short-drama-video-prompts

Short Drama Video and Timeline Music Prompts

Write the confirmed shots and start frames as bounded, generatable, directly copyable 剧集/<EP>/视频提示词.md. Use MOTION-... title for each shot and reference the corresponding SHOT-...。

Quick Start

用 $short-drama-video-prompts 把 EP001 分镜逐镜写成可直接复制的视频提示词

Entry and Boundaries

With the current storyboard, frozen keyframes, and necessary visual facts, reference readiness can be checked; image prompts are not proof that images exist. Motion prompts own the change from the confirmed start point to the end point: performance actions, prop states, camera movement, rhythm, and sound cues. First choose the path according to the storyboard's "Input Reference Image":

When there is a readable real REF-... image, it isimage-to-video; carry over the slots, order, project-relative paths, Chinese names, purposes, control scope, and must-not-control scope from the storyboard as-is. Static identity, styling, location, and composition may be jointly owned by the reference frame and the visual setting. This line is the shot's asset list: the path answers "which one to send",用途 answers "what it is responsible for in this shot",顺序 answers "what label it is called in the copyable body". Slots that cannot answer these three questions are not ready.
Only when it is "Input Reference Image: None (the creator has explicitly chosen text-to-video)" is ittext-to-video; the video document must include a non-empty "Static Visual Anchor" and incorporate it as-is into the copyable body, so that the prompt itself conveys the facts necessary for this shot among characters, styling, location, composition, and lighting. Do not write only preserve/keep/保持已有外观, because the execution end has not received an image to preserve.
When there is a PLAN-... slot, it is likewiseimage-to-video: the image is attached by the creator during generation, and this kit has no file to send. Write the body as image-to-video as usual,顺序 is the order in which the creator attaches images,用途 explains what each image is responsible for in this shot. These shots cannot be converted to $short-drama-produce; to put them into production, first put the real files into the project and rewrite them as REF-...。
"Input Reference Image: None (reference image to be supplemented: ...)", existing REF-... or PLAN-... still carrying "; reference image to be supplemented: ...", or a plain "None" all indicatenot ready, not text-to-video authorization. First automatically search for usable real images and ask the storyboard owner to refresh the binding; if anything is still missing, list the shots, missing images, and corresponding IMG-... or frozen keyframes, and do not write the final 《Video Prompts.md》.

Creator-readable explanations follow the project language. Before starting, read the version, execution mode, prompt dialect, native duration, and reference method in the target model profile; these choices simultaneously constrain the upstream storyboard and the downstream adapter, and cannot be guessed only at the body stage. If the creator names the target model in this request (for example, "write according to MiniMax H3") but the project profile does not yet have this item, first switch to $short-drama to write this choice into short-drama.json 's production_profile, then return to this stage; if a named reference in a session line is not recorded in the file, the next round will fall back to the generic path, and the dialect, native duration, and reference method will have to be guessed again. For standalone tasks without a project configuration, write the named model and the resolved dialect on the first line of this document for downstream verification. The video-copyable body preferentially follows short-drama.json#/creator_authority/production_profile/choices/video_prompt_language; when undeclared, fall back to short-drama.json#/format/prompt_language. When there is no short-drama.json , confirm the body language based on the target video model named by the user; when the target model is also undeclared, follow the user's current language, not fixed to English. The spoken language of dialogue, VO, and OS is still determined by the script and sound direction, and cannot be inferred from the prompt body language.

When the target is seedance-2.0 , write according to Seedance 2.0 dialect , when the target is seedance-2.5 , write according to Seedance 2.5 dialect , when the target is minimax-h3 , write according to MiniMax H3 dialect , when the target is wan-3.0 , write according to Wan 3.0 dialect . Read only the one that is hit. Other targets or versions do not apply the syntax of similar models, and continue writing according to the generic path of target model capability profile .

Follow the existing creator-first skeleton:## MOTION-... · 中文名, with field names in order using "shot, duration, generation method, input reference image, static visual anchor, start frame, state chain, endpoint", and the body title using ### 可复制提示词. The static visual anchor follows the resolved video prompt language of this shot and enters the body as-is; every non-empty line of the body is composed with > into a Markdown quote block. The rewritten content must not be renamed, have alternative fields created, or be changed into ordinary paragraphs.

When the person speaking in this shot needs to maintain their timbre and the target model accepts reference audio, add on the line below "input reference image" - 参考音频：REF-<slot>（顺序：<n>）· <音频路径>《<中文名称>》（用途：音色；角色：<人物名>；控制：<范围>；不得控制：<范围>）, connect multiple with ； (VID-27). For the path, copy exactly the - 声音参考：registered for that person in "Visual Settings.md"; the character must speak in this shot's body the lines recorded under their name in "Script.md"; if someone else in this shot's source scene also said the same line, name the speaker in the clause before the quotation; 不得控制 write at least 台词、语气、情绪(separate items with enumeration commas or commas, do not connect with "and"). The audio's 顺序 is numbered separately from images starting from 1, which is the @音频N、<Audio N> or 音频Nin the body. Audio does not go into "input reference image", and the shot does not write this line either.

Workflow

First check the "input reference image" of each shot. When the user has not manually named images, search for real images in the user input, "production results", and media already pointed to by the document, based on this shot's frozen keyframes and the consistency needs of visible characters, locations, and key props. Bind automatically only after checking content and purpose; do not guess images from similar file names.
When matching images already exist, within the current request first have the shot owner refresh in place the REF-... slots of the affected shots, then return to this stage; bind only the images this shot truly needs to maintain, and do not fill slots based on asset count. In the same refresh, if these shots are still missing the required field "visual basis" (old documents often lack it), backfill it together according to the already finalized frozen keyframes— both read the same frame, and separating them into two rounds will only leave shots with bound images still unable to state clearly who is in the frame. New slots must use verbatim REF-<slot>（顺序：<n>）· <项目相对路径>《<中文名称>》（用途：<用途>；控制：<范围>；不得控制：<范围>）；用途 to take one of 身份|造型状态|地理|构图|尺度|效果|起始帧|结束帧|风格 , and must not invent a slash-separated form. For multiple images use only ； Connect, and after writing, first read it back in the format, then copy it into the video document.
When necessary images are still missing, if there are none at all, write the storyboard as "None (reference images to be supplied: ...)"; if only some are missing, keep the already verified REF-..., and append "; reference images to be supplied: ..." at the end. List the missing shots, image purposes, and the corresponding IMG-... / frozen keyframes in one go, then stop before the video prompt. Do not fabricate paths, do not treat image prompts as images, and do not silently downgrade to text-to-video. When reporting gaps, state clearly all three paths the creator actually has, not just "generate reference images" and "switch to text-to-video": put existing images into the project and bind them as REF-...; have the creator generate images in their own tools and for this round first write it as PLAN-... an image attachment plan; or explicitly switch to text-to-video. The gap list itself is the draft of that image attachment plan; write clearly shot by shot which images are needed and what each one governs.
Only when the creator has explicitly chosen not to use reference images, record the storyboard as "None (creator has explicitly chosen text-to-video)". After preparation is closed, copy it as-is into the storyboard's "input reference images", and mark accordingly "generation method: image-to-video" or "generation method: text-to-video". This field contains only these two exact values; do not append model names, dialects, or explanations such as full-reference; keep model explanations in the body structure. Whether to use first frame, first and last frames, or multi-slot reference is read from each slot's 用途 combination: only one 起始帧 uses first frame,起始帧 plus 结束帧 uses first and last frames, 起始帧 when present together with character/location/prop images, or when only these images are present, uses multi-slot reference. The matched model dialect is responsible for translating this combination into its own asset roles and body structure; for models where multi-slot reference and first/last frames are mutually exclusive in the same request, the starting frame also follows the whole group into multi-slot reference and is not split into two sets of inputs.
Verify that this shot's duration falls within the native range of the target model profile; if it is not within the range, return to storyboard revision. Do not leave it for the adapter to error out after writing the prompt, and do not secretly change the duration in the body. When there is dialogue, VO, or OS, also verify the timing basis in the storyboard's "sound" against the actual available vocal window; if the basis is missing, follow dialogue budget to make the judgment. Correct addition of the second segments does not mean the lines can be fully spoken.
Write the immovable frame: image-to-video is locked by the reference frame and visual settings; text-to-video writes the necessary static facts into "static visual anchors" and copyable body text.
Read the storyboard's "starting point → single action → endpoint" as-is, and confirm that the state transitions of the character, both hands, and held objects all occur within the shot.
Write "static anchors → starting point → single action (trigger, contact, landing point) → secondary reactions → camera movement → sound → endpoint"; write camera movement according to the storyboard's "camera movement" line, and segment by seconds when necessary; the next shot continues from this endpoint, without creating another off-screen transition.
Delete repeated descriptions only when the actual reference frame has already explained the static appearance; the locked aspects of the continuity lock in "Visual Settings.md" are not among those that may be deleted.
Write the confirmed relationships that restrict the action space as positive visible states running through this shot, and arrange only action channels allowed by the relationships; write a change process only when the storyboard explicitly changes the relationship. Prefer visible states to explain boundaries, and retain explicit exclusion conditions when necessary.
Per-shot text obligation. Unless the script or creator explicitly requires otherwise, the main text must directly state, in the parsed video prompt language for this shot, that no non-diegetic subtitles, captions, or dialogue text overlays are generated at any time; it cannot merely state the abstract "text layer is empty"; this default is independent of diegetic text such as flags, signage, and screens. Diegetic exact_readable Specify verbatim the characters, language, carrier surface, and position; do not use a global no-text override.
When sound is in the same track, write the permitted vocalizations as a complete timeline by source, speaker, spoken language, and order; each line runs from the first character to the last character of the original text, and verbatim dialogue appears only once in the replicable main text of a single shot. The visual action paragraph only references the sound event by time and speaker, and does not copy the dialogue again. Outside the events, return to the declared ambient sound or silence. When the Brief, script, or sound direction has already explicitly prohibited foreign-language lead-ins, filler words, extra voices, or repetition, put theseconfirmed vocalization boundariesin the complete timeline and consolidate them once at the end; do not rely on "only includes" to make the execution side guess, and do not add new prohibited categories on your own. The timeline summary lists only the sound layers actually present in this shot. For shots without dialogue, clearly state the background sound or silence and its coverage interval; mouth movements are performed according to the script, and closed mouths are not enforced by default. Non-diegetic music is declared according to the target dialect, and when needed, explicitly write empty.
When finalizing, check whether the actions, sound, and constraints are clear and compatible. Positive states help express the goal, and necessary negative conditions may be retained; do not judge errors by the number of negative sentences. Confirm that it does not cross into the next shot, does not require mutually exclusive actions, and does not cause characters or props to teleport or deform.
If the user wants the whole episode, complete the whole episode; shot groups are only internal batches.
When adjacent formally generated segments are in the same continuous time-space and the target profile supports continuation, choose "continue from the previous segment's generated result" by default; one shot one generation only determines that one shot is loaded each time, and does not mean each shot is independently text-generated. Only when the creator explicitly requests independent regeneration, or the upstream has accepted a time/place jump, should the chain be restarted. After choosing continuation, this is the only formal execution route for that chain. From the second segment onward, simultaneously bind the previous segment's actual videoand the actual tail frameobtained from that video: the video is responsible for action, rhythm, and sound continuity, and the tail frame is responsible for the visible starting point of the next segment. The specific material role and continuation wording are determined by the matched model dialect; do not directly translate general responsibilities into vendor fields. When the real artifacts have not been obtained, mark subsequent segments at 视频提示词.md as "pending continuation, not submittable", state clearly that these two real inputs are missing, and retain the draft content for the next segment; do not downgrade it to an independent text-to-video generation, and do not use planned tail frames, keyframe prompts, or textual descriptions to impersonate relay inputs. Per-shot MOTION-... remains the member boundary and content basis, not another submission scheme parallel to the formal container route.

Prompt requirements

Start from a visible starting point and end at a verifiable endpoint.
The replicable main text fully implements the storyboard's "starting point → unique action → endpoint"; handoffs, changes of hands, and prop placement are not omitted across shots.
The subject, action object, direction, amplitude, speed, and sequence are clear.
Performance is written as actions and reactions, not as abstract results that cannot be executed.
Camera movement and character action do not compete for attention; complex camera movement must have a narrative reason.
Sound only includes the dialogue, voice-over, ambient, or transition responsibilities needed for that shot, and lyrics are not added without authorization.
Non-diegetic subtitles, captions, and dialogue text overlays are not generated by default; the body directly declares that these contents are not generated at any moment in the entire shot, without replacing execution instructions with abstract workflow terminology. Only enable them when the script or the creator explicitly requires it. Precise diegetic text is a separate allowed set; it does not disappear because subtitles are turned off, nor is unapproved text allowed to mix in.
For dialogue/VO/OS on the same track, simultaneously write the verbatim text, spoken language, and vocal boundaries; do not add sound sources or duplicate restrictions that the source does not have.
Shots without dialogue likewise state the background sound or silence; the timeline summary is consistent with the actual sound layer. When H3 has no non-diegetic score, write non_diegetic_music: N/A。
In the body, text enclosed in quotation marks or <d> Chinese must come verbatim from "Script.md"; diegetic signage text may come from "Visual Settings.md"; diegetic text and sound written by the storyboard itself also count.creator_markdown_check.py will verify quotations of four or more characters (VID-25). If the dialogue is not in the script, go back to the writing skill to revise the script; do not change it in the prompt.
Upstream continuous relationships are expressed through persistent states and allowed action channels, avoiding duplicate restrictions or conflicting actions.
For the continuity locks hit by this shot, the lock surfaces appear verbatim in the copyable body; cross-shot invariant facts such as color, material, and form do not rely on model memory.
The body can be copied directly and contains no placeholders, process descriptions, file paths, or QA conclusions.

On-demand knowledge

By default, read only this SKILL, the current storyboard, and necessary visual facts. When encountering the corresponding problem, open only one:

Stage boundaries and rule grading:Stage contract
Minimal recipe for single-shot motion body:Motion recipe
Explicit action syntax required for external execution:Production prompt syntax
Multi-person performance, lip sync, and precise timing:Performance and action timing
Camera movement, sound, and continuity between adjacent shots:Cinematography and sound continuity
Action load, deformation, and generation risk:Generatability
Capabilities of the target execution end such as duration, reference, and same-track sound:Target model capability profile
Chinese storyboarding, asset, and sound syntax of Seedance 2.0:Seedance 2.0 dialect
Long-form narrative, timestamp, and task type syntax of Seedance 2.5:Seedance 2.5 dialect
Structured body, reference modes, and per-cut keyframe container of MiniMax H3:MiniMax H3 dialect
Chinese body, asset numbering, and multi-shot time segments of Wan 3.0:Wan 3.0 dialect
Delivery methods for multi-shot containers or static comic dramas:Delivery forms
Boundary checks before completion:Review and examples

When the user explicitly requests timeline music, add a separate section to this document, writing the usage interval, narrative function, entry/exit, dynamic curve, and prohibited items; lyrics must come from text provided or explicitly accepted by the user. When a static comic drama only needs keyframe switching and voiceover, the per-shot video motion prompt may be skipped, and you may proceed directly to production preview.

Completion and production

Each named shot has a clear starting point, main change, timing, camera movement/sound boundaries, and a verifiable endpoint, and is continuous with adjacent shots, then it is complete. Actual generation turns to $short-drama-produce, first show the exact job and obtain explicit confirmation; this skill does not call external services. Once the five creative documents are complete, you may switch to $short-drama to perform a mechanical check of the cross-document structure; content quality is still reviewed by the creator.

Installation and maintenance

Run only when installing, upgrading, or troubleshooting python3 scripts/selftest.py。