unsketch

View saved version →

---
name: caption-video
description: 動画をカットし、音声から時間付き字幕を作成・修正してMP4で保存する。動画の仕上げ・字幕編集に使う。
---

Open open_video_editor. The user imports the actual video or adds an existing generated output through its widget. Never invent local clip IDs, durations or captions for unheard speech.

1. Read unsketchEditProject from widget context. Apply requested trims and reorder with edit_video_project using exact IDs and validated source times. Captions use source-relative seconds, so trimming does not shift their meaning.
2. For speech transcription, direct the user to the explicit 音声を送って自動字幕 button. It sends only the selected audio range to transcription, at most 60 seconds per clip and six new clips per account per day. Existing jobs are recovered, never resubmitted. This allowance does not spend generation credits. Do not call the app-only transcription tools yourself or claim recognition from silent frames.
3. Alternatively, the user imports SRT/VTT or supplies a timed transcript. Use clip.captions=[{start,end,text}] for timed cues and clip.caption only for a fixed overlay. Keep all cues within source duration. Preserve the user's wording unless correction or translation was requested.
4. Let the user review transcript text and timing; correct recognized names without inventing speech. Apply edits through edit_video_project. The widget offers MP4 by default and WebM when selected; export depends on browser codec support.
5. Preview and save after export. Do not call subtitles complete merely because text was inserted. If source audio is absent or unsupported, report the limitation and use the supplied transcript/SRT/VTT without guessing.

No new media generation is required for cuts or captions. Do not offer purchases or publish media.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: image-to-video
description: 商品や人物の画像から短い動画を作り、動きや構図を修正して保存する。画像の動画化・短い商品動画に使う。
---

Preserve explicit user requirements. Establish the subject, desired motion, destination and duration from the conversation; ask only for missing consequential details. Do not invent product claims.

1. For a chat attachment use import_chat_images with medium video. For a generated image use the exact source messageId/index and conversationId from get_generation. Never infer image IDs or substitute another output.
2. Use list_video_models to choose an available image-capable model. Choose only supported aspect ratios, duration and audio options. Describe one feasible shot, camera motion and what must stay unchanged. Avoid promising lip sync or identity consistency beyond the model's results.
3. Call prepare_video with the reference and prompt. The user clicks the priced generation button. Poll/recover the same job with get_generation and never duplicate a pending request.
4. Review the resulting motion and preservation of the reference if the output is visible. Otherwise ask the user to assess it. Use the existing output's conversation and source for a supported edit; new generation requires another quote and priced click.
5. Deliver the video save action. If the user wants a cut or captions, use the widget's video-edit action and edit_video_project with actual imported clip IDs/durations. Finish with MP4 export; do not claim the saved file exists before rendering/download succeeds.

Template registration belongs in Studio. Do not initiate checkout.

Referenced files: 1 — file references are metadata, not package contents.

View saved version →

---
name: product-ads
description: 商品写真から広告画像の案を作り、選んだ案を修正して保存する。商品広告・SNS告知・広告画像の複数案に使う。
---

Use the user's language and supplied brand facts. Follow explicit instructions over defaults. Ask only for missing facts that affect the image: product, audience, destination, mandatory text. Do not invent discounts, endorsements, performance claims or product features.

1. Import user-attached product photos with import_chat_images using host-issued files. Preserve returned attachmentIds. If no photo exists, ask for one when product likeness matters; otherwise proceed with an explicitly agreed concept.
2. Propose up to three distinct visual directions in a compact list (for example product close-up, use scene, graphic layout). Use list_image_models to select an available reference-capable model; do not require the user to know model names. Use a suitable aspect ratio from its actual capabilities.
3. Use prepare_image with the chosen brief and reference IDs. For distinct directions, prepare separate quotes; count>1 with one prompt does not guarantee distinct designs. Every generation needs the widget's priced click. Never claim generation has started from a quote.
4. Recover each existing job with get_generation; never resubmit because polling is slow. Ask the user to choose a result by its actual output index. Revise with prepare_image and the owned conversationId/source messageId/index so the product photo is retained.
5. Check the result against supplied identity, claims, text, composition and destination. If you cannot see the media, ask the user to confirm; never invent a quality inspection. Point to the per-output save button. Offer image-to-video only when requested.

Template recipes are private. Registration/publication belongs in Studio. Do not offer purchases or commissions in this plugin.

Referenced files: 1 — file references are metadata, not package contents.