as the target, and do not merge by concatenating old main texts.
A segment is one generation and can contain multiple camera shots; do not cross episodes. The model capacity is 4–30 seconds, default smart duration is -1; specific delivery specifications follow the project, user manual duration is preserved. First determine whether dialogue, action, and reaction can be completed naturally; when overloaded, split based on action completion, handover completion, or character choice; do not delete lines, do not uniformly speed up, and do not use 30 seconds to reverse-engineer a per-second timeline.
Camera shot subtitles, plot summaries, and dramatic responsibilities are only kept in production information; do not include them in the first-frame or video submission main text. Multi-shot main text can use "Shot 1: Specific image..." as a separator; do not write "Shot 1 | Summary Title: ...". Before deleting a title, check whether the necessary events, objects, and time points within it are already expressed in the concrete description; if missing, supplement them into the corresponding actions; do not lose plot information. Shot numbers, reference responsibilities, lighting, and necessary continuity tags can still be retained; do not expand title cleanup into deleting valid instructions.
The main text for each shot uses short sentences to indicate time and place, then immediately goes into Shot 1 without a separate opening scene. Follow ai-video-directing's "Division of Labor Among Scene Description, First Frame, and Shot Main Text"; detailed static composition stays in the first-frame plan; starting premises that affect action are written in the relevant shot; a first frame that has not been produced cannot be treated as submitted information. Each camera shot clearly defines shot size, camera position, the object being viewed, one main movement and its endpoint, action and received reaction; cuts convey information or action motivation. Simple events can be completed in a single paragraph; do not force-fill templates. Preserve the original event time points and brief pauses; prompts do not guarantee frame-level control.
When organizing each shot, execute ai-video-directing's "Paragraph Structure with Cinematography First": for each shot, first write the shot size, camera angle, whether it is static or has an object and endpoint movement, then write concrete actions, line of sight, character emotions, and original dialogue in chronological order. Relocate cinematography descriptions scattered later; remove duplicate actions; handovers, dodges, and clothing changes mixed in old sentences cannot be deleted along with cinematography prefixes. Scene lighting is explained centrally; there is no need to repeat focal length, force a movement, or fabricate seconds for each shot. Keep shot numbers in the main text; summary titles remain only in the production column.
When checking the relative direction of a sword and body, disambiguate using action intent and actual motion plane: specify axis, blade tip, lateral clearance, and contact area; do not use "no injury" to obscure a path that would pass through the wrist. The first frame, landing state, and subsequent drawing are established separately; only write the viewing angle when plane projection parallelism is needed; do not fabricate precise angles or on-set safety guarantees.
Write action including the force-applying entity, path, contact point, resulting change, and aftermath; ensure continuity of left/right hands, prop holder, gaze, clothing, and injuries. The official suggestion of small, continuous actions is for reducing unnecessary complexity; it cannot change a required fight into a stationary confrontation. Reduce concurrent changes within the same segment; maintain boundaries for high-risk effects to allow independent retakes.
For multi-person contact and mechanism force application, compile according to the custom directing skill's Limbs Attribution and Dynamic Framing Special Section: explicitly identify easily confused action subjects, the limb applying force, and the body part receiving the action; track control according to transitional phases: secured, handover, withdrawal. Check necessary reachable processes such as standing up, freeing a hand; then verify whether the action starting point, contact change, and landing point are within the frame. First distinguish between what must be in the same frame, what can be shown sequentially, and information already established in previous shots; choose between widening the frame, following movement with a clear endpoint, or keeping a close-up. Do not rely on long modifiers to guess wrist ownership, and do not only change the shot name. Correct the synchronized first-frame plan and start/end fields; if a real first frame already exists, verify against the actual image; do not automatically replace or redo due to text changes.
Write performance into the actual shot: visible trigger, object of gaze, one key expression / physical reaction, then the subsequent choice and the person addressed. Turn "understanding, shock, wanting to flee" into filmable behavior; do not write mind-reading explanations; do not make every actor react simultaneously; do not replace performance with adjective stacks. For multi-person or multi-prop shots, use consistent character names and specific objects to disambiguate; avoid consecutive "he / the other / over there". First establish space, then write "same side"; separate an actor's left/right hand from the screen direction.
When refining, distinguish between visible performance, vocal tone, and psychological explanation according to Performance Landing and Final Submission De-duplication. "Contemptuous, calm" are not forbidden words; preserve the original specified tone; do not summarize intents already expressed by actions; do not mechanically add frowning or gritting teeth. Public image/sound rules are only stated once in the final submission main text; named assets and real number mappings are not deleted by de-duplication; when the app automatically appends them, do not repeat in the creative main text; when delivering the full text outside the app, supplement them. Do not use runtime blanket deletion rules to rewrite user free text.
After merging, reorganize shots, dialogue, and reference mappings; check the number of speakers, total audio duration, opening adaptation, and complex changes. Each line of dialogue appears only once, written directly in the shot where it begins; if it continues across shots, state how far it extends; do not provide another copy of repeated lines. Remove review explanations unrelated to the image, lengthy exclusions, and repeated start/end states; keep necessary state constraints as positive visible facts. Prompt length has no universal success threshold; reduce information conflicts and ambiguity, not key process steps to meet character count.
For refinement and full rewriting, use ai-video-directing's "Image Focus and Main Text Refinement": first determine the object of viewing, character intent, and relationship changes for each camera shot, then compile into a natural image description. Simplify ordinary actions; retain details for dangerous contact and support handovers; explicit cinematography changes do not add new generation tasks. Centralize reference responsibilities; unify lighting for the same scene; the first frame only presents the static starting point. Do not consider refinement complete just by reducing characters, increasing shot numbers, or writing focal lengths for each shot; the source text, shot main text, first frame, and dialogue mapping must be verified together.
When a real first frame has been selected, execute Action Main Text After Selected First Frame: write action directly after cinematography; do not restate the stance, clothing, and grip list in Shot 1. Write necessary direction and contact relationships in the action area; centralize sleeve continuity once. Asset numbers and voice colors must still be compiled according to real attachments; joint reference does not mean pixel-locked first frame.
For consecutive segments in the same scene and episode, prioritize previewing the real clean last frame of the previously selected original footage; verify the actual state and explicitly select it as the opening reference for the next segment; this is not uploading the previous segment's video for extending. When no real last frame is available, keep a "to be prepared" state. Use an independent starting point for episode beginnings, transitions, or situations where the last frame is unsuitable.
When compiling state constraints, use ai-video-directing's "Effective Phase of State Constraints": the opening only writes the present state; continuity across the entire segment only writes facts that hold throughout; phased conditions are written after the corresponding contact or change; the ending state is not written back as a posture maintained throughout. For example, if the sword is still in hand, do not write "the sword is always in the door frame"; first write the release into the wood, then the state after embedding. Check whether reference images prematurely show the final state, whether the first shot can start from the first frame, and whether adjacent segments repeat already completed actions. Do not append redundant action summaries at the end of a segment for handovers and landing states already clearly written in the shot. Retain weight-bearing and emotional pauses; do not use "continuous movement" to counteract contradictory instructions.
Sound and Main Text Boundaries
Dialogue must match the current source text word for word (including revisions authorized by the user), clearly specifying the speaker, designated language, and tone. Use Character says in specified tone{original text}, <action sound effect>, (music) to distinguish sounds; do not add new music or sound sources simply because symbols are used. Reference the actual bound audio for character voice color and freeze the version; maximum 10 segments, each 2–30 seconds, totaling no more than 30 seconds; if missing or exceeding, re-prepare; do not silently switch voices or shorten to reuse old confirmations.
Aspect ratio, resolution, total duration, model, audio toggle, price, and post-production notes remain in parameters / production information. Only add constraints like "no off-screen subtitles" when the user or project has requested them; do not convert such main text requirements into fictional model API parameters. When subtitles are handled by external editing or separate commissioning, preserve the original dialogue, speaker, and shot mapping; do not add subtitle post-production configurations to the generation interface, request, or database; do not auto-recognize or burn subtitles. When later commissioned to produce precise subtitles, verify against the actual audio track; do not attach unspoken lines or pre-generate timecodes.
When the user only requests rewriting, return only one main text; when the user requires a generation plan, deliver production information, real sequential attachments, first-frame prompt, video main text, parameters, and unresolved items. When saving to Loop, use proper authentication and revision checking; read back the main text, attachment order, and parameters; do not automatically submit generation.
Draft and Upscale
New videos default to 480P Draft, smart duration -1. Only prepare 1080P final video when the user explicitly requests an upscale. Draft and upscale are counted, confirmed, and billed separately; the upscale references the original draft task ID, inheriting its real prompt, attachments, seed, sound, duration, and aspect ratio; do not silently carry the current rewritten main text into the upscale. Rewritten content should generate a new draft and be confirmed again each time. The validity period is calculated from the supplier's creation time; expired videos cannot be automatically resubmitted.
For reference generation, explicitly use reference; for editing, explicitly specify the unique master, modification scope, and items to preserve; lock ratio=adaptive,duration=-1; for forward/backward extension, explicitly specify the boundary action, sound, and direction; lock ratio=adaptive. Parameters remain in production information; the main text only expresses events and asset responsibilities.
Submission and Review
This skill does not call generation services by default. For each new or redone real video, first read from Loop the generated/total count, current segment and episode number, actual call count for this segment, current attempt number, and ungenerated count; display frozen character, voice color version, real attachments, specifications, estimated cost and pricing basis; then wait for the user's explicit confirmation for this segment and attempt. Confirmation of changes requires re-review; do not batch-carry over authorizations; do not automatically resubmit requests with unknown results.
Before output, check each item: characters, original dialogue, causality, action capacity, start/end states, attachment responsibilities and numbers. After generation, save the original footage and watch it in full; review identity, action, sound, and continuity; technical success does not equal acceptance testing. Report whether the official last frame is returned and saved; if there is original footage but no official last frame, extract frames locally and note the source; if no original footage, state that no frame can be extracted. Fees are calculated based on estimated and returned usage at frozen unit price; unknown cannot be recorded as zero; actual deductions are based on supplier invoices; do not add manual real-payment registration.
At the end of text revision, also execute ai-video-directing's "Dialogue, Image, and Cross-Field Review", especially checking off-screen sound after a door closes, whether cross-shot lines are still being spoken, and whether lighting or reference descriptions re-introduce actions that have been removed. Programmatic validation cannot replace these semantic checks.
For directional requirements of swords, tools, door slots, and wheels, execute Sword Direction and Contact Direction Phase Constraints: only use the audience's left/right for a fixed perspective; after a cut, use physical references. Clearly define the contact surface, support, and effective phase; necessary turns must occur after breaking contact. Do not use a blanket "no reversal" rule to overwrite normal actions or mechanism recoil.
When reviewing camera position, first-frame dependencies, supporting characters, and sound together, execute Minimum Initial State and Readable Scene: first check the example for factual accuracy; embed the necessary initial state in the action; in the readable scene, distinguish the subject from already established stationary objects; cuts only take key states; separate voice color and performance responsibilities.