Street Pitch × Supercar Twist · skill pubblica Open Skill

Street Pitch × Supercar Twist

ENGLISH Combine a natural product pitch with a dramatic twist: a front view, side details, a high-angle wide shot, an urgent call to action, a conceal-and-reveal cut, and a final product shot. Keep the presenter and product consistent, the speech continuous, and each cut before contact. What you provide: an adult presenter reference, a clear product image, a vehicle reference, the brand and product name, verified selling points, the setting, the spoken language, and the actual shopping link or button used on your target platform. What is included: guidance for the three reference images, a 12-second six-shot plan, adaptable prompts, continuous original audio and editing guidance, local retiming to fix incorrect vehicle motion, and a shot-by-shot quality checklist. About the sample: this 12-second film was previously generated by the creator in LibTV and personally edited. It demonstrates the creative structure and pacing. The workflow has been adapted for Digen, but generation with new inputs and reuse results on Digen have not yet been tested. 中文 将自然商品介绍与戏剧反转结合：正面展示、侧面细节、高位全景、紧张CTA、遮挡转场和商品落版。保持人物与商品一致、台词连续，并在接触发生前切镜。 需要提供：成年人物参考、清晰商品图、车辆参考、品牌与商品名、已确认的真实卖点、场景、口播语言，以及目标平台实际使用的购物链接或按钮。 模板包含：三张参考图的职责划分、12秒六镜规划、可替换提示词、连续原声与剪辑指导、修正错误车辆运动的局部重定时方法，以及逐镜验收清单。 样片说明：这条12秒影片由作者先前在LibTV生成并亲自剪辑，用于展示创作结构与节奏。工作流程已整理为Digen版本，但Digen使用新输入素材的生成和复用效果尚未实测。

@TANFEIZAO

Promo

0 0 1

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4274163/2026-10-13/eca3zdsf.png

# Street Pitch × Supercar Twist

## When to use | 适用场景
- The user wants a street-style talking-head × supercar near-miss product-reveal short: a continuous-voiceover ad where a car repeatedly rushes at a presenter and hard-cuts just before contact, then a blocking cut reveals the product on the road.
  用户需要街头口播与跑车反转结合的商品短片：口播连续，汽车从不同机位反复逼近人物，在接触前硬切，再通过遮挡转场揭示路面上的商品。
- The user uploads an adult character image, a product image, and a vehicle image, plus brand/product name, confirmed selling points, voiceover language, scene and CTA.
  用户提供成年人物图、商品图和车辆图，以及品牌与商品名、已确认的卖点、口播语言、场景和行动引导。
- The user wants to revise one stage (a shot, the voiceover, the cut) or re-render.
  用户需要修改某个制作环节（镜头、口播或剪辑），或重新生成。

## Reference files (read on demand, don't read all at once)
- `references/inputs-and-workflow.md` — inputs, validation, and the full step order. Read first when starting a job.
- `references/six-shot-prompts.md` — the six-shot structure with source prompts and the variable template. Read before generating shots.
- `references/quality-and-edit.md` — per-shot acceptance and editing delivery rules. Read before QC and before delivering the cut.
- `references/example-provenance.md` — the sample's source and untested boundaries. Read before claiming any capability.

## Steps (overview; details in inputs-and-workflow.md)
1. First use write_todos to list the steps, update the whole list as you progress.
2. Collect and validate the three images and the text fields. If only a portrait without the target clothes, first generate a correct first frame wearing the clothes and get user confirmation — never generate the full video with wrong clothes. If product facts are missing, ask rather than fabricate.
3. Verify the current platform channel actually supports the main route: three-image reference, a single ~12s video, multiple shots, and the same continuous original sound with lip-synced dialogue. Do not assume Digen supports it just because the original LibTV run did.
4. If supported: generate ONE video with sound using the full six-shot prompt (see six-shot-prompts.md), then locally trim / re-time problem shots while keeping the continuous original sound.
5. Only if the channel does NOT support the main route AND the user explicitly accepts the segmented plan: enable segmented generation, unified dubbing, and real stitching. Explain this is an alternative flow; lip sync still needs sync and acceptance — don't just overlay dubbed audio on the picture and claim a faithful reproduction.
6. QC per shot (see quality-and-edit.md), then deliver with an honest report.

## Media channels & canvas conventions
- New images: `generate_image` with `model='i2i'` (image-to-image, 1–3 refs). The default requirement is explicit image2 + native 4K; `i2i` is only the channel name and does NOT by itself satisfy "image2". Before each new-image step, verify the channel's actual model and native resolution satisfy BOTH; if you can't verify, reuse the user's provided correct image, or clearly state the limitation and let the user choose an alternative — don't generate before the alternative is confirmed. Never claim a physical model, 1K, upscaling, or a "4K" written in a prompt as satisfying 4K. Historical model names are only source references, not Digen models.
- Video: `generate_video` with `model='ref2v'` (reference-to-video, three reference images for consistency) — but first verify it supports the main route (single ~12s, multi-shot, continuous original sound + lip sync). Video generation does not support NSFW/adult content on any channel.
- Voiceover: `generate_tts` with `model='tts.design'` (text + voice_instruction) for the alternative segmented flow. Only call `model='tts.clone'` (text + reference_audio_url) when the user provides a sound reference AND chooses cloning.
- Stitch: `mix_clips` (clip-mixer), used only in the alternative segmented flow.
- Canvas values (fixed per stage, non-conflicting):
  - storyboard frames → `entity_type='storyboard_images'`
  -