When installing or determining producible capabilities, first readInstallation and Dependencies. Code and necessary resources are provided with the package; Python, FFmpeg, browser, Node.js, Jianying, and generation services are prepared separately according to the route. Do not read the developer's local tools or private case directories.
First read Storyboard Engineering Guide、Production Stage and Material Contract. Knowledge and tutorial tasks also use Knowledge Director。
First choose the output route
Before plan design, clarify whether the user needs an MP4 rendered by the browser or a draft exported by the user with Jianying. If unclear, first introduce these two options and ask, and do not choose for the user by default; new CLI must fill in --output-target.
- web：Browser Production Skill, which can choose seven types of storyboards according to the actual implementation, produce program animations, and output MP4.
- jianying：Jianying Production Skill, by default limited to ST-001～004 and checks the specific native configuration, outputs a draft + material package, and the user exports in Jianying.
Jianying Draft Box Pre-check Requirements: When the user chooses the Jianying route or the prompt contains "Rendering method: Jianying", the Skill must first check whether the user has configured the Jianying draft box path (check hvp jianying-config status); if the user has never filled in the Jianying draft box folder location, it must first proactively ask the user and obtain the absolute path of the draft box (prompt the user to check and tell it in Jianying global settings, and save it to the local configuration). It must not proceed with subsequent plan confirmation and draft delivery without obtaining the draft box path.
Script, style, color scheme, Flow tasks, plan confirmation, material review, voiceover, and subtitles share this master control process. The output route enters the complete plan, and is not checked only at the time of compositing; configurations unsupported by Jianying must not be automatically converted to web pre-rendering. Specific shot pre-rendering requires real user consent and the current plan signature, seeJianying Rules。
Unique Classification
Use hvp catalog --kind shot to read catalog/shots.json. The shared directory has only seven types: ST-001 single-image animation, ST-002 video, ST-003 background + transparent material animation, ST-004 full-image limited-frame animation, ST-005 Remotion animation, ST-006 simo version whiteboard animation, ST-007 SRT version whiteboard animation.
Each shot chooses one main type, and multiple layers can be used internally. AI designs the audience, goal, script, and overall visual and motion direction according to the requirements, then selects the type shot by shot. Art style, image type, and color scheme are common visual settings; subtitles, sound, transitions, and output method are shared settings. Business names must not be used to select the implementation.
ST-001 single-image animation lists only full-screen mode and picture book mode: the former organizes the full-screen image with a complete main image, and the latter uses a background, illustration area, and text area. Keyframe animation and environmental effects are both optional; environmental effects are checked according to the specific implementation and are not added by default. These two modes belong only to ST-001, and no new layout gallery or other mode directory is added.
Picture Book Video Request
When the user says "make picture book videos", "animated picture books", or "narrated picture books", they must readPicture Book Production Guide, design a multi-page story according to the ST-001 picture book mode; both the browser and CapCut routes can be used. By default, reference the reading layout of "The Eaves of a Leaf": paper-page whitespace, a complete illustration centered, a short sentence below the illustration, and an optional subtitle; do not default to a left-right image-text explanation column. What is referenced is the layout and reading rhythm; the story, characters, art style, and color palette are designed according to the needs of the current issue, and the hedgehog or watercolor style is not fixedly reused.
First confirm the output route, then submit a complete plan containing each page's events, illustrations, short sentences, voiceover, page intervals, and transitions. When no image type is specified, by default use one complete illustration without text per page, with titles/short sentences/subtitles layered; when an image type is specified, generate complete image-text pages according to the hand-drawn image generation rules, and do not split them without authorization. Use multiple ST-001 shots for multiple pages, and do not classify page turning as an ST-004 character action sequence. For CapCut, create a draft using native keyframes and text; for the browser, create an MP4 using a deterministic time interface; optional environmental effects must comply with the capabilities of each route and the confirmed plan. Use the same asset contract, and do not require the user to first generate an entire picture book video because of the CapCut output.
When any of the seven categories is insufficient, you may write an independent type, asset contract, and HTML implementation in this issue's local_types . First complete technical verification and actual review, and let the user decide whether to include it in the shared library; a new name does not equal a new capability.
Stage gates
Research and plan → user confirmation → visual asset preparation → all visual assets pass review → voiceover and animation → compositing → acceptance.
"Help me make a video" only authorizes the planning stage. A complete plan includes audience, goals, basis, title, script, estimated duration, visual and motion direction, per-shot purpose/frame interval/action/type/source, asset responsibilities, sound, and delivery. Before confirmation, existing assets may be inventoried; visual assets, voiceover, animation code, or previews must not be generated.
After confirmation, illustrations, reference images, screen recordings, and external generation prompts may be prepared. Before all required images, screen recordings, videos, chart data, and action assets pass technical checks and actual visual review, do not generate voiceover, do not write animation for this issue, and do not pre-render or composite. The same applies to subprojects. See stage records。
Storyboard asset preparation and completion
All video and image prompts and image generation follow thehand-drawn image generation rules. When the user specifies an image type, fully execute the rules of the original image type and image-text mode; do not arbitrarily change it to a text-free illustration plus post-production text because of video use. Do not repeatedly overlay the same copy on an already generated complete image-text card; explain separately the non-editability of text within the image and the independent subtitles.
Read according to the ST numbers used in this issue:storyboard asset preparation directoryand the corresponding independent documents;catalog/shots.json'smaterial_guidepoints to that document. Inventory the required assets shot by shot, including specific format, dimensions, transparency requirements, existing files, and missing items, and distinguish pre-production visual inputs from production-stage outputs.
After the plan is confirmed, for every missing asset that can be generated, you must deliver a complete, copyable prompt for that shot and save it to this issue'swork/material-prompts/; at the same time provide the asset number, file name, format, actual reference image, return location, and review points. Do not deliver only generic templates or ask the user to find images themselves. When reference images are missing, first create a reference image task, and after it passes review, generate the related action frames. When real screen recordings, factual data, user originals, and actual subtitle timings are missing, provide collection or supplementation requirements, and do not pass off generated content as real. Only after all inputs pass review may you proceed to voiceover and animation production.
Asset strategy and external video
auto may contain zero videos;no_video prohibits video footage input;video_only Each shot's main picture comes from the original-speed video, and text/sound/subtitles can be added. The internal shot cache is not an external video input, nor does it bypass policy. Operation tutorials use real screen recordings as operation evidence; generated videos are only for illustration or narration.
When there are Flow or other platform generation tasks, deliver independent prompt and reference image task packages, and wait for the user to send them back. Save versions, candidates, hashes, review and selection records. If not sent back, too short, damaged, or unqualified in identity/action/continuity, production is blocked. Do not loop, freeze, or change speed to pad duration.
Implementation and output
New projects only use composition-v2。episode.json as the full-film plan,composition.json as shots. Old projects, finished films, and original materials are retained, not executed or converted.
new / validate / render / build / export-plan / export use the same model.export use the prepared voice, without calling voiceover.build can be executed according to explicit voiceover settings, but the specified service must not be silently replaced. Volcano Engine uses the bundledVolcano Voice Skill，hvp ttsdispatch by provider and generate actual subtitle timings.hvp volc-configSave the account submitted by the user,hvp volc-statusonly view configuration status; keys are saved in the user directory outside the repository, and updating the Skill does not overwrite them.
Each frame is recalculated from the frame number. The web implementation retains renderFrame(t); CSS animations are paused and positioned, and JS does not depend on wall-clock time. Remotion and stateful whiteboard generate content-addressed caches shot by shot, then decode and composite frame by frame. For SRT whiteboards, review the line art and input data first; when there is no SRT, after review passes, do voiceover, then align with real subtitles, without guessing timings.
Jianying rules: full-film output and per-shot implementation are recorded separately, but must obey the selected route. Default to Jianying native; when native is not possible, block and report, without automatic pre-rendering. Only after the user agrees to specific shots may video caches be generated per shot; explicit native requirements still cannot be fallen back from. Do not forcibly rewrite effects.
Jianying delivery uses the built-in pyJianYingDraft and the absorbed Jianying editing methods, and does not read external reference projects. The in-package offline checker and installer depend only on the Python standard library; all playback materials are actually packaged and installed only after rebinding checks.hvp jianying-check <包或草稿目录>Check files, references, timings, and paths;hvp jianying-resources <关键词> --kind transition|filter|effect|character_effect|intro|outrosearch names and real IDs, without automatically selecting or downloading. Transitions/effects/filters selected by the plan are recorded in shot.jianying and automatically written into the draft; the supported range, parameters, and transition scope are in the Jianying guide. Native effect revisions participate in the plan signature and require re-confirmation.
Users can use hvp jianying-config --draft-dir "directory" to save the local Jianying draft folder. The configuration is in the user directory rather than the repository, and updates and Skill synchronization do not overwrite it. Installation uses the saved value by default, and --draft-dir only overrides it for this time; do not hard-code user paths into shared Skills or distribute them to delivery packages.
Evidence and delivery
When delivering or correcting previews/drafts, readshot review and delivery checklist. Check the actual case version specified by the user, distinguish original images, full pages, and screenshots; split tracks by editable object, and combine action frames of the same sprite into one track. Update and check previews, actual drafts, installation packages, and instructions separately. When screenshots are restricted or the software has not been reviewed, truthfully list pending verification items; coordinate/hash/JSON passing cannot replace visual passing.
The catalog status must point to actual verification evidence. Technical verification, application opening/export, and user aesthetic acceptance are recorded separately. Deliver MP4/draft, preview, subtitles, cover, keyframes, project, and production records. The storyboard gallery only displays numbers, PNG screenshots, storyboard descriptions, and rendering methods; cards and pop-ups do not include tested cases and do not load videos; do not treat editing proxies as actual handwriting. Skill installation includes only screenshots, excludes demo videos and private-period project materials, and does not download videos based on verification records. See Verification records、Source license。