SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsframevideo-ai-production

framevideo-ai-production

Design & creation

View repository

Updated 10-15 17:33

About this skill

FrameVideo AI Production Workflow

When To Use

Use this skill when the user requests:

  • Script-to-storyboard — Generate Clip-level storyboards from a raw script or story synopsis
  • Video task planning — Aggregate Clips into Shot-level video generation tasks
  • Prompt polishing — Polish a short idea into a final prompt that can be fed directly to the model
  • Chanjing AIGC generation — Use the Chanjing platform to generate images/videos
  • End-to-end AI video production — The complete process from idea to final FrameVideo composition

Do NOT Use

Avoid using this skill when:

  • Only editing FrameVideo HTML — Use framevideo skill
  • Only downloading Chanjing music/sound effects — Use framevideo-media skill
  • Only voiceover/subtitles — Use framevideo-voiceover-ssml + framevideo-media
  • Digital human video — Use chanjing-digital-human skill
  • Manual editing with complete existing materials — Use framevideo skill

Quick Start

Scenario 1: Generate storyboards from a script

用户: 帮我把这个剧本转成 AI 视频分镜：
【剧本内容】
场景一：清晨，阳光洒进房间...

System behavior:

  1. Read references/storyboard-script-generation.md
  2. Generate a Clip-level storyboard script
  3. Output a structured Clip list

Scenario 2: Plan video generation tasks

用户: 我有 10 个 Clip，帮我规划成视频生成任务

【Clip 列表】

System behavior:

  1. Read references/video-task-planning.md
  2. Aggregate Clips into Shots (each Shot ≤15 seconds)
  3. Output a Shot-level task plan

Scenario 3: End-to-end production

用户: 从这个剧本直接生成视频，用蝉镜 AIGC

【剧本内容】

System behavior:

  1. Read workflow-and-handoff-formats.md
  2. Script → Clip → Shot → Chanjing AIGC Plan
  3. Hand off to chanjing-digital-human to execute generation
  4. Backfill into the FrameVideo composition

Core Units

  • Script: Raw script, plot synopsis, dialogue, scene description, or unstructured story.
  • Clip: A fine-grained storyboard unit; usually a Clip has only one main action, visual focus, or emotional beat.
  • Shot: One video model generation task, formed by aggregating one or more consecutive Clips, usually no more than 15 seconds.
  • Final Prompt: The final prompt that can be fed directly to the video model, including subject, scene, action, dialogue, camera language, lighting, environmental dynamics, and reference asset declarations.
  • Chanjing AIGC Plan: A structured task plan for Chanjing AI Creation, including creation_type, model selection rationale, payload field sources, idempotent fingerprint input, asset output directory, and FrameVideo backfill location.

Progressive Disclosure Rules

Read only the necessary reference according to the user request:

  • When storyboard scripts need to be generated from a raw script, short drama, web drama, or film/TV segment, read references/storyboard-script-generation.md。
  • When existing Clips need to be merged, split, grouped, or planned into video generation Shots, read references/video-task-planning.md。
  • When a single-video story, reference image tags, or a short plot need to be polished into a final model prompt, read references/single-video-prompt-polisher.md。
  • When you need to use Chanjing AIGC to actually generate images/videos, convert Shots into AI Creation tasks, or write generated assets back into a FrameVideo composition, read references/chanjing-aigc-handoff.md, and read as needed ../chanjing-digital-human/references/ai-creation.md。
  • When you need to determine the workflow, convert formats, chain together multi-stage artifacts, or handle end-to-end production, read references/workflow-and-handoff-formats.md。
  • When the user already has output from a previous stage and requests partial modifications, identify the level at which the modification occurs, read only the reference for the corresponding level, and do not load other files.

Only when the user requests the "full workflow," "from script to final prompt," or "all stages" should you read multiple references in sequence. Otherwise, keep reads to a minimum.

Entry Point Determination

  1. The user provides Script, and requests storyboards, shot scripts, AI image-generation storyboards, AI video storyboards, or storyboards:
    • Read storyboard-script-generation.md。
    • Output a Clip-level storyboard.
  2. The user provides a numbered storyboard, Clip list, shot script, or storyboard, and requests video generation planning, aggregation, merging, or splitting:
    • Read video-task-planning.md。
    • Output a Shot-level video task plan.
  3. The user provides a single-video concept, short narrative, or reference image asset tags, and requests prompt polishing or direct video prompt generation:
    • Read single-video-prompt-polisher.md。
    • Output the final prompt.
  4. The user requests using Chanjing AIGC for image generation, video generation, B-roll generation, or shot asset generation, or requests continuing from Shot to actual generation:
    • Read chanjing-aigc-handoff.md。
    • If the input is still a Script, first read storyboard-script-generation.md to generate Clips.
    • If the input is still a Clip, read video-task-planning.md to aggregate Shots.
    • Output a Chanjing AIGC Plan by Shot, then hand it off to Chanjing AI Creation for submission/polling/download.
  5. The user requests end-to-end production:
    • First read workflow-and-handoff-formats.md, and confirm the format handoff.
    • Then read storyboard-script-generation.md to generate Clips.
    • Then read video-task-planning.md to aggregate Shots.
    • If only prompts are needed, read single-video-prompt-polisher.md to generate the Final Prompt.
    • If actual generation or FrameVideo project implementation is needed, read chanjing-aigc-handoff.md to generate a Chanjing AIGC Plan, then hand it off to FrameVideo for orchestration, preview, and validation.
  6. The user already has output from a previous stage and requests partial modifications, rewriting a certain Shot or Clip, adjusting pacing, or redistributing dialogue:
    • Identify the level involved in the modification (Clip / Shot / Final Prompt).
    • Load only the reference for the corresponding level; do not rerun the entire pipeline.
    • Preserve and output the parts not involved in the modification as-is, and change only the scope specified by the user.
    • If the modification spans multiple levels (such as rewriting a certain Clip and updating the corresponding Shot), process them in order from top to bottom.

Standard Pipeline

Execute end-to-end requests in the following order:

Script
-> Clip 级分镜
-> Shot 级视频任务
-> Final Prompt 或 Chanjing AIGC Plan
-> 本地生成资产
-> FrameVideo composition

Each stage must preserve:

  • Original character names, character relationships, and event order.
  • Original dialogue; unless the user explicitly requests rewriting, the dialogue content must not be polished.
  • The original format of asset tags such as reference images, character images, and prop images.
  • Key continuity information: character identity, costume, props, staging, light direction, spatial state, and action direction.

Output Selection

Output to which layer depends on the depth the user wants:

  • Storyboard only: stop at Clip.
  • Video task planning needed: stop at Shot.
  • Prompt usable by the model needed: output Final Prompt.
  • Real Chanjing AIGC generation needed: output Chanjing AIGC Plan, and hand off to Chanjing AI Creation.
  • Full plan needed: output in order the workflow positioning, Clip storyboard, Shot planning, Final Prompt or Chanjing AIGC Plan, and self-check.

Handoff Principles

  • When going from Clip to Shot, prioritize whether the core generation state is stable, rather than mechanically splitting by shot size or shot-reverse-shot.
  • When going from Shot to Final Prompt, merge the action chain and emotion chain, but do not force the goals of different segments into the same prompt.
  • When going from Shot to Chanjing AIGC Plan, model parameters must not be hardcoded. First have Chanjing AI Creation obtain the models available to the current account, then based on the selected model's params_config.fields construct the payload.
  • If a single Shot contains complex group scheduling, strong action conflict, a large amount of dialogue, complex hand movements, spatial transitions, or strong special effects, you must evaluate whether to split it.
  • If the user's input lacks key information, prioritize reasonable inference based on the text; only ask when the gap would change the plot, character relationships, or generation format.

FrameVideo Handoff

When the user's goal is to "generate a video" rather than "only storyboard/prompt", this skill does not stop at Final Prompt. Continue executing:

  1. Generate Chanjing AIGC Plan.
  2. Use ../chanjing-digital-human/references/ai-creation.md 's idempotency rules to submit and sync the Chanjing AI Creation task.
  3. Only use the assets/ai-creation/images/ or assets/ai-creation/videos/ assets downloaded locally.
  4. Return to the framevideo skill to write the composition, subtitles, transitions, music, and final validation.

Global Self-Check

Before final delivery, briefly check by layer. For the detailed self-check of the storyboard layer, see the end of storyboard-script-generation.md .

Clip Layer

  1. Whether each scene has an opening anchor point and a closing anchor point.
  2. Whether key dialogue satisfies the basic 2-shot system, and whether high-energy dialogue satisfies the enhanced 3-shot system.
  3. Whether dynamic actions are split into initiation, execution, and result, without being covered in a single shot.

Shot Layer 4. Whether the Clip is too fragmented (should be merged), or the Shot is too long and complex (should be split). 5. Whether visual continuity is stable: characters, props, light direction, axis, and spatial state do not jump within the same Shot without reason.

Final Prompt Layer 6. Whether the original plot, original dialogue, and asset tags are fully preserved, with no unsupported plot added. 7. Whether the final prompt explicitly prohibits subtitles, and whether the output layer meets the user's requirements.