Communitygithub.com

product-demo-video

Turn a product demo or prototype into PRD walkthroughs or an assembled product video. Inspect runnable interactions or supplied design states, propose ordered scenes, and guide the user through approving, making, and reviewing one scene at a time. Supports labeled Figma state demonstrations, optional Doubao narration, reusable media, and MP4 export.

O que é product-demo-video?

product-demo-video is a Claude Code agent skill that turn a product demo or prototype into PRD walkthroughs or an assembled product video. Inspect runnable interactions or supplied design states, propose ordered scenes, and guide the user through approving, making, and reviewing one scene at a time. Supports labeled Figma state demonstrations, optional Doubao narration, reusable media, and MP4 export.

Funciona com~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/baijianyu1024-cmd/product-demo-video/tree/main/skills/product-demo-video

Perguntar na sua IA favorita

Abre um novo chat com esta habilidade de agente já pré-carregada.

Documentação

Product demo video

Start with the complete demo and its entry point, assets, or prototype link. Use any accompanying PRD for context. Walk runnable interactions and inspect related source or states when needed. For Figma designs with incomplete links, inspect the relevant before/after states and propose a labeled state demonstration using the capture guidance.

Before proposing scenes, establish the requested feature and audience. A design page can contain several unrelated features. Carry explicit user choices forward. Propose the few missing project choices together: on-screen copy language, narration language and method, cover wording and sound, and composition. Draft explanations in the user's conversation language unless requested otherwise, keeping the product UI language. Reuse an approved voice setup. Show the chosen settings before the first render.

Choose Chinese or English

Before the cover, offer a Chinese or English video and confirm the missing language choices with the other project settings. A choice such as “英文版” sets both on-screen copy and TTS to English; allow separate choices such as English copy with Chinese TTS. Carry explicit choices forward. The conversation language and the original UI language do not determine the video language. Generate only the selected version unless the user requests both.

The on-screen choice applies to cover text, scene titles, explanatory headlines, authored annotations, prototype labels and chapter titles. Keep product names, URLs and quoted UI labels accurate. Preserve the original product UI unless the user requests and supplies or authorizes its localized variant. Keep review headings and explanations in the user's conversation language, and show the exact selected-language copy and narration for approval using the confirmation format.

Record the on-screen and TTS languages in the project checkpoint. For Chinese TTS pass --language zh-cn; for English pass --language en, using a voice approved for that language as described in the narration guide. When switching languages, review the translated wording, line wrapping, narration duration and pauses before rendering that version. Reuse the source states, captures and approved motion; recalculate scene timing and final progress segments for its actual audio. If both versions are requested, keep separate wording, audio, scene manifests and exports with zh/en suffixes.

Guide the user one step at a time

  1. Propose the ordered scenes. Name each scene and the result the viewer will see. Keep the same sample item or scenario where it connects two scenes. Confirm the order and whether the user wants PRD clips, an assembled demo, or both. Treat scope, order, and deliverables the user explicitly specified as confirmed.
  2. Make and confirm the cover first. For a full demo, propose the cover title, appearance and exact narration or silence, then generate a low-cost still for visual approval before making body scenes. Render its motion or narration after approval, reusing that artwork. The cover establishes typography, palette and spacing. Keep it as an unnumbered opening without a product interaction or body progress bar. Body numbering starts at 1. Honor an explicit request to skip a cover or deliver only a standalone PRD action clip.
  3. Confirm the current scene in chat. Use the scene confirmation format: scene title, 顶部文案 followed by the exact line in a blockquote, and TTS · voice name followed by the exact narration in a separate blockquote. Then briefly describe the starting screen, action, and result. Show the next scene's proposal after the current scene is accepted. Keep the confirmation in chat.
  4. Make and show the current scene after approval. Treat the first body scene as the interaction and framing sample, following the approved cover style and reserving space for the final progress footer. Let the user judge framing, text size, click feedback, result dwell, and voice or silence on the actual clip. Revise the relevant picture, copy, or audio, and reuse approved material when it still fits. Wait for the user's acceptance before making the following scene. Accept a clear approval or exact edit with permission to render without asking again.
  5. Deliver the requested scope. An approved standalone scene can be delivered as a PRD clip. For a full demo, assemble approved scenes with a visible segmented progress bar along the bottom. Use one segment per body scene, sized by its actual exported frame duration, showing completed, currently filling and upcoming parts. Exclude the cover from body numbering and progress. Read render and export for the assembly options. Show a preview for a pacing pass, then export the agreed cut. Look for repeated input, long waits, and unnecessary backtracking in the joined preview. Keep the visible result that proves the interaction.

Guide the user through the decision they can make now. For example, ask them to approve the current scene's wording and action before capture, then show the playable clip and give them the choice to revise it or proceed. Read-only exploration can happen while the plan is taking shape.

Before writing each proposal, check the product relationships involved: an item versus its steps, the trigger for an entry, which fields are prefilled or editable, and what stays visible when a panel expands or scrolls. Use the demo's evidence and confirmed user corrections. Ask a specific question when a missing behavior would change the scene. Carry the answer into later scenes. In the proposal, include the connection to the previous scene when needed. If a scene starts from a list after the previous one ended in details, show or briefly explain returning to that list.

Narration setup, if chosen

Choose the narration method once per project. A PRD clip should communicate its main action and result when muted. Use a user-provided recording or previously approved audio when suitable. For AI narration, read the Doubao setup and synthesis guide and check the local binding before the first voiced scene. Reuse the configured key, approved voice, and pace. Include a proposed change of voice or provider in the user's current review. Ask for key setup or an audition only when needed. Synthesize the approved line for the current scene and reuse it across picture edits. Keep credentials in the supported local secret source and retain the resolved voice settings with the audio.

Record, edit, and verify

  • Use the runnable prototype and a reproducible sample state. For capture methods, orientation, click timing, and demo data, read capture guidance.
  • Keep original captures or exported design states in full. Default to one product view. An approved single-view push-in or pan can emphasize a detail while retaining the context needed to understand the current action. Keep one complete sentence on screen across the action. Show the key click or gesture and leave the result on screen long enough to read. Follow approved composition, text, and pacing across scenes. Mark prototype or simulated outcomes when their status matters to the viewer.
  • Keep raw clips or design states, approved narration, and editable scene timing in the task's working area. For a multi-scene project, keep a short checkpoint with the approved order, shared sample state, on-screen and TTS languages, narration settings, scene versions and wording, confirmed behavior corrections, and the current review step or explicit pause.
  • Read render and export for the local helper and its scene input. Default to a 16:9, 1920×1080, 30 fps H.264 MP4, with AAC if applicable. Use the raw capture and approved audio for the final render when available. Use the resolution of the source footage when judging readability.
  • Before delivering a scene, view the starting screen, key action, and result. Check the behavior this scene is meant to demonstrate, such as a preserved answer, updated step indicator, editable prefill, or independent scroll area. Check sample data consistency, the actual click target after camera changes, on-screen text, voice alignment, speaker level, and complete spoken words. Listen to every occurrence of an uncertain product name using the narration checks. Verify the MP4 plays through. When assembling, check continuity, timing, chapter cuts and the visible progress bar in the exported file: no bar on the cover, correct body segment count, correct boundaries and complete fill at the end. Give the user the playable file or its absolute path.

Example

In the Clips demo, “click the citation and see the two notes used in the answer” is one scene. Start with the answered question visible, mark the citation click, then hold on both source notes. Use one on-screen sentence across the interaction. A subsequent scene can start from that same answer and save it as a new note. Reuse its approved recording and final title when a later scene demonstrates finding that note.

Habilidades Relacionadas