---
name: aigc-seedance-prompt
description: Use when the user asks for Seedance, Doubao, or Dreamina video prompts, text-to-video, image-to-video, reference-based generation, video edit, extension, shot bridge, prompt optimization, duration compression, lip sync, or scene continuity.
---

# AIGC Seedance Prompt

## Workflow

Act as a director and Seedance series prompt engineer for connected animation segment production. Infer the scene's real creative goal, viewing priority, emotional beat, rhythm, spatial relationship, and shot organization, then turn that judgment into a prompt Seedance can execute.

CHECKPOINT - Asset And Task Gate:

- For pure text-to-video, proceed from the written brief.
- If the request names a project, script, storyboard, episode/scene/shot identifiers, shot range, or project package, route to `aigc-script-context` first unless the user explicitly asks for a standalone text-to-video prompt.
- For image-to-video, video editing, video extension, reference-based generation, or shot bridging, the actual image/frame/video/reference asset must be present before writing the final prompt.
- If the user only asks why a frame looks weak, route to `aigc-visual-diagnose`.
- If the user asks whether a frame can enter video, route to `aigc-shot-diagnosis-pipeline`.
- If the user only asks to make prompt language more natural, route to `aigc-natural-language-prompt`.
- If the creative idea has no shot purpose or visual strategy yet, route to `aigc-creative-director`.

1. Identify the task type: new text-to-video prompt, image-to-video prompt, reference-based prompt, prompt optimization, diagnostic review, video edit, video extension, or shot bridge.
2. For image-to-video, reference-based generation, video editing, video extension, or shot bridging, confirm the actual source image/frame/video/reference asset is present in the current context before writing a prompt. If the task depends on a missing asset, do **not** write from a text-only handoff summary; ask the user to re-attach the asset and confirm the frame first. A handoff summary is context, never a substitute for the asset. For pure text-to-video, proceed without blocking.
3. When using a shot context card or handoff, preserve locked facts, source priority, character identity, shot function, previous state, and risk notes before drafting; do not reinterpret them unless the user gives a newer instruction. Map context-card fields directly: `源优先级` -> source priority, `当前画面事实` and `参考图角色` -> locked facts, `人物表演` -> performance constraints, `上一镜承接` -> previous state, `本镜剧情功能` -> shot function, and `禁止偏移` -> risk notes.
4. Identify the medium and style target before drafting: live-action photoreal, 2D animation, stylized 3D, illustration, game cinematic, product render, or mixed media.
5. Judge the whole segment structure first: one-shot or multi-shot, task continuity, reference roles, output mode, and per-shot attention load.
6. Ensure the single segment can generate well before optimizing long-form continuity: subject, action, space, camera, emotion carrier, and any necessary continuity anchor must be clear.
7. Apply prompt principles before final wording: translate abstract intent into visible subject, action, space, camera, light, sound, performance beat, and concrete visual change.
8. Apply director judgment. If the intended shot is too complex to generate reliably, simplify the shot organization while preserving the core expression.
9. Apply Seedance-specific rules for duration, reference asset mapping, shot wording, continuity, video editing, and stability. For later Seedance versions, use the current Seedance 2.0 rules as the default unless the user provides newer constraints.
10. Write each shot as one natural execution paragraph starting with `镜头N：x秒，景别。`: keep shot number, duration, and shot size explicit and end the lead-in with a period, then continue in flowing Chinese sentences that include only the camera, movement, action path, camera-subject relationship, and continuity details needed for this shot to generate clearly. Do not fill fixed slots.
11. For long-form work, preserve segment function, starting state, and the next segment handoff only when those details affect the next generated clip. Do not force an ending-state sentence when the action can naturally continue.
12. Before outputting any final prompt, run an internal AI-flavor and logic scan: remove template voice, abstract boosters, decorative connectors, forced summary endings, unsupported off-screen causes, and parameter stacking. Use `aigc-natural-language-prompt` only when the user asks for natural-language cleanup or the language problem is the main task.
13. Output the final Seedance prompt in one and only one fenced code block. Put any judgment or recommendation outside that code block.

### Failure Branches

- If duration is too short for the requested actions, compress the action chain or split the shot; do not squeeze multiple locations, reveals, and dialogue beats into one unreadable clip.
- If a reference asset could be used for identity, environment, style, and composition at the same time, assign explicit reference roles before drafting.
- If camera movement and subject movement conflict, keep the instruction that best preserves readability and remove the contradiction.
- If the requested prompt depends on a missing image/video reference, stop and ask for the asset instead of writing from a text-only summary.
- If a Yellow production decision is carried over from `aigc-shot-diagnosis-pipeline`, state the risk outside the code block before drafting only when the user explicitly accepts that risk.
- If dialogue is requested but the mouth is not visible or the shot is too short for lip sync, adjust framing, reduce dialogue, or state the risk before drafting.
- If the user demands one continuous shot but the action chain requires hard cuts, locations changes, or simultaneous reveals, preserve the strongest beat and simplify the rest.
- If a simple request starts accumulating unnecessary reference, camera, and avoid sections, collapse it back to total duration, one shot paragraph, and only essential avoid notes.
- If multiple references, strict character counts, foreground occlusion, focal length, or composition percentages are present, use the complex reference/composition structure and map reference roles before writing shots.

## Output Modes

- **Default**: give 1-2 concise judgment bullets only when they materially improve the prompt, then provide the final prompt.

- **Direct draft**: when the user signals they only want the prompt, output only the final fenced code block — no preamble, no postamble, no judgment bullets. Trigger phrases include: `直接出稿`, `只给提示词`, `只要 prompt`, `不用解释`, `直接给我`, `不要解释`, `prompt only`, `just the prompt`, `don't explain`, or any clearly equivalent phrasing.

- **Diagnostic mode**: only when the user explicitly asks to optimize, inspect, compare, or diagnose an existing prompt. Use this fixed output order:
  1. **当前问题** (2-4 short bullets, each one sentence, naming the specific failure: missing emotion carrier, compound camera movement, abstract style label, bare reference label, etc.)
  2. **改进 prompt** (one fenced code block, only the prompt body)
  3. **关键修改** (1-3 short bullets, each pointing at one change: `把 X 改成了 Y,因为...`)

- **Creative guidance**: if the user only has a vague idea and asks how to design it, provide the key problem, 2 practical directions, and a recommended direction. When enough information is available, also provide the final prompt at the end.

## Output Structure Selection

Choose the lightest structure that will keep the prompt executable. Do not force a heavy template onto a simple shot, and do not compress a reference-heavy or composition-critical request into one loose paragraph.

Use **simple structure** when the request has one clear subject, one space, one action chain, no complex reference mapping, and no hard composition ratio:

```text
本视频总时长 X 秒，单镜头。
[一句话写清楚风格、主体、空间、声音规则]

镜头1：X秒，景别/机位。
[镜头在哪里拍，先看到什么，主体做什么动作，最后停在什么状态]

规避：
[只写最容易跑偏的1-3点；没有明显跑偏风险时省略]
```

Use **complex reference/composition structure** when any of these are present: multiple reference images or videos, strict character count, reference-role separation, foreground occlusion, voyeur/hidden-camera framing, explicit lens/focal length, precise subject position, composition percentage, monster/prop/environment reference separation, multi-shot continuity, or a user-provided structured prompt they want preserved.

```text
本视频总时长 X 秒，单镜头 / N个镜头。
[全局风格、人物数量、声音规则、不可新增内容]

参考图引用：
@图1（角色 / 外貌 / 服装参考）作为……参考。
@图2（角色 / 道具 / 生物参考）作为……参考。
@图3（人物位置 / 构图关系参考）作为……参考。
@图4（环境 / 光线 / 材质参考）作为……参考。
@图5（环境 / 道具 / 怪物 / 色调参考）作为……参考。

摄影与构图总要求：
[机位、焦距、景深、前景占比、主体位置、背景层次、运动方式]

镜头1：X秒，景别，机位 / 焦距。
[镜头放在哪里，向哪里看，先看到什么，主体如何进入或动作如何发生，最后停在什么状态]

镜头2：X秒，景别，机位 / 焦距。
[多镜头时逐镜头写清楚动作衔接]

规避：
[不要新增人物，不要正面海报感，不要广角变形，不要参考图混用，不要改变核心构图]
```

In both structures, section headings are allowed inside the fenced prompt only when they help the generation model parse global constraints before shot execution. Keep headings short and concrete. The shot body still must use natural Chinese sentences with visible subjects, verbs, spatial relationships, and action order.

## Prompt Detail Budget

Use the shortest Chinese wording that preserves the user's intent and Seedance generation stability. Length is decided per shot or action unit, not per whole video. Do not reveal the `simple` / `standard` / `complex` labels in the final prompt.

## Short Duration Compression

For clips of 15 seconds or less, compress before drafting:

1. Keep one main location, one main action chain, one camera strategy, and one clear ending beat.
2. Reduce simultaneous subjects, background business, dialogue lines, and camera moves before adding detail.
3. Preserve the user's must-have elements first: protagonist identity, core gag or emotional beat, reference role, spoken line, and ending action.
4. Move resolution, frame rate, lens brand, aspect ratio, and other platform settings out of the prompt body unless the user explicitly asks to include them.
5. If the request cannot fit the duration, say what was reduced in one short note outside the final code block.

### Simple Shot

Use one short Chinese sentence when the subject, action, and continuity are obvious:

- One clear subject or one clear edit.
- One main action with no layered blocking.
- Little risk of confusing reference roles, spatial relationships, or emotional intent.
- Atmosphere, light, sound, and camera do not change the user's meaning.

For simple shots, write only the needed subject, action, and essential continuity. Do not add extra camera, lighting, mood, material, sound, or stability language just to make the shot look professional.

### Standard Shot

Use one to two compact Chinese sentences when the shot needs moderate control:

- A clear subject plus atmosphere, space, expression, prop, or reference-image role.
- A simple action whose meaning depends on gaze, posture, timing, contact point, light, or environment.
- A one-shot segment that still needs duration, audio policy, and stable identity.

For standard shots with any performing subject, write the key body part, gaze target, contact point, or expression change that makes the performance readable. Mention only the details that reduce ambiguity, improve generation reliability, or make the acting beat visible.

### Complex Shot

Expand only when detail prevents likely misunderstanding:

- Multiple subjects, layered actions, or action handoffs.
- Foreground/midground/background relationships, occlusion, entrances/exits, or position changes.
- Camera movement, reveal order, transition logic, or continuity across shots matters.
- Reference assets have overlapping roles or could be mapped incorrectly.
- The user is choosing between a conservative stable result and a more ambitious visual effect.

For complex shots, write clear subject, space, action order, camera behavior, and the continuity anchor that prevents likely misunderstanding. Keep the detail purposeful; do not pad with generic quality terms.

## Positive Direction Discipline

Treat every generation attempt as a fresh first run for the video model. When the user is reacting to a failed clip, convert the failure into the desired visible shot instead of carrying a long corrective blacklist into the next prompt.

Why this matters: Seedance, Hailuo, and similar video models do not know that a new prompt is a repair of a previous failed generation unless the actual prior clip is provided as an edit source. Negative lists can still make the unwanted concept salient inside the fresh prompt ("do not show X" still names X). Positive, visible staging tells the model what to render now: where the camera starts, what subject is in frame, how the action moves, and what state the shot should end on.

- Replace failure notes with positive staging. Use `the camera starts at table height beside the old desk, watching the fox students from their side-front` instead of `do not use the previous reference angle`.
- Use avoid notes only for hard, current-run constraints: visible text/UI from references, forbidden identity changes, unsafe content, or a small number of known visual artifacts. Do not build an avoid section from every previous model mistake.
- Keep the positive instruction denser than the negative instruction. If an avoid note is needed, make sure the corresponding desired action, composition, or ending state has already been stated clearly in the shot body.
- For multi-shot continuity, describe the inherited visible state at the start of each shot and the concrete ending state of the previous shot. Do not assume the model remembers an earlier failed generation.

## One-Shot vs Multi-Shot Decision

Use one continuous shot when uninterrupted performance, immersion, POV, or a single action path is the main expression. Use multiple shots when the request needs separate locations, distinct reveals, dialogue coverage, or action beats that would overload one generation.

If the user asks for `一镜到底` but also asks for multiple incompatible beats, keep `一镜到底` only when the main action can stay in one location and one camera path. Otherwise explain the tradeoff briefly and draft the most stable version.

For one-shot prompts, use only `镜头1：x秒，景别。` and describe the internal beat order in natural prose. Do not list hard cuts, montage transitions, or separate camera resets inside a one-shot prompt.

## Camera Movement Detail

Camera movement detail should also scale by shot complexity.

- For simple shots, omit camera movement unless it is central to the request. Use `固定机位` when stillness improves stability, symmetry, or quiet atmosphere.
- For standard shots, write one main camera movement and its purpose, such as `缓慢推近` for expression or object detail, `横向移动` for spatial reveal, or `跟随拍摄` for a clear subject path.
- For complex shots, specify the starting frame, subject relationship, movement path, reveal order, and next-shot anchor only when these details are needed to prevent confusion.

Do not stack multiple major camera moves in one shot unless the user explicitly asks for that complexity. Avoid combining push-in, pan, tilt, crane, zoom, and handheld movement in the same shot.

## Medium And Style Branch

Identify the intended medium before choosing prompt vocabulary. Do not let live-action cinematography language override a non-photoreal target.

- **Live-action photoreal**: use grounded lighting, exposure, lens distance, atmosphere, and material response terms when they help.
- **2D animation / illustration**: prioritize clean silhouette, line/design consistency, readable color blocks, stable character shape, and rhythmized pose changes. Avoid film grain, IRE values, film-stock names, and photoreal skin/material language unless the user asks for them.
- **Stylized 3D / game cinematic**: prioritize stable model identity, readable staging, clean material hierarchy, soft but controlled lighting, natural ear/tail/cloth/hair motion, and clear action timing. Use engine or game-cutscene style terms only as global constraints; shot bodies still need concrete actions and spatial relationships.
- **Product / object render**: prioritize shape accuracy, logo/mark preservation, material response, contact shadows, reflection control, and camera-object relationship.

If the user's references imply a medium, inherit that medium without renaming it. If style is unspecified, keep the prompt neutral and execution-focused.

## Shot Line And Execution Body

Each shot must read as a director's shooting note written for a real crew — natural Chinese prose, not a slot-filling template. Hold this discipline tightly: comma-chained parameter lists are the most common failure mode of generated Seedance prompts and break the script-like read the model handles best.

1. **Shot lead-in**: start each shot with shot number, duration, and shot size in one short structured opening, such as `镜头2：2秒，中近景。` End with a period (not a comma or 顿号) so the prose body starts on a clean break.
2. **Composition sentence**: write angle, camera position, main camera behavior, and visual focus as **one complete sentence with a verb**, not as a chain of bare parameter phrases. Prefer `侧前方低角度的固定机位俯视云海` over `侧前方低角度固定机位，俯视云海，仰角拍摄`.
3. **Execution body**: continue in flowing natural language, the way a director talks a shot through to the camera team. Describe what the camera sees first, where the subject starts, how it moves through frame space, how it passes the camera, where it exits or lands, what the environment does in response, and only the state the next shot truly needs to inherit.

Within the execution body, use temporal and spatial connectives only when they clarify how beats connect rather than pile up: `随后`, `紧接着`, `与此同时`, `此时`, `下一刻`, `画面中`, `镜头前`. Use `最终` sparingly, only when a concrete action endpoint matters. Break long stretches with periods — if a single sentence runs past 4-5 comma-separated clauses, split it.

Natural prose is not permission to over-expand. For simple shots, keep the prose short; add only the verbs, connectives, and visible controls needed to make the action executable.

Avoid bare parameter strings like `中景，固定机位，主角站在画面中央，背景为雨夜`. Translate them into a sentence with a verb: `中景，固定机位从正面拍摄主角，他站在画面中央，身后是雨夜街口`.

Use Chinese-only shot and camera wording in the final prompt by default. Do not add English shot abbreviations or English camera terms such as `EWS`, `MCU`, `locked-off`, or `tracking` unless the user explicitly asks for English labels. Platform-facing reference labels such as `Image 1` are optional and should be used only when the target UI needs them.

Compare the same shot written both ways:

- **Too list-like**: `镜头2：2秒，中近景，侧前方低角度固定机位，仙剑从画面上方高速下落，剑尖朝下，冷白拖尾先穿过云雾，再从镜头前景掠过并向下冲出画面。`
- **Natural prose**: `镜头2：2秒，中近景。侧前方低角度的固定机位仰望云海，一柄仙剑从画面上方高速坠下，剑尖朝下，拖出一道冷白色的尾焰先穿过云雾。紧接着剑身从镜头前景一闪而过，最终向下冲出画面。`

Both carry the same information, but the second reads like an actual shot description rather than a parameter dump.

For action, VFX, object, flight, impact, or transformation shots, include these controls when relevant:

- **Entry and exit**: where the object appears, passes, exits, lands, or ends.
- **Camera relation**: whether the camera is fixed, follows, tilts, pushes, tracks beside, tracks behind, or holds while the subject crosses frame.
- **Path and speed**: vertical drop, horizontal sweep, diagonal crossing, acceleration, pause, impact, rebound, or continuation.
- **Environmental reaction**: clouds torn open, water displaced, debris falling, light blooming, shadow moving, mist clearing, or waves rising.
- **Continuity anchor**: posture, gaze, object position, movement direction, light state, or camera position only when the next shot depends on it.

Avoid contradictions. If the camera is `固定机位`, the subject may cross or exit the frame, but the camera should not also follow. If the camera follows a sword, vehicle, character, or energy trail, write `跟随拍摄` and specify whether it follows from above, behind, side, front, or close to a specific body/object part.

## Performance And Blocking Detail

For shots with a human, animal, anthropomorphic character, creature, hand, face, or other performing subject, make the performance accurate and vivid by default. Scale detail by shot complexity, but do not reduce performance to a bare state label such as `sad`, `happy`, `stares`, or `walks`.

- For simple shots, include the one visible action and one readable performance cue, such as gaze target, hand contact, posture shift, or expression change.
- For standard shots, write a short action chain: starting pose, active body part, contact point, movement direction, gaze target, and continuity anchor when relevant.
- For complex shots, add action order, eye-line logic, foreground/background blocking, and how the shot begins from the previous shot's ending pose or gaze.

Prioritize performance controls in this order:

1. **Body and contact**: name the active body part and object/body contact point, such as `right hand holds the bowl base`, `left thumb rubs the bowl rim`, or `both paws rest on the table edge`.
2. **Gaze and attention**: state who or what the subject looks at, the gaze path, and whether the eyes return, avoid, or fail to focus.
3. **Expression transition**: describe the change, not only the final mood, such as `relaxed smile fades into concern`.
4. **Movement path**: describe direction and endpoint, such as `leans forward half a body length`, `raises the bowl toward the table center`, or `turns from the empty pot to the boy`.
5. **Continuity handoff**: when shots connect, state how the new shot inherits the previous pose, gaze, or action.

## Dialogue And Lip Sync

When the user requests dialogue, speech, lip sync, or visible mouth movement:

- State who speaks, the exact spoken line, and whether the mouth is visible in the frame.
- Keep dialogue short enough for the duration. For 15 seconds or less, prefer one or two short lines.
- Give the speaking subject enough stable face time; avoid hiding the mouth behind fast camera motion, back view, heavy occlusion, or a cutaway.
- Keep default audio policy unless the user asks otherwise: no music, no voiceover, no subtitles, and no dubbing; only diegetic speech and necessary action/environment sound.
- If subtitles are requested, include them only when the user explicitly asks and keep them out of the prompt otherwise.

When a reference image is only used for environment, style, or identity, say that explicitly with a soft reference role. Preserve any literal platform reference anchor that starts with `@`, including ordered labels such as `@图1` and file-name anchors such as `@庠序场景.png`; natural cleanup must not remove `@` or rewrite the anchor as `参考图1`, `图1`, or a plain file name. If the source only says `图1` / `参考图1` without `@`, normalize it to the platform label the user is likely using, such as `@图1`. If the written shot design should override the reference image's camera angle or composition, write that priority in the prompt, e.g. `@图1（房间、道具、光线和材质参考）作为空间质感参考；镜头位置与构图以文字描述为准`。

## Strong And Weak Prompt Words

Use strong control words before weak descriptive words. Expanded prompts should add control information, not decorative adjectives.

Strong words are visible, executable, and reduce ambiguity:

- subject identity, position, action verb, action order
- active body part, contact point, gaze target, expression transition
- camera movement, spatial relationship, useful continuity anchor
- reference asset role and continuity constraint

Weak words are mood, taste, or atmosphere helpers, such as cinematic, lonely, mysterious, premium, tense, dreamy, epic, or beautiful.

Weak words may be used only when anchored to visible carriers. Do not rely on weak words alone.

- Bad: `镜头很电影感，氛围孤独。`
- Good: `镜头固定在雨夜街口，人物独自站在路灯下，身后街道空旷，积水反射冷白灯光。`

For each shot, write strong controls first, then add weak atmosphere only if it changes the intended image or emotion.

## AI-Flavor And Logic Pass

Before returning a final prompt, do a concise language and logic scan. For general natural-language rewrites or teaching requests, use `aigc-natural-language-prompt`; inside this Seedance skill, apply the same standard while preserving Seedance-specific duration, reference mapping, and shot-bridge rules. Natural does not mean casual or vague; it means the shot reads like a coherent visual action instead of prompt-engineering syntax.

- Each sentence should name a visible subject and a real verb. Replace noun piles such as `中景、冷色、孤独、电影感` with filmed relationships such as `中景固定拍摄，人物独自站在冷白路灯下，身后的街道空旷`.
- Keep cause and sequence readable. Use `先`, `随后`, `此时`, `最终` only when they clarify the action order; do not add connectors as decoration.
- Use adjectives only after the concrete carrier is clear. `压抑` should become tight spacing, low ceiling, held breath, blocked doorway, heavy shadow, or another visible/audible cue.
- Do not invent an off-screen source or cause. If the source is not visible in the current shot or clearly established by a previous shot, write only the visible result, such as `额前碎发被轻轻吹开`.
- When a shot cuts from a previous view, state the current frame relationship only when it prevents confusion, such as `从上一镜头的远景切到桌前右侧中近景`.
- Remove prompt-flavored filler before output: `高质量`, `大师级`, `极致细节`, `电影质感`, `氛围拉满`, `高级感`. If the idea matters, translate it into camera, light, blocking, material, sound, or movement.
- Remove AI-flavored structure before output: "不只是...更是...", rule-of-three padding, generic conclusions, decorative `最终`, and any sentence that explains creative intent instead of controlling the visible shot.
- Do not write a closing state merely to make the prompt feel complete. Add a continuity anchor only when it prevents confusion in the next shot or connected segment.
- Check that the shot can be acted or animated. If a phrase cannot be seen, heard, performed, lit, framed, or timed, rewrite it before placing it in the final code block.

### Confirmation Policy

Default to self-judgment and produce the prompt directly. Ask the user before finalizing only when ambiguity changes the actual creative direction or generation strategy:

- The main character or reference asset role is unclear.
- Two reference images or videos conflict in identity, style, action, or setting.
- The same request could reasonably become different shot designs.
- The prompt requires a tradeoff between stability and a complex visual idea.

Do not interrupt the user merely to ask whether an ordinary simple shot should be short.

## References

Load only the reference needed for the task:

- Read `references/prompt-principles.md` when the task needs creative completion, shot design, style handling, long-form continuity, prompt efficiency, or translation from abstract intent into visible action.
- Read `references/single-segment-quality-control.md` before finalizing new or optimized video prompts when a scene has multiple subjects, large actions, occlusion, complex blocking, unclear camera movement, or weak continuity.
- Read `references/seedance-2-rules.md` for all final prompt drafting, reference image/video handling, text-to-video, image-to-video, video edit, video extension, shot bridge, and official Seedance 2.0 constraints.
- Read `references/task-patterns.md` when the request targets a specific format such as product ads, UGC, creative VFX, dialogue drama, music beat sync, one-take, educational visualization, or multi-video fusion.
- Read `references/examples.md` only as an optional calibration aid when the output shape is unfamiliar. Do not load examples for routine Seedance prompts.
- Use `aigc-natural-language-prompt` as the cross-skill standard when the user is primarily asking how to make prompt language more natural, director-style, visible, and non-parameterized rather than asking for Seedance-specific final drafting.

## Prompt Contract

Write the final prompt in Chinese by default. Do not include English shot-size abbreviations or English camera movement terms unless the user explicitly asks for bilingual camera labels.

The final prompt should normally start with duration and scene overview, then write each shot as a natural Chinese paragraph beginning with `镜头N：x秒，景别。` and continuing in flowing sentences with verbs and connectives. For simple requests, use the simple structure from Output Structure Selection: total duration and scene setup, one shot paragraph, and only necessary avoid notes. For complex reference-heavy or composition-critical requests, use the complex structure: total duration and global constraints, reference-role mapping, camera/composition requirements, shot paragraphs, and avoid notes. For a true one-shot design, use only `镜头1：x秒，景别。` and describe the internal continuous action order in prose. Default audio policy: no music, no voiceover, no subtitles, and no dubbing; keep only environment sound, action sound, and necessary diegetic sound.

Keep the prompt visible, executable, and stable: one main action per shot, clear spatial relationships, clear subject identity, clear reference asset roles, readable performance beats, and no internal reasoning or rule explanations inside the final code block. Use production shorthand and abstract taste words only for internal planning; translate them into visible shot, action, light, space, body/contact detail, gaze, expression transition, camera-subject relationship, action path, environmental reaction, sound, and only the edit-handoff details that matter.

### Chinese Shot And Camera Terms

Use Chinese terms consistently across all shots:

**景别**：大特写、特写、中近景、中景、中远景、远景、大远景。

**角度/机位**：高空俯拍、低角度仰拍、侧前方角度、正面角度、背面角度、贴近海面、贴近主体上方半侧、远处海平面视角。

**运镜**：固定机位、手持、缓慢推近、推近、后拉、横摇、垂直下摇、跟随拍摄、横向移动、升降、环绕。

Use only one main camera movement per shot. Put the shot size, camera position, and movement near the beginning of the shot paragraph, then explain the real movement in natural language.

## Common Failures To Avoid

These are the failure modes most likely to break a Seedance prompt. Scan the final draft for each before returning it:

- **Parameter-list writing style** — e.g. `镜头1：5秒，中景，固定机位，主角站在画面中央，背景是雨夜，主角抬头，雨水打湿肩膀`. Comma-chained slots without verbs break the script-like read Seedance handles best. Rewrite as flowing sentences with verbs and connectives, ending the structured lead-in with a period: `镜头1：5秒，中景。固定机位从正面拍摄，主角站在画面中央，身后是雨夜的街口。他抬起头，雨水打湿了他的肩膀。`
- **Writing aspect ratio, resolution, or frame rate inside the prompt** — these belong in the platform UI, not the prompt body. Only include if the user explicitly asks.
- **Bare reference labels** — e.g. `@图1 走向画面中央`. Always attach a semantic role: `@图1（白衣少年角色参考）走向画面中央`.
- **Unscoped reference intent** — e.g. `参考 @视频1`. State whether the reference serves as camera movement, action, edit rhythm, effect behavior, sound, or character performance reference.
- **Environment reference overriding shot design** — if a scene image is only an environment reference, state that it does not set camera angle, framing, or starting image; otherwise the model may copy its composition.
- **Compound camera movement in one shot** — e.g. mixing push, pan, and tracking-like following in a single shot. Pick one main movement; if multiple are needed, split into multiple shots.
- **Conflicting camera or edit instructions** — e.g. requesting `固定机位` and `环绕镜头` in the same shot, or `一镜到底` while also listing hard cuts. Resolve the priority before drafting.
- **Duration-complexity mismatch** — e.g. placing several locations, transformations, dialogue beats, and camera moves inside 4-5 seconds. Reduce actions, split shots, or extend the segment.
- **Inventing style labels not established by the user or references** — e.g. `三渲二`, `UE5风格`, `照片级写实`, `cel shading`, `cinematic`. If style is unspecified, write neutral execution quality only.
- **Negative tag lists** — e.g. `不要模糊，不要变形，不要失真`. Replace with positive boundaries: `主体保持清晰可辨，身体结构自然，动作物理合理`.
- **Internal reasoning inside the fenced code block** — rule names, planning notes, or explanations of why a choice was made. The code block contains only the executable prompt body.
- **Abstract taste words without a visible carrier** — e.g. `氛围感强烈`, `极具张力`. Translate into specific gaze, posture, light direction, spacing, or sound.
- **Flat performance labels** — e.g. `he looks sad` or `the fox acts funny`. Replace with body part, contact point, gaze path, expression transition, movement direction, and only the anchor needed for the next beat.
- **Music, voiceover, subtitles, or dubbing** unless the user explicitly asks. Default audio is environment sound, action sound, and necessary diegetic sound only.
- **Plot synopsis** — describing what happens before or after the clip, character backstory, or narrative arcs the camera cannot see. Stay inside what the camera frames during the segment duration.
- **Identifiable real people, celebrity likenesses, trademarked characters, or protected IP** — keep generic or ask the user for rights-safe handling.
- **Treating reference-image prompts like text-to-video prompts** — when references exist, preserve literal `@...` anchors such as `@图1` or `@庠序场景.png`, name what each reference is used for, and state what the written shot overrides.
- **Overusing natural-language cleanup** — do not run a separate cleanup pass unless the user asks or the draft has visible language defects.