Communitygithub.com

Ethanxwang/shisan-product-video

Provides natural UGC/KOC and concept demonstration creation methods, director prompts, and few-shot examples for product promotions. Ideal for unboxing, trial sharing, and product showcases, helping the agent craft creative and generate content based on product and needs.

What is shisan-product-video?

This skill, named 'Thirteen Product Video Factory,' is designed for AI agents (like Claude Code, Cursor, Codex) to transform simple product promotion requests into actionable image generation and video prompts, executing them within the current creative platform. It emphasizes preserving user aesthetics and selected assets, avoiding generic advertising scripts. The skill supports two primary creation methods: natural UGC for authentic user sharing (unboxing, KOC, daily trials) and concept demonstration for polished product showcases. It includes detailed references for director prompts, few-shot cases, and situational variations, ensuring agents can intelligently plan and produce diverse video content tailored to specific products and user preferences. The skill also provides clear guidelines for input requirements, image-to-video workflows, execution within creative platforms, output verification, and precise case reproduction, making it a comprehensive toolkit for AI-driven product video creation.

Works with~Claude Code~Codex CLI~Cursor
npx skills add Ethanxwang/shisan-product-video

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Thirteen · Product Video Factory Skill

Turn simple requests into executable image generation and video prompts, and complete authorized generation within the current creative platform. Preserve user aesthetics and selected materials; do not turn every video into the same set of advertising copy.

Responsibilities & Production Methods

The Agent is responsible for understanding requirements, deciding the quantity and variation of deliverables, selecting tools, arranging generation, and summarizing results. The Skill provides professional creation methods, complete prompt examples, and quality judgment, without fixing the number of versions, scene combinations, or batch scheduling flows.

Select the method based on the required content. Can be used individually or combined by the Agent within the same task:

What the user wants to doMust-read referenceOutput
Real-person sharing, KOC, unboxing, daily trialNatural UGCNatural explanation frames and video prompts; complete sharing or short behavioral clips as needed
Concept reel, product showcaseConcept DemonstrationProduct assets and concept video prompts; shots designed per aesthetics and product

Read the corresponding method, then select a case closest to the need from the few-shot cases, read its complete prompts — do not use summaries as substitutes. When exploring different usage scenarios, additionally read Situation & Action Variations. Do not load all cases every time; "make one KOC" should not automatically append multiple versions. If satisfactory images already exist, proceed directly to the video stage without regenerating characters.

Minimum Input & Direct Execution

Required: identifiable product image, or already selected holding-shot image/product asset. Optional: selling points, target audience, language, character, scene, duration, model, and aesthetic reference.

  • First observe the image, record product shape, color, size ratio, material, cap/opening, readable brand, and text. Do not fabricate blurry small text. The product identity provided by the user takes precedence over brand names in examples.
  • When the user provides only a product image and a single sentence request, a feasible plan should still be presented first. If unspecified, default to Chinese, adult East Asian trendy creator, 9:16, 720p. The 15-second complete unboxing, 5-second behavioral clip, and 13-second concept reel in the built-in cases are only duration references; the Agent selects based on content and tool capabilities, not fixed templates. User's explicit choices take precedence.
  • When selling points are not provided, design around visible appearance and actual usage actions. Do not fabricate scent notes, longevity, efficacy, user reviews, discounts, or sales data.
  • Do not let the user write technical prompts. You complete the director's requirements, actions, natural dialogue, and shooting relationships. Only ask necessary questions when product input is missing, capabilities are unsupported, or quantity/cost exceeds authorization.
  • User only wants a plan/prompts/preview: output only text, no image generation, no video generation. User requests production: first state the number of images/videos and model for this delivery; respect the current platform's cost confirmation process, do not add extra generations without authorization.

Two Reuse Modes, Not to Be Confused

Remake with the same image: Read the saved complete video prompts and parameters, inherit them verbatim. Use the same selected image, do not use the previous finished video as a reference video, and do not stealthily alter actions, dialogue, duration, or resolution. Generation randomness allows different results.

New product / New requirement: Learn from few-shot cases "how to transform requirements into creative choices and how to write choices into complete prompts." Reorganize actions, dialogue, and shots based on the current product, not limited to replacing template slots; continue to retain already approved characters, scenes, and aesthetic requirements. Product changes must run through all shots and end card; if the new product does not spray, remove the spraying wrist/neck action. For non-perfume products, first read Natural UGC's "Judgment when changing products." Record key adaptations.

Case text and prompts can proactively serve as learning demonstrations; case images are only used as generation input when the user requests "use built-in examples / reproduce cases." Do not replace the user's product with example perfumes. The built-in main reel and behavioral clips use black perfume, the concept reel uses a WANG13 amber bottle — these two sets of product assets cannot be mixed.

Image Generation & Video Workflow

  1. Select a natural and aesthetically pleasing reference image. If the user already has a reference, continue using it; if a reference needs to be found and internet is available, prioritize Pinterest for real photography/natural explanation frames, do not search for posters as UGC substitutes. If internet is unavailable, state this, do not claim to have seen external references.
  2. Convert the reference into text descriptions of scene, character, clothing, camera position, action, and lighting. Default to pure text-to-image generation, do not upload the aesthetic original to the model, and do not use failed generation images for continuous image-to-image modifications. Only switch when the user explicitly requests an image reference, and record the input method. If the product image is actually used as an image constraint input, it must be honestly recorded, not referred to as pure text-to-image.
  3. Character static frames should look like a paused moment in an explanation, not defaulting to selfie with phone, bottle pressed against face, or label facing the camera. Preserve delicate facial features, voluminous hair, outfit, and real skin; natural does not mean removing character aesthetics.
  4. After the image is selected, generate using this single image + complete video prompts. Write pre-unboxing, bag-rummaging, etc., into the video prompts, do not default to creating separate character boards/first frame per shot. A single image can be used for identity and environment, not necessarily replicating the first frame's pose; if the tool forces a first frame, explain this limitation and adjust pre-actions rather than claiming the model can ignore the first frame.

Execution in the Creative Platform

  • Use the host's actually provided material, image, video, and audio tools; first read the tool's parameter constraints, do not fabricate APIs, tool names, or model IDs. This Skill does not install CLI, does not require account tokens, and does not call external private paths.
  • The Agent selected by the current user is responsible for planning. Image and video generation use the model selected by the user on the platform, do not substitute without authorization. When reproducing built-in cases, keep the selected image, complete prompts, duration, and aspect ratio consistent; when using a different model, follow the director's structure, the visuals and examples will differ, requiring re-viewing and re-selecting images.
  • 720p is the regular target; if the model does not have this setting, explain available settings and confirm, do not automatically upgrade to 1080p.
  • Use the generated reference image as video input, not directly using a regular white-background product image as the character's first frame. For UGC, enable the model's native dialogue/audio (if supported); if not supported, explain alternatives, do not claim dubbing as native synchronized audio.
  • Default to generating once per authorized version. When there is failure, insufficient quality, timeout, or uncertain returns, first check existing tasks, do not automatically resubmit. When a retry is needed, point out specific issues and adjustments, and wait for the user's decision.

Output & Verification

When showing results to the user, keep it brief: output, main action/visual changes, visible issues, next steps. Retain in the background: input images, complete image/video prompts, actual model and parameters, task status, original final output, and modification records for each version.

  • UGC: Character and product are continuous; actions are actually happening; dialogue matches actions, allowing pauses; face-hand relationship is natural; not a single long-distance bottle-holding script-reading shot; variants differ in usage behavior as well.
  • Concept reel: Real glass and material feel; rotation, camera movement, focus changes are described and visible; fast/slow, sharp/soft focus have contrast; no continuous glaring flicker; do not automatically make it a smooth high-end commercial just because it's "cool."
  • Verification distinguishes between: file playable, visual actions, audio track/lip sync, user aesthetic confirmation. Do not claim to have heard the audio without listening; do not write "pass verified" without running on the current model.
  • If the final output needs post-processing, only handle user-approved shot selection, assembly, or audio. Select shots at original speed, do not speed up all versions to fit a fixed duration; retain original files and in/out points. Do not attach music of unknown origin.

Few-Shot & Precise Reproduction Entry

Case Index organizes demonstrations by "Requirement → Creative Choice → Complete Prompt → Key Effect." For new tasks, learn from judgment methods, do not fixedly copy product, character setting, dialogue, number of shots, or colors. Only when the user explicitly requests to reproduce the original case, use the accompanying images, complete original text, and original parameters; generation randomness still applies.

Related Skills