执行前先读取同套件 chatcut-plugin-basics 和 widget-forms，检查 ChatCut 连接器，确认目标项目并使用千问支持的提问方式。不能直接输出其他宿主的交互标签。

App Promo

What this skill covers

Create professional short promo videos for software products: SaaS, apps, and websites. Confirm the target duration with the user before scripting.

Two generation paths, which may be combined with separate TTS and music when the confirmed brief calls for it:

Motion Graphics (MG) - Generate segmented MG clips, TTS voiceover, background music, and timeline assembly. Best for high information density, text emphasis, and strict brand consistency.
AI video - Generate video using the models and duration limits currently supported by the video-gen skill and tool schema. Best when visual impact matters most, a realistic scene texture is needed, and text requirements are light.

Required inputs

Before production starts, collect:

Product information - Product name, main selling points, target audience, target platform, etc.
Visual assets - Logo (required), product screenshots, feature images, demo video.

The source can be either path. Ask which path the user wants as the first step of Phase 1:

A. Product website URL - The agent uses Firecrawl to extract information and assets automatically.
B. User-provided materials - The agent asks for product information and has the user upload assets.

What shapes the promo

Before writing the script, align with the user on these dimensions:

Target platform + length + aspect ratio - YouTube, Twitter, WeChat Channels, TikTok, etc.
Tone - Premium, energetic, minimal, cinematic, etc. Clarify vague wording with concrete options in the user's language.
Generation path - MG or Seedance, based on asset type and style preference.
Audio driver - TTS voiceover mode vs music-only beat-sync mode.
Narrative structure - Introducing, Problem-Solution, Copy + UI Loop, Pure Visual Impact, Feature Deep Dive. See the narrative-mode table in Phase 2.

Preset mode, if present

If the prompt context contains <preset_directive>, the user selected a preset from the template entry. The preset also provides these context variables, and you must follow them instead of bypassing them:

<creative_guidance> - Creative direction template: style, narrative, visual tone. In Phase 2, adapt the script to this direction by writing a tailored version; do not freestyle.
<execution_path> - Generation path identifier: MG, Seedance, etc. In Phase 3, follow this path and load the matching reference; do not ask the user to choose MG vs Seedance again.
<reference_assets> - Internal production reference assets. In Phase 1, do not register these in the asset library because the user should not see them. In Phase 3, pass them directly as URLs for generation, through the reference fields supported by the current submit_video schema.

What still needs user alignment - The preset already fixes creative direction and execution path, but you still need to align on target platform if the preset did not specify it, duration, and product information confirmation.

Flow

There are 3 phases. You must get user confirmation at every key checkpoint before continuing, unless the user explicitly asks you to run the whole process end-to-end without stopping. Key checkpoints:

Product information and assets (end of Phase 1) - Is the information accurate? Are the selected assets correct?
Creative direction (inside Phase 2, before writing the script) - Are platform, duration, tone, generation path, etc. aligned?
Full script (end of Phase 2, before generation) - Is the script ready to generate?
Generation result (end of Phase 3) - Is the result OK, or should it be iterated?

Do not bundle multiple checkpoints into one reply. Send one checkpoint per reply and wait for the user's response before moving to the next. Do not show product information + creative direction + script all at once and ask for one combined confirmation.

Upstream mistakes make downstream work useless: wrong product information -> wrong script; wrong script -> wasted generation credits.

Phase 1 - Gather and align product info & assets

Ask for the source path - Use the host question UI or a plain-language question; this is a classic two-option choices scenario. Write option labels in the user's conversation language. English example:

Ask whether the user wants to provide a product URL or upload assets. Use a native question control if available; otherwise ask in plain conversation.
Path A: URL
- Call the available web_browser connector tool to extract branding + images + product information. See the Firecrawl command in references/info-gathering.md.
- Extract the logo and product screenshots. See the logo extraction rules and data URI handling in references/info-gathering.md.
Path B: user-provided materials - collect everything in one widget flow: uploaded assets (logo + product screenshots + feature images), product name, main selling points, target audience, target platform, and desired tone. Multi-dimensional inputs + file upload is the textbook widget case; do not split it into multiple rounds. See the widget-forms skill and load it before rendering the form.
Download and batch-register assets into the asset library. See references/info-gathering.md. This step is a hard prerequisite for confirmation: users cannot confirm assets they cannot see in the asset library, and text descriptions such as "img1: front view" are not a substitute. Unregistered assets are invisible to users and cannot be referenced for generation.
Save the brand as this project's Design Style. In a promo the brand is the visual language, so it must live in the project's Design Style, not in chat text. Map branding.colors into colors[], branding.typography.fontFamilies into fonts[], and the extracted logo into logos[], then call manage_design_style with action: "create", named after the product. Read the saved project Design Style and explicitly apply its brand values to direct MG code and video prompts. A promo with a known brand never goes through the catalog style picker - presets are the fallback for a promo whose brand cannot be obtained at all. See the mapping and the fallback rule in references/info-gathering.md.
Confirm product information, brand, and assets with the user. Before confirming, make sure the assets render in the asset library.
- Is the product description / target audience / main selling point accurate? Should anything be corrected or added?
- Are the brand colors, fonts, and logo captured correctly? Name the saved Design Style and list what went into it, so a wrong extraction is caught here rather than after generation.
- Are the logo and product images the right assets to use? Should any be replaced, removed, or added?

Preset reference assets inside <reference_assets> must not be registered in the asset library. They are internal production references. Pass them directly by URL during generation, through the reference fields supported by the current submit_video schema, and do not show them to the user.

Phase 2 - Plan and confirm script

Align creative direction using the five dimensions in "What shapes the promo". Ask only for key missing inputs and provide concrete choices. If the preset provides <creative_guidance>, adapt to that direction; do not freestyle.

Confirm creative direction with the user - Platform, duration, tone, generation path, audio driver, and narrative structure must be aligned before moving to step 2 and writing the script.

Choose narrative structure based on available assets. In preset mode, follow the structure from <creative_guidance>:

Asset situation	Recommended structure	Notes
Multiple clear feature images	Copy + UI Loop	One-sentence intro -> Slogan -> Feature loop with alternating copy and visuals -> Logo + CTA
Clear pain point to solve	Problem-Solution	Hook pain point -> Product appears -> Outcome after use -> CTA
Product launch / homepage	Introducing	Logo + product name -> One-sentence positioning -> Core capability showcase -> CTA
UI design is the core selling point	Pure Visual Impact	Logo entrance -> Rapid multi-screen switching synced to music -> Slogan + CTA
One killer feature	Feature Deep Dive	Hook + slogan -> Deep demo of the main feature -> Briefly cover other features -> CTA

Write copy. For detailed rules and examples, see references/script-templates.md:
- The first 3 seconds decide everything. Use the PAS framework: Problem -> Agitate -> Solve.
- Word count guide: 15s is about 30-40 Chinese characters or 25-35 English words; each sentence should be <= 10 Chinese characters or 8 English words.
- Say outcomes, not features. One idea per sentence. Read the full script aloud as a self-check.
Plan the storyboard and show it to the user as a markdown table. See references/script-templates.md for format:
```
| # | Time | Visual | Copy | Reference Asset |
```
Each row: Visual (visual content + motion), Copy (on-screen text or voiceover), Reference Asset (@Logo, @img1, or mark pure MG as -).
Confirm the full script with the user. Generation consumes credits and is irreversible. Show the full User Script, then end the reply with the host question UI or a plain-language question for the next step. Write option labels in the user's conversation language. English example:

Ask in ordinary conversation, preserving the choices and waiting for the answer.

Phase 3 - Generate visuals

Execute according to the generation path. In preset mode, route by <execution_path>; otherwise use the choice aligned with the user in Phase 2:

MG path -> references/execution-mg.md for TTS mode, music-only beat-sync mode, brand injection, and timeline assembly.
Seedance path -> references/execution-seedance.md for script-to-prompt conversion, omni_reference, and text-rendering precautions.

In preset mode, pass <reference_assets> directly into generation as URLs, through the reference fields supported by the current submit_video schema; do not route them through the asset library.

Confirm the generation result with the user - Show the result and ask whether it is OK or should be iterated.

Core principles

Brand consistency

The video is a visual extension of the user's product, not an independent artwork.

Brand elements - colors, fonts, border radius, spacing, button styles, etc. - must be extracted from the user's page or provided by the user. Do not invent them.
The project's Design Style is where the brand lives. Save it there in Phase 1 and read it back before writing MG code or generation prompts. Do not assume brand values are automatically injected into direct JSX authoring.
All generation calls, MG or video, must use the same brand specification.
For Web MG, follow create-motion-graphics to author JSX directly. Use confirmed brand elements and storyboard timing; do not call the built-in agent MG generator.

Asset rules

Always extract or obtain original images first. Screenshots are not a substitute for original images.

Logo -> must be extracted or provided by the user. Never redraw it with AI.
Original images -> required for product images, feature images, and template previews. They are the core visual generation assets.
Functional demo assets -> extract if available, including GIFs, videos, and animated previews.
Page screenshots -> optional, only for showing the overall web interface.

Usage rule: Feature showcase -> use original images | Interface showcase -> use screenshots | Logo -> use the real logo.

Pacing

Fast pacing is the core trait of a promo.

One visual change every 2-3 seconds.
15s video: 6-8 change points. 30s video: 12-15 change points.
A single MG should not exceed 3 seconds.

Fast pacing does not mean fragmentation: changes should stay around one theme, use rapid switching to reinforce the message, and keep the visual style consistent.

Copy principles

The first 3 seconds decide everything. Whether it is a pain-point hook, an "Introducing..." title, or a striking visual, the opening must create curiosity or desire.
Say outcomes, not features. "Write smarter. Ship faster." beats "AI-powered writing assistant with style learning."
One idea per sentence. Keep it short, forceful, rhythmic, and natural to say aloud.
Every second must earn its place. A 15s promo has room for about 40 words; delete any sentence that does not add information or emotion.
The ending should close the idea, not hard-sell. The CTA should be a natural conclusion. Do not introduce new information in the last 3 seconds.

Less but better

Do not over-add motion effects or decorations. One right element beats three forced elements.

Tools

Phase 1: information gathering - web_browser connector tool + asset-import for cloud asset registration.
Phase 2: script planning - reasoning + the host question UI or a plain-language question to confirm the script; see the example at the end of Phase 2.
Phase 3: visual generation
- MG path: create-motion-graphics + voice for TTS / music + timeline edits.
- Seedance path: video-gen (Seedance 2.0) + timeline edits.

References

references/info-gathering.md - Firecrawl command details, data URI logo handling, parallel download, batch registration.
references/script-templates.md - Narrative-mode details, copywriting examples, storyboard table format.
references/execution-mg.md - Full MG generation flow: TTS mode, music-only beat-sync mode, brand injection, examples.
references/execution-seedance.md - Seedance prompt conversion, brand integration, text rendering.

Error handling

Scenario	Handling
Firecrawl extraction failed	Check the live web_browser schema for supported retry options. If unavailable or still failing, ask for product screenshots or brand assets.
Logo extraction failed	Ask the user to provide it manually. Never redraw it with AI.
Style extraction incomplete	Use a default modern style plus the extracted brand colors.
MG style inconsistent	Check the actual MG code and saved brand values, then revise the code and verify the rendered result.

制作产品 / App 宣传片

Files5

SKILL.mdqwen/skills/app-promo-guide/SKILL.md