冬のラブレター｜雪原のシネマ短編 · публичный навык Open Skill

冬のラブレター｜雪原のシネマ短編

顔写真を1枚もとに、雪原を舞台にしたオリジナルの恋愛短編を制作します。柔らかなフィルムの光、自然な演技、現在と思い出の対比を通して、ひとりで過ごす今と、ふたりで笑い合った時間を描きます。 元の30秒の物語をもとに、10秒・15秒・30秒の3つのバージョンを用意しました。各バージョンには、単独で使える完全版プロンプトと、カットごとの時間配分が含まれています。 ・10秒版：5カット。ひとりの時間、カメラを構える瞬間、幸せな思い出を凝縮し、短い時間で感情を伝えます。 ・15秒版：7つの主要カット。風に舞う封筒を受け止める動作や、雪の上で追いかけ合う場面を加え、物語をより豊かに描きます。 ・30秒版：全13カット。現在では、走る、封筒を受け止める、誰もいない場所にカメラを向けるまでを。思い出では、雪合戦、追いかけっこ、一緒にカメラを確認する時間、相手を撮影する瞬間までを描きます。 どのバージョンも「ひとりで過ごす現在→ふたりで笑い合う思い出」という構成を保ち、長さに合わせてカットと動作を組み直しています。使用するモデルが対応する長さに合わせて、プロンプトを選べます。 標準のスタイルは、短い黒髪のボブ、黒いウールコート、落ち着いた赤いマフラー、シルバーと黒のミラーレスカメラ。冬の雪原、柔らかな逆光、35mmフィルムの質感で雰囲気をつくり、人物の顔立ち、衣装、小道具、光の連続性も大切にします。 利用者自身の顔写真をアップロードし、物語に合わせて人物のスタイルを調整できます。標準ではセリフ・ナレーション・字幕を入れず、演技、環境音、映像で物語を伝えます。

@TANFEIZAO

Film

4 3 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4274163/2026-10-15/9f9rf4bk.jpg

# Winter Love Letter — production workflow

Create an original Japanese-inspired winter romance short from the end user's uploaded identity reference. It contrasts a woman alone in the present (A) with joyful memories with an adult man (B). Do not copy a named film's characters, dialogue or music.

## Language and inputs
Respond in the language of the user's current message. Publication language never overrides conversation language; this creator communicates in Chinese even when testing the Japanese listing. Default no dialogue, narration or subtitles. Ask only for missing essentials: identity photo and a target duration of 10, 15 or 30 seconds. All three default to an A+B present/memory story in landscape 16:9. The midpoint hard cut is at 5, 7.5 or 15 seconds respectively. A-only or B-only 15-second stories and cross-cut edits are optional only if requested. Never embed a creator's private reference as a preset for all users.

Read references/storyboard-prompts.md, references/input-continuity.md and references/acceptance.md before production. Select the complete 10-, 15- or 30-second prompt from storyboard-prompts.md before generation. Preserve its timing and intentionally reduced action list. Plan multi-clip work only when needed and expose intermediate clips with clear A/B names.

## Model preflight
Call list_models for video. Report the exposed channel, mapped model if available, durations, resolution, aspect support and credit estimate if exposed. If the user requires a specific model/version or matching the original, verify that exact version before generating. A generic channel or a marketing statement is not task-level proof. If unavailable/unverifiable, pause generation and ask for an explicit alternative; never silently substitute Mini/Fast. Use only documented model/channel parameter values. Never invent a Seedance selector. Do not repeat paid generations without a concrete failed-shot reason.

## Identity and look
Use the uploaded image only for facial identity, proportions and age impression. Default character is an adult. Do not inherit studio background, grey camisole or long hair from the reference. Preserve the original story styling: black, airy jaw-length bob with wispy bangs and subtle bends, black below-knee wool coat, muted deep-red chunky scarf, black winter boots. The user can explicitly override styling. A single silver-black compact mirrorless camera with a narrow black neck strap stays with her. Never a phone, selfie, V-sign or duplicated camera.

## Generation and edit
Prefer a verified reference-to-video channel with the uploaded identity bound as a tool input, not merely mentioned in prose. Generate supported-length clips, each with one main action and enough handles to trim. The storyboard timings are edit targets, not claims that a model can generate the entire 30 seconds in one call. If multi-image support exists, bind consistent accepted character references; otherwise disclose continuity limits and correct only affected shots. No forced new-image stage. Optional new image generation requires verified image2 and native 4K support when requested under this creator's production settings; if unavailable, skip it and use uploaded material.

For one-call generation, use the selected complete prompt and bind the uploaded identity as a real tool input. If the user asks for one test, generate exactly one supported clip and stop for review. For the 30-second multi-clip workflow, generate A1–A6 and B1–B7 as storyboard_videos; for shorter edits use the corresponding shorter shot list. Repeat identity, look, camera, light and prop locks in every call. Verify each shot, trim to the selected table and assemble as final_videos only when needed. Hard-cut A to B at the selected midpoint, without dissolve, white flash or morphing. Do not silently substitute A-only for A+B or a shorter duration for the request.

Use natural winter sound, footsteps, breathing, cloth/paper and camera shutters. A has one person; B