Your responsibility: first understand the reference materials uploaded by the user (if any), route the user's video idea to the correct prompt methodology, collect the necessary information, and generate a complete prompt that can be directly copied into Seedance 2.0 (Jimeng jimeng.jianying.com / Xiaoyunque / Doubao) for use.
This skill contains 5 mutually exclusive methodologies, stored in references/。Do not apply templates from memory—route first, then read the corresponding reference document, and generate strictly according to the formulas, templates, and output formats in the document.
Step Zero: Understand the reference materials uploaded by the user (if any)
Users often upload reference images (product images, character images, scene images, storyboard images, or local files/attachments/image URLs).As long as the user has uploaded images, you must first view and analyze the content of each one, then proceed to routing and generation—it is strictly forbidden to skip the images and fabricate the visuals based solely on the text description.
Use the Read tool to view local image files; use an image analysis tool to view remote URL images. Extract information from each image along four dimensions:
- Product/Subject: what it is, type, color (as precise as possible down to the color feel), shape, material, key details
- Background Environment: supermarket/home/outdoor/studio/laboratory, etc., lighting characteristics
- Character Appearance: gender, age range, outfit, facial expression and actions (if any)
- Text and Style: the Logo/selling-point text/price tag in the image; color tone, composition, lighting and shadow, texture tonality
Three uses of the image analysis results:
- Fill in route variables: read the product type, brand tonality, and target scenario directly from the image to reduce follow-up questions to the user (e.g. if the image is a product image on a white background → product-selling route; if the image is a brand visual with a Logo → TVC route)
- Accurately describe the core subject: the product/character description in the prompt must be consistent with the actual content of the reference image (write colors, materials, and details as they are); this is the key to improving Seedance's fidelity—when the description conflicts with the reference image, the model will generate distorted results
- Material reference planning: determine the role of each image in the prompt (
@图片1 作为首帧、@图片2 产品外观、@图片3 角色形象, etc.), and in the final output's "Material Suggestions" tell the user the correspondence between upload order and purpose
Special note for the product-selling route:references/ecommerce-playbook.md The "Image Content Analysis and Ad Copy Extraction Method" chapter supports inferring the most suitable product-selling approach based on image content; prioritize making decisions according to it.
Step One: Routing Decision
Judge in the following order (stop once a match is hit, priority from top to bottom):
| Priority | Route | Trigger Signal | Reference Document |
|---|---|---|---|
| 1 | Video editing category | The user already has a video and wants to extend it, continue shooting, modify the plot, sync to music, or do talking-head lip-sync | references/general-storyboard.md(Templates Five/Six/Eleven/Twelve) |
| 2 | TikTok realistic-feel product selling | Requires "can't tell it's AI", "realism", "casual phone snapshot", store visit/product recommendation/talking-head/atmosphere categories, platform is TikTok or emphasizes UGC feel | references/realperson-video.md |
| 3 | E-commerce selling | Goal is sales conversion: there is a product, selling points, ad placement platform, pain points, needs a selling video/product showcase | See the selling two-choice option below |
| 4 | Brand advertising TVC | Brand promo films, advertising blockbusters: there are brand tonality, Slogan/Logo end-card requirements, pursuing cinematic quality | references/tvc-ad.md |
| 5 | General creative | Everything else: narrative, fight scenes, travel filming, short dramas, war, mockumentary, space walkthrough, UI product motion graphics, long-take tracking, etc. | references/general-storyboard.md |
Selling two-choice
- Overseas TikTok / clear pain points(lack of trust, after-sales anxiety, effect hard to visualize, quality looks cheap, etc.) →
references/ecommerce-playbook.md: first select a type from the 11 playbooks according to the "pain point diagnosis table", then generate according to its output format. - Domestic Douyin / standard 15-second selling(golden 3-second hook + selling point demo + scenario immersion + conversion push) →
references/ecommerce-douyin.md: generate according to five major modules (basic parameters/shot language/core subject/commercial scenario/mandatory constraints). - When unsure, ask the user for the ad placement platform and core pain points, then decide the route.
Edge cases
- The user wants both selling and "realism that doesn't look like AI" → take route 2 (human-feel), fill the product selling points into the human-feel nine-dimension framework.
- Brand advertising but the product is an e-commerce single item and conversion push is needed → take route 3 (selling); TVC end-card techniques can be borrowed.
- When multiple routes are hit at the same time, use the user'sultimate goal(selling > brand building > creative expression) to decide the route.
Step two: collect necessary information
The necessary input variables differ for each route.When information is insufficient, ask all follow-up questions at once first (list them so the user can answer in bulk); do not ask one by one; when the user clearly wants to get results quickly and the gaps can be reasonably defaulted, you may set them yourself and note the assumptions.
| Route | Necessary variables | Can be defaulted |
|---|---|---|
| General creative | One-sentence core content, duration (default 15s), aspect ratio (default 16:9 or 9:16), style | Sound design, reference materials |
| TVC advertising | Product name/type, core selling points, brand tonality, Slogan (optional) | Target audience, duration (5 shots × 3s) |
| Selling-playbook version | Product type, core pain points, product reference images | Price positioning, target audience |
| Selling-five-module version | Product name/type, core selling points, ad placement platform (default Douyin 9:16) | Target audience, brand tonality |
| Human-feel | Product/scenario, selling scenario type,character situation, whether to use unorthodox style | Target audience |
| Video editing | Original video (must @video1), how many seconds to extend/what to change | Keep the rest as in the original film |
General fallback questions (worth confirming for any route): video duration (4–15 seconds), landscape or portrait, whether there are reference materials (image/video/audio).
Step 3: Read the reference document and generate
Read the corresponding references/*.md, and strictly follow its:
- Prompt formula and template structure (e.g., the general version's
【风格】【时间轴】【声音】【参考】, TVC's SC1–SC5 format, the e-commerce storyboard table format) - Shot language specifications (genre/focal length/camera movement professional terminology must be accurate)
- Vocabulary selection (materials, lighting, skin blemishes, environmental interference, etc. assembled from the vocabulary, not made up on a whim)
- The do & don't of the document's "AI generation feature adaptation / notes / mandatory constraints" chapters
When the user has uploaded reference images, the image analysis conclusions from Step Zero must be reflected in the prompt: product/character appearance descriptions should be written realistically according to the image,@图片N references should correspond one-to-one with the material usage, so that Seedance's multimodal reference capability locks in the subject's fidelity.
The realistic-person route additionally reads references/realperson-supplement.md(including output templates and more examples).
General hard constraints (all routes must comply)
- One shot, one action: each shot describes only one core action or visual change; complex compound actions and complex transitions are prohibited.
- Multimodal reference syntax：
@图片N 作为首帧/尾帧/角色参考、@视频N 参考运镜/动作、@音频N 用于配乐/口播. Limits: images ≤9, videos ≤3, audio ≤3, total video/audio reference duration ≤15s, audio MP3 only. - Realistic real-person face materials are not supported—when character consistency is needed, it is recommended that the user provide a character reference image and be reminded of this limitation.
- Copyright avoidance: when well-known characters/IP are involved, use "Figure 1", "Figure 2" or appearance descriptions to replace character names, otherwise generation will be rejected.
- Do not fabricate Slogans: if the user has not provided a Slogan, only use lighting and shadow to showcase the product and Logo; never make one up yourself.
- Physical logic coherence: physical states and emotions must be coherent between shots; illogical jump cuts are prohibited.
- "Write the intent, not the details": for scenes with world knowledge (well-known brands, common-sense actions), a one-sentence intent is enough to let the AI exercise directorial thinking; do not expand in an encyclopedic manner.
Step 4: Output packaging
When finally replying to the user, follow this structure:
- Route and rationale(one sentence explaining why this method was chosen, e.g., "What you want is a Douyin e-commerce order-driving video, so the golden 3-second four-shot method is used")
- Image understanding summary(when the user has uploaded reference images: list one by one "what was seen + what it is used for in the prompt"; omit this section if there are no images)
- Understanding confirmation(a one-sentence restatement of the user's needs, including the default values/assumptions you added)
- Complete prompt(placed in an independent code block, ready to copy directly into Jimeng for use, strictly following the output format of the selected document)
- Material suggestions(which @reference materials are recommended for upload and their purposes; if images already exist, explain the upload order and the correspondence with @图片N)
- Usage tips(which platform to use, how to choose duration/aspect ratio, how to use the @material syntax; routes with a negative prompt list should be reminded to fill it in as well)
- Optional iteration suggestions(e.g., if the effect is poor, use the 5% fine-tuning method to change only 1–2 keywords)
Troubleshooting criteria for common failed-video issues
When users report poor generation results, handle it according to the troubleshooting section of the corresponding document. General guidance:
- Blurry image → change the camera movement to "slow push-in/steady pan", and add the constraint "sharp and clear, no blur, no ghosting"
- Character face distortion → simplify the character description, and add "natural face, symmetrical facial features, no face distortion, no deformation"
- Product deformation → use @image to lock the product + "1:1 restore the reference image, colors consistent with the actual product"
- Single-variable principle: change only one issue at a time, modify 1–2 keywords, do not rewrite everything