# Storyboard Lite > Generate a production-oriented storyboard, draft asset-generation pack, continuity maps, and paste-ready AI video prompts from a script. Use for shot breakdowns, short-drama storyboards, or video prompts, including Seedance 2.5 long-form and multimodal-reference workflows. - Skill: `towardsyoung/storyboard-lite` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add towardsyoung/storyboard-lite` - Raw SKILL.md: https://api.skillmd.com/api/skills/towardsyoung/storyboard-lite/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: towardsyoung (https://skillmd.com/u/towardsyoung) - Updated: 2026-09-22 - Page: https://skillmd.com/skills/towardsyoung/storyboard-lite --- # Storyboard Lite Turn script text into a practical production handoff: reusable asset definitions, scene continuity, a readable storyboard, and video-generation prompt groups. Preserve the low-input workflow while supporting professional asset packs and Seedance 2.5. This skill is standalone. It does not upload media, invent platform asset IDs, or depend on an external project database. ## Inputs Required: - Script text. Optional: - Story type and art style. - Target platform or model, such as `generic` or `seedance-2.5`. - Generation mode: `generate`, `reference`, `extend`, or `edit`. - Preferred group duration or duration mode: `auto`, `short`, or `long`. - Existing asset pack, approved assets, reference media, or real platform `@tags`. - Output format and prompt language. - Whether to include the asset list, draft asset-production pack, storyboard, or video groups. Defaults: - Prompt content language: Chinese, unless the user explicitly requests another language. - Structural headings such as `SCENE CONTEXT`, `ACTIVE REFERENCES`, `LOCATION MAP`, `ACTION TIMING`, and `AUDIO` may remain in English. - Generation mode: `generate` without references, otherwise `reference`. - Duration mode: `auto`. - Output: Markdown with asset list, draft asset-production pack, scene maps, storyboard, and video groups. - Asset IDs: never generated. Do not silently choose a story type, visual style, or target model. Use the confirmation checkpoint below before producing the storyboard or prompts. ## Mandatory Style and Model Checkpoint Work in two phases. ### Phase 1: recommend and pause When the user has not explicitly confirmed both the visual style and target video model in the current request or immediately preceding context: 1. Read the script and identify its period, genre, emotional register, action density, continuity risks, material requirements, and intended viewing context. 2. Recommend one primary visual style and one target video model. If an asset pack must also be generated, recommend the primary asset-image model or model route. 3. Recommend the aspect ratio and duration strategy when they materially affect production. 4. Give a brief script-specific reason and mention only consequential tradeoffs or exposed platform limitations. 5. Ask the user to confirm or revise the recommendation. 6. Stop. Do not output the asset list, asset prompts, scene maps, storyboard, or video prompts in this phase. Use a compact response: ```markdown ## 制作建议（待确认） - 推荐视觉风格：... - 推荐资产模型：...（需要生成资产时） - 推荐视频模型：... - 推荐画幅：... - 推荐时长策略：... - 推荐理由：... 请确认以上方案，或告诉我需要调整的风格、模型、画幅或时长策略。确认后我再继续生成完整资产与视频提示词。 ``` ### Phase 2: produce Proceed only after the user confirms the recommendation or has already explicitly supplied both the visual style and target video model. Treat clear instructions such as “按推荐方案继续” as confirmation. Preserve the confirmed choices as the source of truth throughout assets, storyboard, and video groups. ## Model Profiles Use one capability profile as the source of truth. Do not scatter duration limits through unrelated rules. ### Generic profile - Planning range: 4–15 seconds until a target platform limit is known; this is not a verified model capability. - Prefer multiple consecutive same-scene shots per group. - Use the unified video-group format below for every model. ### Seedance 2.5 profile Read [references/seedance-2.5.md](references/seedance-2.5.md) whenever the user selects Seedance 2.5 or requests its long-duration, multimodal-reference, extension, or timestamp-editing behavior. Core profile: - Single-generation duration: 4–30 seconds. - References: up to 30 images, 10 videos, and 10 audio clips, with up to 50 total materials where the active platform exposes the full model capability. - Supports timestamp control, multimodal reference, motion/camera reference, clay-render guidance, extension, and targeted video editing. - The active platform UI wins if its exposed limits differ. Thirty seconds is a ceiling, not a target. Choose duration from the combined shots and reference complexity. Verify the selected model and active platform limits before finalizing submission-ready groups; do not infer capabilities from a model name or use a different version’s limit. If unavailable, label the limit and grouping as provisional. ## Asset Handoff If the user supplies an asset pack: - Treat approved identity, state, voice, behavior, location GEO, prop, and label data as source of truth. - Preserve real platform `@tags` exactly. - Do not redesign locked assets merely to make the storyboard more decorative. - Report a blocking contradiction if the script requires a state or prop that the locked pack cannot represent. If no asset pack is supplied: - Extract a compact asset list. - Generate a draft `## 资产生产包（草案）` before the storyboard. - Include usable prompts for recurring characters, every location, and plot-critical props. - Include character state variants, fixed voice prompts for recurring speakers, short behavior master profiles, location anchors, GEO maps, and reference roles. - Mark all proposed labels as `建议标识，上传后确认`; do not present them as real platform tags. - Follow the asset-generation and state-separation rules from `$shortdrama-asset-pack` when that skill is available. Otherwise apply the compact rules below. For a long script, create the complete registry once and batch detailed assets or storyboards by episode. Do not silently omit assets. ## Asset Label Contract - `名称` is human-readable. - `建议标识` is a stable project-local label based on the exact script name, not an uploaded asset. - `平台标签` is supplied by the user or platform and must remain exact. - `资产 ID` is never invented. Suggested asset identifiers use the script's exact Chinese or original-language asset name without generic prefixes or transliteration. For example, character `何朝宗` uses `@何朝宗`, location `何朝宗作坊` uses `@何朝宗作坊`, and prop `观音瓷像` uses `@观音瓷像`. Add only a short script-grounded suffix when a visual state must be separate, such as `@何朝宗_守窑`, `@观音瓷像_未烧紧握`, or `@龙窑_黎明开窑`. Use the shortest qualifier needed to avoid name collisions. Base identity and temporary state must be separate: - Character: base, wet, wounded, bloodied, disguised, alternate costume. - Location: day, night, rain, fire, destroyed. - Prop: intact, open, broken, bloodied, concealed, close-up. ## Draft Asset-Production Pack The inline draft uses this minimum structure: ```markdown ## 自动提取资产清单 | 名称 | 类型 | 描述词 | 状态 | 建议标识或平台标签 | 标签状态 | 出镜位置 | |---|---|---|---|---|---|---| ## 资产生产包（草案） ### 全局视觉标准 ... ### 角色资产 #### {角色名}｜{标签} **平台参数：** ... **身份锚点：** ... **资产生成提示词：** ... **固定声音档案：** ... **行为主档案：** ... **状态变体：** ... ### 场景资产 #### {场景名}｜{标签} **平台参数：** ... **资产生成提示词：** ... **固定锚点：** ... **GEO SPATIAL LAYOUT：** ... **状态变体：** ... ### 道具资产 ... ``` Asset prompts must be production-ready. Keep platform parameters outside prompt prose. Character sheets use a large 3/4 portrait, neutral grey background, real untreated skin, live eye catchlights, a headless front wardrobe/body view, and a back full-body view unless the user or target model requires a conventional full-body front view. Locations use a 3/4 perspective, stable anchors, and one primary light direction. Local state changes preserve the approved original instead of causing a full-image redraw. Write all asset-generation, edit, acting, voice, and video-prompt content in Chinese unless the user explicitly requests another language. English structural headings may remain, but their body content defaults to Chinese. ## Scene Continuity Masters Before the storyboard, create one source-of-truth block per recurring location: ```text GEO SPATIAL LAYOUT — {location} CAMERA SAFE SIDE = ... 180° AXIS = ... ENTRY = ... PRIMARY ANCHOR = ... SECONDARY ANCHOR = ... ACTION ZONE = ... BACKGROUND BOUNDARY = ... PRIMARY LIGHT = ... SHADOW DIRECTION = ... ``` Keep the GEO block unchanged across groups in the same location. It describes space, not character action. Shot-level blocking places characters relative to these anchors. Read [references/directing-and-continuity.md](references/directing-and-continuity.md) when generating scene maps, dialogue blocking, acting, physics, multi-shot prompts, or continuity locks. ## Storyboard Format ```markdown ## 分镜表 | 序号 | 画面描述 | 场景 | 关联资产/状态 | 时长 | 景别 | 运镜 | 角色动作 | 表演指令 | 朝向与视线 | 空间关系 | 情绪 | 台词 | 音效 | |---|---|---|---|---|---|---|---|---|---|---|---|---|---| ``` Rules: - Follow script order. Do not add plot events or omit dialogue. - Copy dialogue verbatim. - Every shot has a scene and names all visible characters and plot-critical props. - `关联资产/状态` uses the correct base or state asset; do not combine mutually exclusive states. - `角色动作` begins with `(开篇)` or a concrete carried state such as `(承接上镜：右手仍压住协议，呼吸未稳)`. - `表演指令` describes objective, obstacle, tactic, tactic change, physical business, and visible reaction. Do not rely on emotion labels alone. - `朝向与视线` separates torso direction from gaze target. - `空间关系` uses frame-left/frame-right, foreground/midground/background, landmark distance, and the scene GEO map. - Dialogue duration uses roughly 4 Chinese characters/second for anger, 3 for normal speech, and 2 for sadness, whispering, or weakness, plus about 1 second for reaction when needed. - Non-dialogue shots are usually 6 seconds or shorter unless a deliberate long take needs time to read. - Sound contains only dialogue, diegetic sound, Foley, and environment unless the user explicitly requests music. - Merge shots that repeat information; split shots that overload one short interval. ## Duration Routing Use these ranges to pace shots and dramatic beats, then assemble groups using the rules below. They do not prescribe one group per beat or require multi-shot groups to reach 20 seconds: | Content | Preferred range | |---|---:| | One micro-performance, one line, or one local action | 4–8s | | Two-person exchange with one tactic change | 6–12s | | Complete dialogue beat or continuous blocking | 10–18s | | Low-complexity same-location long take | 15–30s on Seedance 2.5 | | Controlled multi-shot or occlusion-led scene progression | 20–30s on Seedance 2.5 | | Multi-person combat, complex contact, or precision prop action | Prefer 4–12s segments | Use at most one primary action goal per timing block and no more than three short action sentences. If one second requires movement, dialogue, prop transfer, camera change, and a story reversal, split the beat. Complex actions start from a readable active state when appropriate: mid-throw, mid-impact, door already beginning to break. Put the approach or wind-up in another shot. ## Assemble Video Groups A storyboard shot is one camera shot; an action timing block is a phase within a shot; a video group is one model-generation request containing consecutive storyboard shots. Do not treat these units as interchangeable. Split storyboard rows that conceal several cuts or montage shots into individually numbered shots first. After the storyboard and before writing prompts: 1. Give every shot a stable number and a feasible duration, including verbatim dialogue and reactions. 2. Combine consecutive shots into multi-shot groups, prioritizing the same scene, continuous action, compatible references, and one readable dramatic progression. Do not generate a prompt independently for every storyboard row. 3. Add a shot only if the combined duration stays within the selected model/platform limit and its reference and shot-control capabilities. Split at a motivated boundary for a scene/time change, incompatible references, excessive complexity, or the duration ceiling. Do not fill unused seconds or merge unrelated scenes merely to reach the maximum. 4. A single-shot group is an exception: an intentional long take, a complex action needing isolation, an unavoidable remainder or scene boundary, or a platform unable to control internal cuts. State the concrete reason in the group manifest. Preserve planned cuts; do not silently rewrite several shots into one take. If one shot exceeds the limit, revise its timing or split it at a viable action/dialogue boundary before grouping; never truncate or speed up dialogue to fit. 5. Output a manifest with group ID, included shot IDs, shot-to-local-time mapping, total duration, and boundary/single-shot reason. Each shot appears exactly once in order. Each group's clock starts at zero, and the sum of its shot durations equals its total duration. ```markdown ## 视频组划分 | 视频组 | 包含分镜 | 组内时间映射 | 总时长 | 分组依据／单镜原因 | |---|---|---|---|---| | G01 | S01、S02、S03 | S01 0–4s；S02 4–7s；S03 7–12s | 12s | 同场景连续动作，下一镜转场 | ``` Read [references/video-group-example.md](references/video-group-example.md) when producing video groups for a worked multi-shot example. Its assumed duration limit is illustrative, not a model capability claim. ## Unified Video-Group Prompt Use this structure for generic, Seedance, and other selected models. Model profiles change limits, reference syntax, and supported controls, not the section structure. Produce one complete prompt per group, not one complete prompt per included shot. ````markdown ### 视频组G01｜{model} **包含分镜：** {shot IDs} **组内时间映射：** {shot ID → local interval} **平台参数：** {mode}，{duration}s，{actual aspect/resolution UI settings} **分组依据／单镜原因：** ... #### 当前激活参考素材 | 引用 | 标签状态 | 模态 | 对象 | 继承内容 | 不继承内容 | 作用时段 | |---|---|---|---|---|---|---| #### 生成提示词 ```text SCENE CONTEXT {地点、时间、剧情前因、全组角色名单与确切人数、当前状态} ACTIVE REFERENCES {每个资产单独一行：准确标签；继承内容；不继承内容；作用时段。无参考时明确写无外部参考。} LOCATION MAP {继承场景GEO母版的固定锚点、距离、轴线与摄影机安全侧} FIRST FRAME AND SPATIAL BLOCKING {首镜开场构图、人物位置、朝向、视线、道具持有关系；后续切镜首帧按分镜编号列出并承接状态} FORMAT MODE {受控多镜头序列：分镜数、总时长、全部切点及切法；仅单镜组写连续长镜头、无切镜} OPTICS {按分镜编号写视角或焦距、景深、焦点与焦点变化} CAMERA {按分镜编号写机位、高度、距离、运动路径与幅度；遵守安全侧} ACTION TIMING 0.0–4.0s｜S01：{动作与可见反应} 4.0–7.0s｜S02：{明确切镜；承接动作、站位与道具状态} 7.0–12.0s｜S03：{明确切镜；动作完成与组末衔接状态} PHYSICS {本组适用的重量、接触、惯性、重力与物体状态约束} LIGHTING {主光来源、方向、明暗关系，跨镜光线保持一致} AUDIO {组内时间码、环境声、拟音；说话角色固定声音档案逐字粘贴，台词逐字保留；沉默角色不加词} CHARACTER ACTING {逐角色的目标、阻碍、策略、可见行为与反应；有变化时标注分镜或时段} STYLE {逐字粘贴已确认的全局Style Prefix，不留占位符} QUALITY {与已确认风格相符的清晰度、身份稳定、时间稳定、材质和动作质量要求} POSITIVE CONSTRAINTS {确切人数与出镜时段、道具数量与状态、左右手、轴线、允许的切点、组末状态等本组硬约束} ``` ```` Keep all sixteen sections in this order. For an inapplicable control, state that it is not applicable briefly rather than inventing details or dropping the heading. Resolve all placeholders in the delivered prompt. A text-only group has no invented reference tags; retain ACTIVE REFERENCES and state that there are none. Keep the scene GEO master unchanged and render its relevant geography under LOCATION MAP. Distinguish the group's cast count from the visible count in each shot; off-screen characters remain off-screen, and later entrances retain their timing. Use one asset assignment per line in ACTIVE REFERENCES. Bind optics, camera, cut-first-frame blocking, timing, and audio to the same shot IDs and group-local clock. ACTION TIMING may subdivide a shot into smaller beats without inventing cuts; dialogue text appears only in AUDIO. Shared controls apply to the whole group unless a shot-specific change is explicit. If the approved style has no fixed Style Prefix yet, formulate one from that style in the global visual standard and reuse it verbatim. Do not copy example-specific requirements such as photorealism, 8K, NO CGI, 16:9, 12s, or no cuts into unrelated projects. QUALITY describes the desired image quality, not a promise of unsupported output resolution; actual output settings stay in platform parameters. ## Acting and Voice For every active character, derive scene acting from: - A current objective expressed as a verb aimed at a specific partner. - The obstacle and cost of failure. - A visible tactic and tactic change. - A physical task or hand business. - Reaction beginning before the partner's line ends when natural. - A thought-filled assessment pause before a difficult answer. - Gaze target, blink quality, breathing, posture, distance, and status shift. - Emotional and physical state carried from the prior beat. Use observable behavior instead of `angry`, `sad`, or `shocked` by themselves. The tighter the close-up, the smaller the behavior: eyes, breath, jaw, fingers, and thought before speech. Each recurring speaker has one fixed voice prompt in the asset pack. Paste it verbatim whenever that character speaks and omit it when silent. Scene acting may vary; vocal identity must not. Dialogue appears only in `AUDIO`. Only quoted scripted words are spoken. Silent characters do not ad-lib. Keep lips naturally still when not speaking. Duck ambience under clean dialogue when appropriate. ## Physics and First Frame - The first visible frame contains every subject required at the opening in the correct place. A character with a scripted later entrance may enter at the assigned time, but must not be accidentally delayed or omitted. - Define exact character and important prop counts. - Define prop hand, contact, and state. - Describe mass, inertia, gravity, friction, weight transfer, collision, cloth or hair delay, and liquid behavior only where relevant. - Do not allow floating motion, teleporting, frictionless feet, rubbery collisions, or action reset after a cut. - In multi-shot sequences, every cut preserves identity, state, geography, screen direction, gaze, light, props, hand states, and action progress. ## Reference Rules - List only references active in the current group. - Map each image, video, audio, style, or clay-render reference to one or more explicit responsibilities. - In the paste-ready `ACTIVE REFERENCES` block, put each asset on its own line in the form `@实际名称：继承内容；不继承内容；作用时段`. - State non-inheritance boundaries where a reference could overwrite identity, composition, location, light, or timing. - Do not depend on text labels inside an image to bind a subject. - More references are not automatically better. Remove every material without a defined role. - Use actual user/platform tags exactly. Planned labels must remain visibly marked until uploaded and confirmed. ## Generation Modes - `generate`: text-led generation without required reference media. - `reference`: generate from a purpose-bound set of image, video, audio, style, or clay-render references. - `extend`: describe only the new interval, while carrying forward the exact end-state, geography, active identities, prop states, light, and audio space of the source clip. - `edit`: identify the target source and time range, define one focused change, and preserve all unaffected time ranges and properties. Do not promise a mode or limit that the user's active platform does not expose. ## Final Quality Check - Asset list, draft or supplied asset pack, scene GEO maps, storyboard, and requested video groups are present. - Every production asset has a generation prompt or is explicitly supplied and locked. - Platform parameters are outside prompt prose. - Every dialogue line appears verbatim in the storyboard and relevant video group. - The group manifest covers every storyboard shot exactly once, in order, with no hidden multi-shot rows; default groups combine consecutive shots and each single-shot exception has a reason. - Every group obeys the verified selected model/platform limits (or is explicitly provisional); shot durations sum to the group duration, and local timecodes run continuously from zero to that duration without overlaps or gaps. - Each group has one complete sixteen-section prompt in the required order; shot IDs, cuts, optics, camera, blocking, action, and audio agree across sections. - Every internal cut preserves continuity; the group end state can connect to the next group, and a multi-shot group never also says no cuts. - STYLE reproduces the confirmed global prefix verbatim; QUALITY matches that style and does not claim unsupported output settings. - No timing block is overloaded. - Every reference is declared, used, purpose-bound, and free of stale tags. - Every character uses the correct visual state, fixed voice, position, facing, and gaze. - Every first frame, count, prop hand, landmark distance, physics-critical action, and light direction is explicit where needed. - Acting contains behavior, tactics, listening, and state inertia rather than emotion labels alone. - No unrequested subtitles, narration, music, extra dialogue, characters, or props are introduced. - All prompt content is Chinese unless the user explicitly requested another language; permitted structural headings may remain English. - Every proposed asset label uses the exact script name plus only a necessary state or collision qualifier. ## Options When the user asks for choices, show concise options: ```text 故事类型：都市职场、甜宠恋爱、悬疑惊悚、热血动作、仙侠玄幻、家庭温情、喜剧幽默、科幻末世、历史史诗、心理剧情、恐怖超自然、成长剧情，也可自定义。 美术风格：真人现代都市、真人古装中国、真人武侠、2D 国风、2D 日漫、2D 扁平设计、2D 成熟都市恋爱、3D 动画、3D 国风、3D 国风赛博、3D 黏土定格，也可自定义。 目标模型：通用、Seedance 2.5，或用户指定的平台。 时长模式：自动、短镜头优先、Seedance 2.5 长镜头优先。 ```