Product Video Script

Turn product images and optional selling points into platform-ready short-video scripts, storyboard notes, and generation prompts.

Core Rule

Use research-enhanced generation when possible, but never make research a hard dependency.

• If the user asks for recent, trending, top-liked, platform-specific, or competitor-based scripts and the agent has browsing/search tools, research first and cite sources.
• If browsing/search is unavailable, blocked, or the user wants speed, use the fallback playbooks in references/direction-playbook.md and clearly label the result as 未联网版本.
• Do not invent exact rankings, likes, creator names, or video URLs. Say when exact metrics could not be verified.

Workflow

1. Read the product input

   • Inspect all provided images when image understanding is available.
   • Extract visible product category, appearance, use scene, likely audience, giftability, price/activity clues, and visual risks.
   • Incorporate user-provided selling points. If selling points conflict with the image, mention the uncertainty.
   • Campaign language is opt-in, not forbidden. Use 618, 双11, Black Friday, Christmas, summer, back-to-school, discounts, limited-time offers, or similar campaign framing only when the user explicitly requests it, selects a promotion direction, or the product image clearly contains that campaign. Otherwise keep the script evergreen.

2. Offer directions if the user has not chosen one Present 5-7 concise choices and wait for the user:

   • 带货口播
   • 场景种草
   • 促销活动（only when the user wants a campaign/deal angle）
   • 测评对比
   • 海外 TikTok 风格
   • 直播间真人带货
   • 商业广告/TVC

3. Research or fallback

   • For research mode, search the chosen category + platform + time window + commerce intent.
   • Summarize reusable patterns: hook, pacing, camera style, proof, offer, CTA.
   • For fallback mode, read references/direction-playbook.md.

4. Write the script Use references/output-template.md unless the user requests another format. Default duration is 15 seconds and default ratio is 9:16 when not specified. Include:

   • concept title
   • why this direction fits the product
   • timecoded script
   • voiceover
   • shot/action
   • on-screen text
   • visual/reference notes
   • optional AI video generation prompt

5. Offer next actions After delivering a script, ask or infer the next step:

   • After one script, use the single-script next actions from references/output-template.md.
   • After multiple scripts, use the multi-script next actions from references/output-template.md.
   • For 商业广告/TVC, use storyboard-first as the internal video handoff workflow, but keep the user-facing action as 直接生成视频.

6. Quality checks

   • Hook appears in the first 1-3 seconds.
   • Each claim is supportable from visible product traits or user-provided selling points.
   • Campaign context is explicit. If requested, use it confidently; if not requested, keep the script evergreen and do not add shopping-festival language.
   • On-screen text is short enough for AI video generation.
   • Avoid dense subtitles, tiny price tables, and complex text overlays when the next step is AI video generation.
   • Keep platform tone natural: TikTok should feel native and slightly messy; commercial ads can be more polished.

When Coordinating With Video Skills

If the user asks to generate the video after the script, pass the script and reference-image notes to the video generation skill/tool. Preserve user wording and the selected direction. When image-to-video tools have trouble with text, recommend a no-text or minimal-text version and rely on voiceover.

When the product, platform, campaign context, or language changes from prior work, recommend a clean project/session for the video tool. Add an explicit isolation line to the handoff prompt: This is a new standalone product task. Do not inherit any previous product, campaign, 618, 双11, price, or activity context unless repeated in this prompt.

For multiple scripts, do not merge all scripts into one vague generation request. Create a numbered batch manifest: each item must include concept title, duration, ratio, reference image, full voiceover, shot plan, text-overlay limits, and success criteria. The video skill/tool should submit one generation task per item unless its verified metadata supports multiple video outputs.

For multiple 商业广告/TVC scripts, use serial batch execution by default: submit one TVC, wait for its storyboard images and final video to complete, record the result, then submit the next TVC. Do not submit multiple storyboard-first TVC jobs into the same video-tool session at once.

For 商业广告/TVC, use a storyboard-first handoff by default: generate 5 key 9:16 storyboard images first, then use those storyboard images together as reference images for one complete video. See references/video-handoff.md.

References

• Use references/direction-playbook.md for direction-specific structures when research is unavailable or when you need a baseline.
• Use references/output-template.md for the default script output format.
• Use references/research-guide.md for safe research behavior and source handling.
• Use references/video-handoff.md when the user wants to generate one or many videos from the scripts.