video-master - SKILL.md Agent Skill

name: video-master description: Use when a user wants to turn a video idea, campaign brief, existing assets, story, product concept, or rough requirement into a video pre-production package with intake routing, video-mode confirmation, creative strategy, script/copy/audio extraction, shot list/storyboard, native image-generated storyboard frames, and Chinese or model-specific video generation prompts. Triggers include video-master, 视频脚本, 分镜, storyboard, 短视频, 广告片, TVC, 产品宣传片, AI视频提示词, or video prompt.

Video Master

Video Master turns a user's requirement or source assets into an AI video production package. It behaves like a director's pre-production workflow: identify what the user already has, confirm the video mode, lock production constraints, design rhythm, then create scripts, storyboard frames, audio copy, and copy-ready prompts.

Dependencies

The skill can run with no extra Python packages, but the repository supports recommended dependencies for richer local tooling. When the user allows setup before production work, install dependencies from the project root:

python3 -m pip install -r requirements.txt

Recommended dependencies enable:

PNG storyboard contact sheets with Pillow.
Production workbook export with openpyxl.
Packaged MP4 storyboard animatic previews with imageio/ffmpeg and numpy. The default draft profile uses 12fps and follows the project's aspect ratio, with optional smooth and off profiles.
Optional TTS voiceover generation with edge-tts or a local VoxCPM2 HTTP service.
Stronger JSON/subtitle validation with pydantic and pysubs2.

Global Rules

Match the user's language unless they explicitly request another language.
Default output location: create video_projects/<project_slug>_<YYYYMMDD_HHMM>/ under the current workspace unless the user names a destination.
Follow the serial pipeline. Do not write later-phase deliverables before the current phase's gate is satisfied.
Treat brief/spec_lock.md as the execution contract. Re-read it before writing each shot prompt, generating each storyboard frame, or assembling final deliverables.
Use native image generation for storyboard frames when the user asks for images, 分镜图, storyboard frames, keyframes, or visual boards. Do not substitute SVG boxes or text-only placeholders.
Video Master is built around Codex native image generation. Default to assuming native image generation is available during project work; do not mark it unavailable just because a CLI key, environment variable, or tool listing is missing. Only record image generation as unavailable after an actual native image-generation attempt fails in the current turn.
When using the local WebUI, scripts/serve_webui.py can start a Codex device login and call Codex native image generation for a selected storyboard shot. It can also trigger optional TTS voiceover generation from audio/tts_lines.json through edge-tts or a local VoxCPM2 service. Generated originals are written to 最终交付/01_分镜图/, voiceover audio is written to 最终交付/03_口播与字幕/, and status is recorded in qa/metadata/.
Treat fixed characters as a continuity lock before storyboard generation. When a project includes recurring people, hosts, founders, interviewees, actors, or mascots, create/confirm character design anchors first and reference them from storyboard image prompts and video prompts.
For formal projects that generate storyboard images, enforce style_confirmation_gate: create/confirm the character anchor and the first storyboard frame (S01), set style_gate_status: pending, and do not batch-generate remaining storyboard frames until the user approves the current style.
Treat title packaging as an optional sidecar branch only. It must never change the normal storyboard, script, audio, or video-prompt generation flow.
When the user asks for commercial title cards, lower thirds, number animations, alpha overlays, or packaging text, create separate title-packaging deliverables; do not add packaging notes, title-packaging file paths, or alpha-MOV instructions to copy-ready video prompts.
During the Production Lock, explicitly ask whether the user needs title-packaging images: main title, chapter/section cards, lower thirds/name tags, key data/counter callouts, CTA/end cards, or none. Default to title_packaging_enabled: false only when the user does not need packaging or asks to keep the package lean.
For title packaging, use native image generation for designed packaging images and chroma-key/transparent PNG look development when available. Use scripts/render_title_packaging.py for exact text, verified transparent PNGs, and optional ProRes 4444 alpha MOV overlays.
Default title packaging output is static transparent PNG. Do not generate MOV just for a simple fade, scale, or position shift; create MOV only when the user explicitly asks for animated overlay delivery and provides a meaningful animation need.
If a dedicated imagegen skill/tool is available, follow it for image generation and project-bound save-path handling.
If a real native image-generation attempt fails, create the complete image prompt set, mark each affected frame Needs-Generation, and record the failure reason in qa/metadata/workflow_events.jsonl.
Do not claim a storyboard image file exists until its path has been verified.
Use Seedance 2.0 as the default target video model/profile for video generation prompts. Record target_model: seedance-2.0 and prompt_dialect: seedance-2.0 unless the user explicitly names another video model.
Prefer Chinese final prompts when the target workflow is Chinese or domestic video models are named. Keep optional English camera/style tags only when useful.
Confirm or explicitly assume copy_language, voiceover_language, caption_language, localized_caption_languages, and subtitle_rendering_policy before writing script, TTS, captions, or final video prompts. Prompt language controls the model prompt; copy language controls the spoken/readable words.
Default subtitle_rendering_policy to post-production-only and burned_subtitles_allowed to false for generated video clips. Captions may exist as audio/captions.srt for preview/post-production, but final video prompts must not ask the video model to render subtitles unless the user explicitly approves generated on-screen text.
For domestic Chinese workflows, include a Chinese SRT deliverable even when the voiceover is English. If the VO is English, package both 英文字幕.srt and 中文字幕.srt in 最终交付/03_口播与字幕/.
If narration is external voiceover rather than on-camera speech, do not include the exact VO lines in copy-ready video prompts. Keep spoken copy in audio/voiceover_script.md, audio/tts_lines.json, and SRT files; video prompts should only say the VO is added in post.
Copy-ready video prompts must explicitly say the video model should not generate background music for each clip. Music direction belongs to post-production or whole-film audio planning, because per-clip generated music creates hard seams during assembly.
Every copy-ready video prompt must first request natural synchronous production sound / room tone / environment sound when model audio is relevant, then add per-shot SFX/sound-design cues as accents. SFX cues guide the clip; they must not over-constrain the model into only the manually named sounds. Background music is not generated per clip.
Do not use a Negative prompt / 负面提示词 field in final video prompts. Prefer positive generation requirements and clear policy fields such as 画面文字策略.
Avoid unlicensed celebrity likenesses, copyrighted characters, trademark-heavy style imitation, or deceptive real-person depictions unless the user provides rights and the request is allowed. Convert risky requests into original characters, original brands, or generic style language.
If the user provides reference images or videos for style, treat them as reference_style assets: extract transferable color, lighting, camera, pacing, framing, and packaging rules, but do not copy subjects, plot, branding, protected characters, creator identity, or a living artist/director's protected style. Use the distilled rules and safe reference frames to guide native image generation and final video prompts.
When a user wants a reusable style approach, offer three project style modes: original, use_style_template, or create_style_template_from_reference.
Treat visual_style_preset_id as a lightweight look card for storyboard frames and video prompts. It is separate from template_id: presets lock image look, color, light, texture, and camera feel; templates lock a full director method including rhythm, editing, sound, and prompt structure.
Official style templates live in style_templates/<template_id>/ and are applied as a complete director archive through template_id; do not ask for or write a template strength.
When a style template is selected, user ideas override template defaults: explicit user ideas, supplied assets, brand/copy constraints, and project-specific creative directions come first. Capture these as template_user_overrides and make the output follow the user first.
Do not maintain light / medium / high variants for official templates. If the user wants a variation, treat it as a project-specific override rather than a new template strength.
Do not use a draft style template for a final project unless the user explicitly opts in with allow_draft_template: true.
A style template transfers creative rules such as rhythm, palette, camera language, sound policy, and prompt structure; it never authorizes copying reference subjects, exact shots, dialogue, branding, subtitles, watermarks, or creator identity.

Pipeline

Step 0: Workflow Entry Mode

Gate: the user starts a new video-master project or opens an existing project for major changes.

Offer two entry modes:

autopilot: full delegation. The user gives the brief and assets, then Codex makes reasonable assumptions, logs them, and only asks when blocked by missing rights, unsafe claims, missing core assets, or irreversible creative choices.
guided: collaborative director mode. Codex confirms key information, offers visual/style/rhythm options, brainstorms with the user, then implements after confirmation.

Record the decision in brief/spec_lock.md:

workflow_mode: autopilot | guided
confirmation_policy: ask_only_blockers | confirm_each_phase
assumption_policy: auto_fill_with_log | require_user_confirmation

When the user says to proceed quickly, default to autopilot. When the user wants brainstorming, comparison, or creative control, default to guided.

Step 1: Input Readiness Check

Gate: the user has provided a video idea, source material, or rough requirement.

Classify the input mode before creative work:

idea-only: the user has only a concept; Codex may propose product, story, scene, copy, and visual assumptions.
asset-assisted: the user has partial assets such as product photos, logo, copy, selling points, reference videos, or target platform.
material-locked: the user has approved assets/copy/brand rules; Codex must structure and adapt them without inventing unsupported claims or changing key wording.

Before video-mode confirmation, classify the style route:

original: create a new style from the project brief.
use_style_template: use an official template from style_templates/.
create_style_template_from_reference: analyze reference assets and produce a draft template package for user confirmation.

If using a template, capture template_id, whether draft templates are allowed, and template_user_overrides from any user-supplied ideas or constraints.

If using a scene director pattern, capture scene_director_pattern and any user overrides that change the pattern's default staging or rhythm.

Before storyboard work, classify the visual style preset route:

preset: use one visual style card from references/visual_style_presets.json.
custom: user supplies a custom look in natural language.
reference-derived: derive the look from user-provided reference assets and record safe transfer rules.

If the mode is unclear, ask one concise question. Otherwise proceed with explicit assumptions.

Write:

strategy/input_readiness.md

Capture available materials, missing materials, what may be invented, what must be preserved, and any required source files in sources/.

When the user uploads reference images or videos, label each asset's role explicitly:

reference_style: use only for transferable style rules such as palette, contrast, lighting, pacing, camera language, framing, transition rhythm, and visual packaging. Captions/subtitles may be analyzed as packaging observations, but final image/video prompts must not ask models to reproduce burned subtitles or subtitle styling unless subtitle_rendering_policy and burned_subtitles_allowed explicitly allow generated text.
reference_subject: use for permitted product/person/object continuity when the user owns or provides the asset.
do-not-copy: note any protected characters, celebrity likenesses, brand marks, plots, slogans, or recognizable creator-specific style that must not be replicated.

Step 2: Video Mode Confirmation

Gate: Step 1 complete.

Confirm the video mode before writing the script. Use references/video-modes.md.

Common modes:

fast-paced-tvc
product-promo-short
narrative-short
animation
tutorial-explainer
brand-film
ecommerce-conversion-short

When the brief clearly matches a repeatable scene grammar such as product showcase, live-commerce spokesperson, short drama, science visualization, fantasy action, music beat montage, one-take transition, video extension, video edit, motion poster, or animation action, also read references/scene-director-patterns.md after selecting the main video mode. Treat the scene pattern as a director-method layer that can combine with a full style template and a visual style preset.

Present the recommended mode and the impact on structure, rhythm, audio, storyboard coverage, and deliverables. Wait for confirmation unless the user already specified the mode or explicitly said to proceed.

Write:

strategy/video_mode.md

Step 3: Production Lock

Gate: input mode and video mode are known.

Create the project folders:

video_projects/<project_slug>_<YYYYMMDD_HHMM>/
  sources/
  brief/
  strategy/
  script/
  storyboard/frames/
  prompts/
  audio/
  characters/
    character_bible.md
    character_manifest.json
    reference_images/
  packaging/
    title_packaging_plan.json
    title_packaging_prompts.md
    title_cards/
    alpha_mov/
  references/
  references/reference_keyframes/
  最终交付/
    01_分镜图/
    02_提示词/
    03_口播与字幕/
    04_分镜总览/
    05_预览视频/
    06_制作总表/
    07_title_packaging/
  qa/metadata/

Present the Production Lock as a bundled recommendation and wait for confirmation unless the user has supplied the decisions or explicitly allowed assumptions:

Workflow mode: autopilot or guided, with confirmation and assumption policy
Input mode and asset authority
Video mode
Objective and CTA
Audience and platform
Aspect ratio and target duration
Prompt language and target video model/profile; default target video model: seedance-2.0, unless the user explicitly names another video model.
Copy/VO language, caption language, and subtitle rendering policy (post-production-only by default)
Narrative style, visual style, and pacing style
Visual style preset: choose one preset from references/visual_style_presets.json, custom, or reference-derived. Present 2-4 relevant cards with a recommended default instead of an unstructured open-ended style question.
Scene setting and scene-anchor plan: project-level anchor, per-segment anchor, or none; record stable set/location/tabletop/stage rules and detailed visual style description.
Character/product/brand continuity rules, including whether fixed people need a character-design lock before storyboard generation
Claims/compliance boundaries
Storyboard image coverage: every shot, key shots, or selected scenes
Reference style usage: mimic color grading, camera language, edit rhythm, typography/packaging, or only general mood
Style route: original, use_style_template, or create_style_template_from_reference
Style template fields when applicable: template_id, allow_draft_template, and template_user_overrides
Template application summary: what is inherited from the template, what is overridden by the user's ideas, and what must not be copied
Scene director pattern when applicable: product_showcase, live_commerce_spokesperson, short_drama_reversal, science_visualization, fantasy_action, music_beat_montage, one_take_transition, video_extension, video_edit, motion_poster, animation_action, or none.
Optional title packaging: ask whether to generate main_title, chapter_card, lower_third/name_tag, data_callout/counter, cta_card/end_card, or none. Capture exact copy, style references, PNG-only vs real animated overlay need, and title_packaging_enabled. This is a sidecar branch and does not modify video prompts.

Write:

brief/creative_brief.md
brief/spec_lock.md

Use references/output-contract.md for the required sections.

Step 3.5: Reference Style Analysis

Gate: reference images or videos exist and the user wants them used for style.

Write:

references/style_analysis.md
references/color_style.md
references/editing_style.md
references/reference_style_manifest.md
references/reference_keyframes/ when frames are extracted or selected

Measure objective traits with local tooling when possible: duration, aspect ratio, FPS, scene count, average shot length, keyframes, color palette, brightness, contrast, saturation, and motion density. Then synthesize them into creative rules that can be reused without copying the source content.

The final style rules must answer:

What should be mimicked: palette, contrast, lighting, lens language, shot duration pattern, camera movement, transitions, and packaging. Subtitle behavior may be recorded as an analysis-only/post-production observation, not as a generation instruction unless the subtitle policy explicitly allows generated text.
What must not be copied: people, brand marks, exact scenes, plot, dialogue, slogans, copyrighted characters, or protected creator style.
How the rules affect storyboard image prompts and video prompts.

Step 3.6: Visual Style Preset Lock

Gate: Production Lock is confirmed, and reference-style analysis is complete when reference assets are used.

Use references/visual-style-presets.md and references/visual_style_presets.json. This step must be complete before prompts/storyboard_image_prompts.md is written.

Ask or confirm one of:

One preset card, such as imax_70mm_realism, photoreal_commercial, eastern_fantasy_3d, hyperreal_3d_render, graphic_2d_editorial, soft_storybook_2d, anime_cinematic_light, noir_gothic, or future_tech_clean.
custom, when the user describes a look that does not fit a preset.
reference-derived, when reference assets define the look.

When recommending, present 2-4 relevant cards, name the recommended one, and explain the tradeoff in one sentence. Record the selected preset in brief/spec_lock.md under visual_style:

visual_style_lock
visual_style_preset_id
visual_style_preset_name
medium
realism_level
art_direction
color_palette
lighting
texture
camera_language
storyboard_prompt_rules
video_prompt_rules
visual_style_overrides

If a full template_id is also selected, the template's prompt rules remain the larger director method, while the visual style preset supplies the exact look card unless the user overrides it. Do not ask for exact living-artist or studio imitation; convert those requests into descriptive traits.

Step 3.7: Character Design Lock

Gate: Production Lock and Visual Style Preset Lock are complete, and the project has recurring characters or the user requests fixed人物/主持人/创始人/采访对象/演员/虚拟角色 continuity.

Ask or confirm whether the project needs fixed-character continuity. If not, record character_lock_enabled: false and continue. If yes, lock the visual identity before writing storyboard image prompts:

Define stable character IDs such as host_a, founder_b, or guest_c.
Record each character's role, age range, face/hairstyle/body-shape descriptors, wardrobe rules, temperament, forbidden changes, and allowed variations.
Generate or collect reference images when available: face/front, half-body or full-body, and any required wardrobe or expression references.
Store the character bible and manifest in characters/.
Record prompt rules that require downstream storyboard and video prompts to reference the locked character IDs instead of reinventing the person.
Mark the main reference as the character anchor used by the later style_confirmation_gate.

Write when enabled:

characters/character_bible.md
characters/character_manifest.json
characters/reference_images/

Record the selected policy in brief/spec_lock.md under character_design:

character_lock_enabled
character_lock_status
fixed_characters
character_reference_dir
character_prompt_rules
character_anchor

Step 4: Creative Strategy And Rhythm Map

Gate: creative_brief.md and spec_lock.md exist, and character design is locked or explicitly skipped.

Write:

strategy/creative_strategy.md
strategy/rhythm_map.md

The rhythm map must allocate non-uniform shot durations unless the confirmed mode genuinely calls for uniform timing. For advertising modes, design hook, product memory, proof, and CTA beats deliberately. Do not default to equal shot lengths.

Act like a creative director, not a spreadsheet. For high-motion subjects such as racing, sport, chase sequences, launch films, and fast-paced TVC, design at least one rapid-cut cluster of short shots before or after longer emotional shots. For model-facing timing, use whole-second durations such as 1s, 2s, and 3s; keep sub-second precision for post-production edit notes only when needed. Mark camera energy explicitly: stable, handheld, vehicle-mounted vibration, impact shake, whip pan, hard push-in, POV, pass-by, or locked-off.

Use references/video-modes.md, references/platform-and-model-profiles.md, and any references/style_analysis.md.

If template_id is present, read style_templates/<template_id>/template.md, director_notes.md, rhythm_rules.json, editing_craft.md, shot_motifs.json, and prompt_rules.md before writing the rhythm map.

When a style template is selected, strategy/rhythm_map.md must name the template and explain how the complete template method is adapted to the current subject, duration, and template_user_overrides.

Step 5: Script, Copy, And Audio Extraction

Gate: creative strategy and rhythm map exist.

Write:

script/script.md
audio/voiceover_script.md
audio/tts_lines.json
audio/captions.srt
audio/captions_en.srt and/or audio/captions_zh.srt when multilingual subtitle deliverables are needed
audio/music_sfx_cue_sheet.md
audio/audio_generation_prompt.md

Use references/audio-and-copy.md. Keep audio copy centralized so TTS, captions, and final video prompts stay consistent. If template_id is present, read style_templates/<template_id>/template.md, director_notes.md, and prompt_rules.md before writing script, copy, and audio files so rhythm, sound policy, and copy posture remain aligned with the selected style template while preserving user-provided copy direction. Confirm whether the spoken copy is Chinese, English, bilingual, or user-supplied before writing audio/voiceover_script.md. Treat audio/captions.srt as a post-production subtitle asset unless burned_subtitles_allowed is explicitly true. For Chinese-facing delivery, create Chinese localized captions even if the spoken VO is English, and keep the VO-language transcript as a separate SRT. In audio/music_sfx_cue_sheet.md, map natural synchronous sound / room tone and at least one SFX cue to every shot. Keep background music as a whole-film post-production direction unless the user explicitly asks to generate or mix music later.

Step 6: Shot List And Storyboard Plan

Gate: script and audio files exist.

Write:

storyboard/shot_list.md
storyboard/shot_list.json

Use a compact overview table plus per-shot detail blocks. Avoid a single very wide Markdown table for all fields. Every shot must include timing, beat, visual action, framing, camera, movement, lighting, audio/copy references, continuity notes, image prompt seed, and video prompt seed.

If template_id is present, read style_templates/<template_id>/template.md, director_notes.md, rhythm_rules.json, shot_motifs.json, editing_craft.md, example_shot_list.md, and prompt_rules.md before writing the shot list and storyboard plan. Apply the template as a complete director method while redesigning the subject, plot, characters, product, and brand details around the user's own ideas. If scene_director_pattern is not none, read references/scene-director-patterns.md and use the selected pattern's required locks and rhythm grammar to shape shot staging. The selected visual style preset still controls look, and any full style template still controls the broader director archive.

Step 7: Storyboard Image Prompts And Native Images

Gate: shot_list.md exists and storyboard coverage is known.

Read brief/spec_lock.md, references/storyboard-and-video-prompts.md, references/visual-style-presets.md, characters/character_bible.md when present, and any references/style_analysis.md. If template_id is present, also read style_templates/<template_id>/prompt_rules.md and carry the template's safe prompt rules as defaults. User ideas and approved assets override template defaults when they conflict.

Write prompts/storyboard_image_prompts.md before generating images. Generate storyboard frames with native image generation:

Every storyboard image prompt must carry the locked visual style preset fields from brief/spec_lock.md: visual_style_preset_id, medium, realism level, art direction, color palette, lighting, texture, camera language, and storyboard prompt rules.
For 15-second Seedance workflows, recurring locations, or product environments where space matters, create a wide scene-anchor image before action keyframes. Store scene anchors in references/scene_anchors/, name them distinctly such as SEG01_SCENE.png, and use them to lock environment layout, light direction, set dressing, action-safe space, and visible placement of recurring characters/products when they belong to the scene.
If character_lock_enabled is true, every storyboard image prompt involving a fixed character must reference the stable character ID and the locked character bible. Do not vary face, age, hairstyle, body type, or signature wardrobe unless character_bible.md allows it.
For formal projects, enforce style_confirmation_gate: generate or confirm the character anchor, generate only the first storyboard frame (S01), set style_gate_status: pending, show both to the user, and do not batch-generate remaining storyboard frames until the user approves the style. If the work is explicitly a simulation/test run, record style_gate_status: skipped and the reason before continuing.
If shot count is manageable and the user requested detailed storyboard images, generate one frame per shot.
If shot count is high, generate key frames unless the user explicitly asks for every shot.
If reference_style assets exist, inject the distilled style rules and safe reference keyframe paths into every storyboard image prompt. Use native image generation with reference images when the available tool supports it; otherwise include the keyframe paths and style rules in the prompt text. Never ask the model to reproduce the exact source video or image.
For recurring characters or products, create or identify visual anchors in references/ when possible. If reference-image conditioning is unavailable, tighten every prompt with identical character/product descriptors and visually check for drift.
After generating key frames, inspect continuity before marking the manifest complete. Regenerate any frame whose character, product, palette, or composition clearly conflicts with spec_lock.md.
Copy project-bound images into storyboard/frames/ and final selected frames into 最终交付/01_分镜图/.

Write:

storyboard/storyboard_manifest.md

Step 8: Video Generation Prompts

Gate: shot_list.md, storyboard prompts, image statuses, and audio files are ready.

Write:

prompts/video_prompts.md
最终交付/02_提示词/视频生成提示词.md
最终交付/02_提示词/图片生成提示词.md

If no other target video model is specified, write the final prompts for Seedance 2.0. If the target workflow is Chinese, Seedance 2.0, or another domestic Chinese video model, final copy-ready prompts must be Chinese-first. Keep prompts/video_prompts.md detailed for review, and make 最终交付/02_提示词/视频生成提示词.md easy to copy into a video model.

Use references/storyboard-and-video-prompts.md, references/platform-and-model-profiles.md, references/scene-director-patterns.md when a scene director pattern fits the project, references/seedance2-practical-playbook.md when Seedance 2.0 or 15-second web generation is targeted, and any references/style_analysis.md. If reference_style assets exist, final video prompts must carry the same safe style rules used for native image generation so generated video matches the reference look and editing language without copying protected content. When a style template is selected, storyboard image prompts and video prompts must carry the template's safe prompt rules as a whole, adapted around template_user_overrides and the current subject. For Seedance 2.0, each copy-ready shot prompt must put model, duration, aspect ratio, and reference-frame path in the shot heading, for example ## S01 - 镜头名（3s / 16:9 / Seedance 2.0 / 参考图：references/scene_anchors/S01_SCENE.png + 最终交付/01_分镜图/S01.png）. Do not add standalone 目标模型, 时长, 画幅, or 参考图 lines in the prompt body. Every Seedance 2.0 copy-ready shot prompt must include 场景设定 and 画面风格说明 before 动态时间切片. 场景设定 locks the stable environment, layout, light direction, key props, and non-drift rules; when recurring or high-frequency characters appear, it must also summarize the visible character identity anchors such as role, age range, face/hair/body cue, wardrobe, prop, and starting position, while still referencing the character bible when present. For style-heavy or realistic scenes, split 画面风格说明 into 风格核心, 视觉基调, 色彩与影调, 摄影机与镜头, 材质与特效, 动作质感, and 风格边界. For realistic live-action scenes where the user wants to avoid AI feel, include 超写实, 极致逼真, and Photorealism-真人实景拍摄; do not use those realism anchors for intentionally stylized animation, anime, 2D, toy-like 3D, or graphic motion. Every Seedance 2.0 copy-ready shot prompt must include a 动态时间切片 section. Split the shot duration into rhythm-driven, non-mechanical whole-second ranges inferred from the shot's actual framing, action, environment, props, and emotional beat. Do not use decimal/sub-second time codes in final model-facing prompts, because they imply a precision current video models may not reliably follow. Do not default to equal slices such as five 3-second blocks for a 15-second segment unless the confirmed rhythm deliberately calls for a fixed beat grid. Use 1-second slices for impact, handoffs, reveal hits, and fast transitions, then longer whole-second slices for performance holds, product proof, atmosphere, or emotional payoff. Do not reuse generic slice text across shots. Each segment should describe concrete camera movement, subject action, environment/material motion, synchronous sound/SFX, and the transition of the shot beat. Keep copy-ready video prompts compact and model-facing. Do not paste actual VO sentences, subtitle file paths, packaging file paths, or explanatory post-production notes into the prompt body. Do not add standalone 运镜与焦段 or 光线与风格 lines after 动态时间切片; camera, lens, lighting, and color belong in 画面风格说明. Put stability, audio, subtitle/text, and execution constraints under one 生成要求 section, with compact lines such as 音频：背景音乐不要生成；同期声...；SFX... and 字幕与文字：不要生成字幕、caption、对白文字或烧录文字。. Do not include a 负面提示词 section in final prompts.

Step 8.5: Optional Title Packaging Sidecar

Gate: the user explicitly asks for title packaging, commercial title cards, lower thirds, number animations, alpha overlays, or packaging text.

This branch runs beside the video prompts. It must not rewrite prompts/video_prompts.md, must not edit 最终交付/02_提示词/视频生成提示词.md, and must not insert title-packaging instructions into any copy-ready video prompt.

In the formal Production Lock, ask the user whether they need any of these packaging images:

main_title: main title or campaign title.
chapter_card: chapter, section, location, or day title.
lower_third / name_tag: person name, role, location label, interview ID.
data_callout / counter: key number, ranking, percentage, year, distance, price, milestone.
cta_card / end_card: ending slogan, follow/subscribe, campaign CTA.

If the user is unsure, recommend PNG-only packaging first. MOV is optional and should be used only for meaningful animation such as stroke reveal, route drawing, counter ticking, mask wipe, glow sweep, or designed motion; do not render MOV for a simple static image with fade/position offset.

Write:

packaging/title_packaging_plan.json
packaging/title_packaging_prompts.md
packaging/title_cards/
packaging/alpha_mov/
qa/metadata/title_packaging_manifest.json
最终交付/07_title_packaging/

Use reference packaging assets as reference_style only: extract typography mood, composition, material, lighting, spacing, commercial polish, motion intent, and layout grammar. Do not copy the original title names, exact layout, logos, brand marks, watermarks, or recognizable source artwork.

When native image generation is available, use it to create designed packaging looks. Without API-native transparency, generate the packaging on a pure chroma-key background, remove the key locally, and verify the transparent PNG. For exact Chinese/English text, counters, lower thirds, and editor-ready static overlays, render deterministic PNG assets with:

python3 ${SKILL_DIR}/scripts/render_title_packaging.py <project_path>

The default output is transparent PNG only. Use --alpha-mov only when the user explicitly requests animated alpha overlay delivery and the motion is more than a trivial fade or offset. In packaging/title_packaging_plan.json, set motion_template to a production template such as brush_reveal, mask_wipe, glow_sweep, route_light_trail, odometer, or marker_annotation. For designed main titles, put the native-image-generation result through chroma-key cleanup first, then pass the transparent PNG as design_asset; the script handles exact alpha MOV motion and keeps the asset outside video prompts. The MOV output is a ProRes 4444 alpha overlay meant for editing software, not a video-generation prompt input.

Step 9: Deliverables Package

Gate: script, shot list, audio files, storyboard manifest, generated frames, and video prompts exist.

Write:

最终交付/00_使用说明.md
最终交付/03_口播与字幕/口播稿.md
最终交付/03_口播与字幕/中文字幕.srt and/or 最终交付/03_口播与字幕/英文字幕.srt
Any model/platform-specific final files requested by the user
Optional title packaging files in 最终交付/07_title_packaging/ only when title_packaging_enabled is true or the user requested the sidecar branch

If the user wants a voiced preview and dependencies/network are available, generate a TTS track from the centralized copy:

python3 ${SKILL_DIR}/scripts/generate_voiceover_tts.py <project_path>

If the user already has narration, place it in 最终交付/03_口播与字幕/ or pass it to the preview tool with --voiceover-audio. If the user has approved a local background music track, place it at 最终交付/03_口播与字幕/背景音乐.mp3, audio/background_music.mp3, audio/bgm.mp3, or pass it with --background-music <path>. To use an Eagle asset directly, pass --eagle-background-music-id <item_id> and optionally --eagle-library-path <path>. This mixes BGM only into the local animatic preview and does not change video-generation prompts.

Generate the storyboard overview:

python3 ${SKILL_DIR}/scripts/make_storyboard_overview.py <project_path>

Generate the production workbook:

python3 ${SKILL_DIR}/scripts/export_production_workbook.py <project_path>

Generate the storyboard animatic preview:

python3 ${SKILL_DIR}/scripts/make_animatic.py <project_path>

The default preview profile is draft: 12fps and an output size inferred from brief/spec_lock.md aspect_ratio such as 1280x720 for 16:9 or 720x1280 for 9:16. Use --preview-profile smooth when the user prioritizes playback polish, or --preview-profile off when the user only wants the core storyboard and prompt package. The default motion style is none so storyboard frames stay stable; use --motion-style center-zoom or --motion-style pan-zoom only when movement is intentionally desired. The animatic preview should include an opening card, ending card, shot overlays, burned-in captions when available, any provided or generated voiceover, and an approved background music bed when one is available. Keep per-shot SFX as cue-sheet guidance for now. Background music remains whole-film post-production audio; final segmented video prompts must still request no generated background music per clip.

Generate the local WebUI state snapshot:

python3 ${SKILL_DIR}/scripts/project_state.py <project_path> --write

The first WebUI is a read-only control surface for project inspection. It displays the entry mode, workflow nodes, storyboard frames, prompt snippets, title-packaging status, and deliverables from canonical project files. To inspect a local project visually, run:

python3 ${SKILL_DIR}/scripts/serve_webui.py --host 127.0.0.1 --port 8765

Future UI edit requests should be recorded as workflow events before Codex reconciles them into canonical files. Do not treat qa/metadata/project_state.json as the source of truth.

The 最终交付/ folder is the user-facing package. Work-in-progress files stay in brief/, strategy/, script/, storyboard/, prompts/, and audio/. Internal machine records such as manifests and fallback HTML belong in qa/metadata/, not the user-facing package.

Step 10: QA

Gate: all deliverables exist.

Run the checklist in references/quality-check.md. If project files exist locally, run:

python3 ${SKILL_DIR}/scripts/validate_video_project.py <project_path>

Fix issues before finishing. Final response should list the output folder, the user-facing deliverables folder, generated image count/status, validation result, and any remaining manual actions.

Reference Files

references/output-contract.md: v2 project structure and file schemas.
references/video-modes.md: mode routing and rhythm rules.
references/scene-director-patterns.md: scene-level director grammars that combine with style templates, visual presets, and model profiles.
references/platform-and-model-profiles.md: platform/model prompt language and constraints.
references/audio-and-copy.md: VO, TTS, captions, SFX, and copy extraction.
references/visual-style-presets.md and references/visual_style_presets.json: lightweight visual look cards for storyboard image prompts and video prompts.
references/storyboard-and-video-prompts.md: image and video prompt patterns.
references/quality-check.md: final QA checklist and validator use.
scripts/project_state.py and scripts/serve_webui.py: local read-only project state and WebUI helpers.