AI Video Directing and Generation Segment Design
In this workflow, one take is one AI generation task, which can contain multiple camera shots. Organize generation segments according to complete dramatic changes and readable actions. Prioritize rhythm and effect, then consider cost and rework scope. Specific scripts, characters, episode numbers, aspect ratio, and the current budget stay with the project and are not treated as global fixed templates.
First, Define Generation Boundaries
For a full review, first temporarily remove existing segment numbers, reconstruct the event chain of "trigger - goal - choice/resistance - action - consequence" from the current script, and then determine generation boundaries. Existing segment counts, old budgets, or shot counts are not reasons to retain boundaries. When consecutive turns are fragmented, evaluate merging them; if merging would squeeze out performance, key processes, reference capacity, or single-generation stability, retain or split differently. Explain your reasoning for each item. Do not set a target for compressing the number of segments.
- Moderately merge a single scene's attack-defense, handover, or character choices and their reactions; do not create a task for each line of dialogue, each close-up, or each camera cut. Do not generate across episodes; even continuous scenes within the same scene should be separated at episode boundaries.
- By default, use the target model's native smart duration; for Seedance 2.5, use
duration: -1. Specific parameters follow the official documentation and integrated capabilities. If the user explicitly requests a fixed length, retain manual selection. Shot-level durations serve only as capacity and budget estimates, not to be hard-coded as total length. Do not force merge segments just to use up the maximum. If the model does not support smart duration, clarify the actual boundary. - Decisive effects that are difficult to control stably can be reworked independently, but retain the force, change, and aftermath. When adding new locations, subjects, or layouts, judge whether a single first frame can support this change.
- By default, retain original dialogue, speaker, tone, and causality. When the user authorizes script revisions, first correct unreasonable story facts, information order, and dialogue, then derive shots from the current source text. Do not use "retain original text" as an excuse to not fix script flaws. Leave time for vocalization, character changes, object handling, and reactions. Do not delete words or speed up delivery just to fill time.
Writing and Capability Constraints
When writing or reviewing body text, read Photography, Action, and Delivery Examples. Describe segment by segment: the freeze-start point, camera and action process, character reactions, and net endpoint. For camera cuts, explain the motivation, what is seen after the cut, and spatial continuity. Large scenes are established by space, staging, material, and power, not by piling on "cinematic" adjectives or unmotivated orbiting shots.
When involving a full script review, insufficient character performance, ambiguous actions, or plot segment rearrangement, also read Rehearsal, Performance, and Re-segmentation. Write according to the standard of what can be directed for a real shoot, covered by the camera, and recognized by the audience. For martial arts and other extraordinary results, retain the genre conventions, but the start of the move, contact, force, and aftermath must still be continuous. When the user authorizes script changes, synchronously correct action and performance directions. The camera direction manuscript and the story facts/dialogue should be layered; do not just modify the prompt but leave conflicting script notes.
When refining prompts, first read Visual Focus and Body Text Refinement: first decide who the audience is looking at and what change they see at that moment, then choose camera and action details. Logical correctness is only the baseline; also check if character intentions and relationships are visible. Do not replace directorial judgment with word reduction ratios or fixed shot counts.
When encountering excessive psychological explanation, redundant constraints, or overly detailed mundane actions, execute Performance Landing and Final Submission De-duplication. Retain specified tones and necessary transitions; do not rewrite the entire show into a uniform micro-expression template. Check the actual compiled body text, not just the editing box.
For long dialogues, identity reveals, evidence confirmation, and multi-character choices, re-check according to Division of Shots for Long Dialogues and Relationship Twists. Determine cut points based on information changes. Distinguish between full contact scenes, continuous statements, and scenes requiring shifting the viewing object. Do not automatically treat a fixed camera position or long lines as a problem.
For organizing shot by shot, use Camera-First Natural Paragraph Structure: at the beginning of each shot, sequentially describe the shot size, camera angle, fixed or moving method, and then unfold the action, line of sight, emotion, and dialogue. Camera information must not be buried inside long action sentences, nor should it just be a generic prefix added to old text. Move existing camera direction to the correct position and check that actions are not duplicated or omitted due to rearrangement.
When streamlining a scene or opening, execute Division of Scene Description, First Frame, and Shot Body Text: the body text scene should only list time and place, then go directly into Shot 1. Static compositions are left to the first frame plan. Prerequisites for actions are placed in the relevant shot. Do not assume that as-yet-unproduced images already provide this information.
When first frames have been selected, refine according to Action Body Text After Selected First Frame: directly write movement and performance based on the real image; do not restate the entire image. Grip terms cannot replace direction and trajectory; continuity only retains necessary constraints.
When wrapping up a full script review and revision, execute Dialogue, Visuals, and Cross-Field Review: verify promises vs. actual operation, occlusion and sound source, blocking and reachability, hand-use and prop whereabouts, limb ownership and control responsibility, shot size and evidence coverage, and whether side notes contain residual old events. For multi-person wrist-holding, support-handcuffing, or mechanical force application, independently check "whose which hand is controlling whose which body part" and whether these necessary contact points can actually be seen in the selected shot size. See that section for counterexamples and rewriting methods. When involving character references, check the different responsibilities of opening visibility, subsequent visibility, off-screen vocalization, and mere mention according to the target prompt skill.
When troubleshooting limb ownership and shot-size classification issues across the entire show, execute Dedicated Section on Limb Ownership and Dynamic Framing: from action subject, control handover, and body reachability check to full-process framing and first frame continuation. Based on evidence needs, choose to relax the framing, track with the action, or retain close-ups. Do not just change pronouns and shot-size names.
The current default work uses Seedance 2.5 with 480P previews and smart duration of 4–30 seconds. Upgrading requires explicit user request and confirmation per instance. When using Seedance, combine the official sd25-pe with the custom seedance-2-5-prompt. For short dramas, also reference short-drama-storyboard and short-drama-video-prompts. Retain third-party source text. This skill defines generation segment organization and directorial judgment; it does not replace its materials, body text, or model rules. In other environments lacking these skills, still execute the methods in this document, and check the current provider's official guidelines. Do not claim to have read them.
Read the current project's integrated capabilities, then select text-to-video, first-frame, reference, editing, or extension mode. A single first frame only anchors the opening; internal camera cuts rely on text and previous state deduction; do not pretend every cut has a keyframe. When creating attachments for the first frame, prioritize according to the opening's required identity, clothing, space, and key props. Prefer real images matching the opening state; avoid pre-damaged, pre-opened, pre-delivered, or pre-doffed states. Mark unproduced first frames clearly as plans. Do not treat a character design board directly as a story first frame. Do not treat multi-image references for first frame creation as indicating the video already supports multiple references.
Reduce Non-Informative Pauses
When encountering action and state conflicts, execute Effective Stages of State Constraints: separately check opening facts, invariant facts for the entire segment, stage constraints, and end state. Place conditions like "after clamping", "after unloading", "after steadying" back into the action. Leave the end state in the last shot or production record. Do not just rename "maintain". Do not let the first frame composition restrict subsequent camera cuts. Check each pair of previous segment's endpoint and next segment's starting point to avoid repeated handovers, door-through sequences, or mechanism activations.
Ensure each photography segment advances action, dialogue, evidence, or character choice. When the result is distinguishable, cut to the next beat. Where causality permits, let existing action preparation, vocalization, and receiving reaction be in the same shot to avoid pressing a complete sentence plus extra margin onto a static listener. Do not use a uniform "tight/no pause" suffix, random micro-movement, orbiting shots, or repeated empty shots to fill time.
Retain pauses for weight-bearing, steadying, releasing, and emotional turning points that truly change a choice. If the source text already has clear event times like "half a second", continue to retain them. The default smart total duration does not mean deleting them. Smart duration also does not guarantee it is shorter; the effect must be verified by real output. Do not fabricate universal success rates based on a single sample. Keep optimized new body text and the original frozen task input in separate records; do not overwrite the original task.
Sound, Voice, and Subtitles
When the creator requires a complete audio delivery, use the bound real audio according to the actual speakers in this segment. Freeze character, voice, and audio responsibilities. If missing or unreadable, clearly state this; do not impersonate another sound. Voice references only control the corresponding character's voice characteristics. Listening to an example sentence does not replace the dramatic dialogue. The model output must still be checked for language, age feel, tone, and lip sync.
Adhere to the actual provider's modal mutual exclusion constraints. Seedance 2.5's strict first frame and audio reference cannot be mixed. When using image+audio combined references, the opening image does not promise pixel-level locking; the delivery must indicate the actual routing. Only after parameters and materials are actually passed in can you claim that references were used.
Subtitles are a post-production task, not a sound generation parameter. When the user chooses external editing or outsources post-production, do not add subtitle switches, task configurations, or auto-recognition/burn-in processes. Only retain the original dialogue, speaker, source, and shot-correspondence draft. Do not fabricate timecodes when no real audio track exists. According to project requirements, constrain the body text not to generate on-screen subtitles; this cannot be used as an excuse to delete dialogue, turn off native sound, or forbid necessary text within the story.
When additionally commissioned to produce subtitles, produce them based on the actual vocal recognition/alignment time; do not divide based on word count or shot budget. Check the original text, speaker, errata words, and end-of-sentence time. Do not fill silent sections with unspoken dialogue. Retain the original video. If recognition/burn-in fails, you can correct subtitles based on the actual audio track and real start/end points, marking the correction source. If the model actually speaks extra lines, truthfully mark them as creative deviation; do not rewrite the script to cover it up. Do not treat recognition or correction success as final product qualification. Only redo post-production; do not automatically recreate paid videos. Generate precise subtitles preferentially in post-production; avoid treating model-generated text as proofread subtitles.
Output, Cost, and Review
Deliver separately: Production Information (source, episode number, duration, boundary rationale, start/end states, risks, attachments, and budget) and Submittable Body Text (actual visuals, action, performance, and camera). Resolution, aspect ratio, total duration, model, sound on/off, cost, and post-production notes go only in parameters/remarks. If no digital timeline is requested, use sequential paragraphs. Retain existing precise event timepoints; do not promise frame-level control.
First, complete reviewable prompts. If the product cannot support it, then supplement display, attachment, and persistence capabilities. Then, sync data as authorized and read back for verification. A solution document alone does not constitute authorization for additional images, videos, or dubbing. Before each actual new or reworked video, show that segment, the input for this attempt, and the cost, and wait for explicit confirmation. Historical or batch authorizations do not cover subsequent calls. Sound and precise handwriting are listed as separate deliverables based on current capabilities; honestly mark media that has not been generated or accepted.
Bills use current official pricing, noting the check date, mode, output specifications, estimated output seconds/tokens, first frame and reference image costs, and separately calculate successful output and rework risk. When charged by output volume, a drop in task count does not mean a proportional cost saving. Retain necessary pacing; limit rerunning entire high-risk long segments. Distinguish between successful but unsatisfactory outputs and technical failures according to actual billing rules; do not call regeneration a "free retry".
Review criteria: no cross-episode segments; dialogue and tone are complete; segment length is performable; contact precedes result; left/right hand responsibilities, prop ownership, injuries, and clothing are continuous; attachments are real and consistent with the opening; parameters and body text are separate. Watch the full output before judging pacing and quality. Do not substitute engineering tests or simple file playability for final product acceptance.
For direction requirements of swords, tools, door slots, and wheels, execute Stage Constraints on Blade Direction and Contact Direction: only use the audience's left/right for fixed perspectives; after a camera cut, use physical reference points. Clarify the contact surface, support, and effective stage. Necessary turning must occur after contact is broken. Do not use a blanket "no reversal" ban to overwrite normal actions or mechanism rebounds.
When reviewing camera positions, first-frame dependencies, secondary characters, and sound together, execute Minimal Initial State and Readable Scenes: first check facts against examples. Embed necessary initial states into actions. In readable scenes, distinguish between the main subject and established static objects. Camera cuts should only take key states. Separate voice and performance responsibilities.