animated-voiceover

Animated Voiceover

Overall Workflow

Write, split, and save the confirmed full script. First confirm the topic, audience, and final video duration with the user, then complete the full text, verify facts, and split it into segments by semantics. Before execution, read in full the Narration Script Writing Guide; submit the full script and the split results to the user for review. After the user confirms, immediately save it according to the narration guide as an independent 旁白.mdin the current work's unified document directory, then proceed to visual design.
Plan key characters and generate only character reference images. Read in full theCharacter Reference Image Guide, and based on the finalized script and all segments, determine how many key characters in this video actually need stable identities. The number is determined entirely by the content, not mechanically fixed at 1–3, and reference images are not generated for one-off extras. First establish a character list and a "character—appearing segments" mapping table, then generate for each key character one character reference image with a pure white background showing the same character's front and back side by side; generate as many images as there are characters needed and confirm them one by one. Animated videos do not generate independent scene images, art style reference images, or style reference images.
Write and separately save video Prompts segment by segment. After the narration and all key character reference images are confirmed, read in full the video guidance for the selected style, theVideo Prompt Writing Guideand the Seedance 2.0 Official Prompt Guide, and design the asset responsibilities, scene, shots, actions, and transitions one by one according to segment numbers. Connect only the key character reference images actually needed for the current segment, and at the beginning of the Prompt clearly state image by image "which character reference image N is", using stable features to define character identity. Each segment should rewrite the specific location, era, environment, materials, colors, lighting, and rendering style according to its own narrative needs; do not rely on a unified scene image, and do not require different segments to replicate the same scene content. Prioritize rewriting the ideas in the narration into specific characters, objects, actions, and results occurring in recognizable scenes, and do not use abstract concept imagery to carry the main expression. Each time a segment is completed, immediately save it according to the video prompt guide as an independent Markdown file, and lay it flat together with the 旁白.md, character and reference asset plan, and other production documents in the same work document directory; multiple segments must not be merged into the same file, nor kept only in the conversation.
Choose the voice route and generate only segment 1. Before formally generating the video, read the skill'sbuilt-in reference asset library, and explicitly ask the user whether to use an existing voice from it, a private voice additionally registered by the host project, or create a new voice through segment 1. When choosing an existing voice, list the available options and wait for the user to specify, and connect the selected standardized audio starting from segment 1; when choosing to create a new voice, first confirm the voice description, and segment 1 does not connect an audio reference, letting the voice be generated together with segment 1. Both routes first generate only segment 1; after confirming the task returns successfully, hand it to the user for review, and continue only after obtaining user confirmation.
Lock the unified voice anchor. Read in full the Voice Reference and Audio Conversion Guide. Existing voice routes continue to reuse the same standardized audio in the library, and no replacement is extracted again from clip 1; for a newly created voice route, strip the pure audio from the confirmed clip 1, convert, verify, and upload it as an independent audio asset acceptable to the target platform.
Generate subsequent clips in parallel. Starting from clip 2, all use the same pure audio or the same voice ID locked in step 5; the accurate narration of the current clip is still determined only by the current Prompt. Verify the assets, Prompt, and parameters before submitting in parallel.
Confirm the returns and concatenate. After generation, only confirm that each video task has returned successfully, that its final state is success, and that the node ID, task ID, and resource URL all exist, and download all videos directly to the current work's video/ directory. By default, do not play, extract frames, transcribe, or inspect specific content. After all clips succeed, concatenate them in order; only when the user explicitly requests it, or when a task fails, a resource is missing, or the platform returns an exception, perform targeted content checks. When a cover is needed, separately read Video Cover Generation Guide。

Task Routing

Task	must be read in full
Write, rewrite, or split narration	Narration Script Writing Guide
Plan key characters based on the script and generate or use character reference images	Character Reference Image Guideand the currently selected style file
Select a built-in style, write image or video Prompts	the currently selected style file
Select or use the voice distributed with the skill	Built-in Reference Asset Library
Write, rewrite, or troubleshoot Seedance Prompts	Video Prompt Writing Guideand Seedance 2.0 Official Prompt Guide
Select, establish, convert, or change the voice anchor for this issue	Voice Reference and Audio Conversion Guide
Handle looping actions, repeated camera movement, or stiff cuts	In addition to the above video materials, also read Community Director Practice Notes
Generate, compare, modify, or quality-check covers	Video Cover Generation Guide

The official guide is long; when troubleshooting in a targeted way, search first for:基础公式、多模态参考、编辑视频、延长视频、定义主体、使用分镜时序、动作描述要求、运镜写法、特殊字符规范、音色参考不准、人物 ID 漂移、风格漂移、视频延长 vs 分段拼接、视频结尾有噪音and中文发音不准. Community notes only supplement verified directing practices and do not replace the official guide or this skill's confirmed rules.

Visual Style and Reference Assets

Current built-in styles:

When the user has not specified a style, show these options and wait for a selection; do not silently set a default value. Custom text styles and built-in styles have equal priority.

The style file defines the medium, materials, lines, shapes, colors, and motion language that remain stable across works; it is not a fixed Prompt or shot template that can be copied wholesale. For each work, clip, and shot, the subject, scene, era, composition, lighting, action, camera position, camera movement, and transitions must still be decided anew based on the content; do not repeat fixed image and camera-movement combinations just because the style is reused.

Image guidance using style generates only character reference images: each image contains only the front and back views of the same key character on a pure white background. The number of reference images equals the number of key characters whose identity needs to remain stable in the current film, with no preset upper limit or fixed count; one-off extras do not get reference images. Animated videos do not generate scene images or art style reference images. Detailed planning, generation specifications, and image reference notation are maintained byCharacter Reference Image Guide.

Scenes and art style are defined directly by each segment's text Prompt. Different segments may use completely different locations, eras, weather, props, and compositions, but must maintain overall unity through the chosen style's consistent medium, styling, materials, color principles, and rendering language; do not create consistency by replicating fixed scene content.

The official voices distributed with the skill are registered inBuilt-in Reference Asset Library, whose assets/reference-library/voices/ files are the only official media source for the public voice library. Before formally generating videos, read the built-in voice list and explicitly ask the user whether to use existing voices or create new voices, with no silent default value. Host projects may additionally provide private character references and voices, but must not scan historical work audio to pass it off as library assets. When character references are used across canvases, first download via the official CLI and then upload; do not directly connect nodes across canvases.

Tool Routing and Support Boundaries

The only currently officially maintained and actually verified execution path for images, audio, and video in this skill is LibTV CLI. The complete script, segment splitting, asset responsibilities, character consistency, multi-shot directing, aspect ratio, voice, and task return confirmation methods are themselves platform-decoupled, and can also be used for Higgsfield, Jimeng, or other platforms capable of completing multimodal generation and asset management via CLI.

When using other platforms:

First read that platform's latest official CLI documentation; do not guess or directly translate parameters based on LibTV commands.
Confirm that the CLI is installed and logged in, and query in real time the models, schema, input modes, asset count and format limits, segment duration, aspect ratio, resolution, sound toggle, asynchronous tasks, download, and concatenation methods.
Map this skill's asset responsibilities and steps to the target CLI, and retain task IDs, failure statuses, stderr, or equivalent logs.
Platform-specific commands, model names, and parameter mappings are maintained separately and must not be mixed across platforms. When the latest documentation, login status, or necessary capabilities are missing, stop explicitly; do not fabricate commands or silently switch platforms.

Official entry points:

Higgsfield CLIand itslatest official version
Jimeng CLI Usage Guide(Feishu/Lark permissions may be required)

The currently verified default pacing is to split a 1–5 minute final film into multiple 15-second segments, with approximately 60 Chinese characters of narration per segment and generally 5 shots. Although Seedance 2.5 Pro supports up to 30 seconds, narration length, shot density, continuity, audio stability, and cross-segment strategy have not yet been systematically optimized; before formal verification, do not mechanically double the 15-second rule, nor set 30 seconds as the default.

LibTV Execution Rules

Only in LibTV, parse GPT-Image-2 as Lib Image、Nano Banana Pro as General image Pro、Nano Banana 2 as General image V2、Midjourney as 悠船; other platforms must not follow this. Before formal calls, real-time search and schema remain authoritative.
When passing a model to libtv node ... -s "model=..." , you must use verbatim the libtv model search returned by matches[].modelName, including the spaces, capitalization, and punctuation therein; do not switch to modelKey, and you must not use libtv model <nameOrId> Inconsistencies that may appear at the top level of a complete schema modelName. For example, the search result is Seedream 5.0 Pro, you must pass -s "model=Seedream 5.0 Pro", and must not write it as Seedream5.0 Pro. The complete schema is only used to validate capabilities and parameters.
Query the model and schema in real time; prefer Seedance 2.0 Pro; currently LibTV corresponds to Seedance 2.0 VIP (star-video2). The verified default value is duration=15；ratio、resolutionandenableSoundRead from the user's requirements or project configuration. The model display name,modeTypeand material restrictions must not be guessed based on experience.
When using the Seedance 2.0 series, if the user has not explicitly requested 1080p、4K or another higher tier,resolution use by default 720p; even if the model or an old node carries a higher default value, 720pmust be explicitly passed in the CLI. When the user explicitly requests a higher resolution, first confirm via the real-time schema that the tier is indeed supported; if it is not supported, stop and explain, and do not silently downgrade or switch to another model.
Every time a node is created or regenerated, explicitly pass the actual parameters, and do not inherit old canvases, copied nodes, or model default values.duration、ratio、resolution、enableSound、countandmodeTypeWrite only to CLI parameters, not into the creative Prompt.
When audio serves the role of timbre reference, even if only one audio is connected, generally explicitly select the all-purpose reference mode supported by the real-time schema mixed2video, and must not switch to it merely because there is only one audio in the input audio2video。audio2video is not equal to "single-audio reference timbre mode"; if the platform requires valid visual input, connect the character or video reference actually needed for the current segment, and do not connect irrelevant material just to pass validation. Before formal generation, the input restrictions must still be confirmed in real time.
After parsing the canvas UUID, explicitly pass it in formal node creation, generation, upload, and download -p <projectUuid>; after returning, verify the actual canvas UUID, node ID, and task ID, and retain failure status and stderr.
All independent video segments and the final composite video of the current work are downloaded uniformly to <作品目录>/video/, with files placed directly flat under that directory; use segment numbering to ensure order, for example 片段01.mp4、片段02.mp4, do not create segments/segment-xx-* and other per-segment subdirectories.
Do not assemble LibTV HTTP requests yourself, and do not use a personal API Key to bypass the CLI. If the project tool policy conflicts with this skill, stop first and explain, and do not silently choose one of them.

Quality gates and failure conditions

Execute before generationVideo prompt writing guidepre-submission checks in . After generation, by default only confirm the task return and resource integrity. The following situations must not proceed to subsequent batch generation or final splicing:

The timbre of segment 1 has not yet been confirmed; an existing timbre route has not actually connected the selected in-library audio starting from segment 1; or subsequent segments of any route have not actually connected the locked unified pure audio/timbre ID.
The planning of the number of main characters has not been completed based on the script; any main character requiring a fixed appearance lacks a confirmed character reference image; the character reference images do not simultaneously show the front and back on a pure white background as required; or the reference material roles, numbering, and actual connection order are inconsistent.
The video Prompt still uses abstract space, floating symbols, conceptual transformation, or pure color changes to carry the main expression, without translating ideas into concrete events, character actions, object relationships, and observable results.
The CLI lacks login, real-time schema, necessary model capabilities, or traceable task and node information.
If any video task does not return a successful final state, the node ID, task ID, or resource URL is missing, the resource cannot be downloaded, or the platform reports another exception.

By default, do not play the video, extract frames, transcribe narration, or check characters, art style, actions, actual frame size, or ending audio. Only when the user explicitly requests it, or when a task fails, a resource is missing, or the platform returns an exception, perform content checks related to the exception. On failure, retain the task ID, node ID, parameters, status, and stderr; after correcting the real cause, regenerate, and do not silently switch models or use other fallbacks to cover up the problem.

Delivery principles and information priority

The specific structures of narration and video prompts are based solely on the corresponding writing guides as the only source. By default, narration and other creative drafts are first shown in the conversation as ordinary Markdown paragraphs, not placed in text、plaintextor other code blocks; once the narration is confirmed by the user, it must be saved as an independent 旁白.md, and each video prompt fragment must also be saved separately as an independent Markdown file. The above files, together with production documents such as characters and materials, are all placed flat in the same document directory of the current work, and the file paths are returned to the user. Technical commands, JSON, and parameter examples may still use code blocks when necessary.

When information conflicts, judge according to the following sources: the model display name passed to the node is based on libtv model search of matches[].modelName prevails; input capabilities and parameters are based on the current CLI's real-time schema; Seedance prompt capabilities are based on the official guide; visual language is based on the style selected by the user and the confirmed reference images; pacing such as 15 seconds, about 60 Chinese characters, and generally 5 shots is based on this skill's verified production experience.