Horror and suspense · 恐怖悬疑短片

Every shot carries its own timecode and shows one visible change on a body. The dread comes from the chain — a look, veins, a bite, the next person — and the ending seals a door without settling anything.

This Skill carries one prompt structure distilled from 29 human-verified Seedance cases on goodcase.ai. It is a sibling of seedance-prompt-library (all templates in one Skill); install this one when you only want this kind of clip. Do not invent structure from general video-generation knowledge: follow the structure below and ground the draft in one anchor case from references/cases.md.

Use when

• Outbreak and possession clips, corridor chases, ritual scenes — anything where the fear comes from a body changing on a clock.
• 中文：感染爆发、附身、走廊追逐、驱邪仪式，任何靠一具身体按秒变化来吓人的片子。

Workflow

1. Collect inputs. Ask only for what is missing: subject or product, setting, duration, aspect ratio, whether reference images / audio exist, and which language the final prompt should be in.
2. Pick one anchor case. Read references/cases.md, choose the case closest to the request, and name it in the reply. One anchor, never an average of several.
3. Fill the structure block by block. Every block below must be present in the final prompt; an empty block is where prompts go vague.
   1. Header: runtime, the genre named, and the reference lock on whoever turns first
   2. Shot 1 with a timecode: patient zero in an ordinary seat or bunk, already carrying one symptom
   3. Escalation shots, each adding exactly one visible change: veins, milky eyes, a stiff head tilt
   4. The trigger shot: the attack itself, written as slow motion with impact
   5. Transmission: the bitten person runs the same escalation on a shorter clock
   6. Crowd panic and the barricade: the door, the luggage, hands clawing through glass
   7. Closing shot: a sealed door still shuddering, or a wide exterior, with nothing resolved
4. Apply the guidance and check the pitfalls (next two sections) before returning anything.
5. Return the finished prompt as one copy-pasteable block, the anchor case link, and a three-line checklist of what to verify in the generated video.

Guidance (from the cases)

• Give every shot a timecode and exactly one change. The sleeper-train case fits 26 shots into 30 seconds, where Shot 2 is only dark veins emerging beneath the skin and Shot 3 is only Her eyes cloud milky white.
• Lock the face in the very first shot, before anything happens. The sleeper train opens with <<<image_1>>>, face and outfit matching reference; the zombie-train case uses Character A, matching reference face/outfit; the ritual case uses Keep the Word character's face and outfit consistent throughout.
• Pass the infection on and compress the second round. In the zombie-train case the first person takes twelve seconds from symptom to finished turn; the man he bites is done in seven, across Shots 13 to 16.
• Land the scare on somebody else's reaction. The sleeper train cuts to A sleeping passenger stirs as another blood drop lands on his forehead, and the rooftop case has friends fall silent, chairs scrape back.
• Refuse to resolve it. The sleeper train ends on the train running through the night with chaos in the windows; the Korean ritual case ends on One intact talisman emits faint dark smoke.

Pitfalls

• Cramming a whole transformation into one shot. Split it across four or five: eyes cloud, veins branch, body convulses, inhuman scream, head snaps forward.
• Buying the horror with gore volume. The hardest beats in the top cases are a single blood drop landing on a forehead, held as extreme close-up in slow motion.
• Letting the middle collapse into a brawl where nobody can tell who bit whom. Even the chaos shots name a subject and a target, like The infected turns and lunges at nearby passengers.
• Putting the consistency lock at the end with the style notes. By then the face has already drifted; the lock belongs in Shot 1, and the infected version still has to be the same face.

Language

Reply in the user's language. The prompt itself can stay in English when that is what the user's Seedance workflow expects; ask if unclear. The Chinese version of this structure follows.

中文：恐怖悬疑短片

每一镜都带自己的时间码，身上只发生一个看得见的变化。吓人的地方在于这些变化一环扣一环：一个眼神、浮起的血管、一口咬下去、下一个人。结尾把门关上，事情不了结。

结构:

1. 开头：时长、片种，以及第一个要变的人的参考图锁定
2. Shot 1 带时间码：零号病人坐在普通的座位或铺位上，身上已经有一个症状
3. 递进镜头，每一镜只加一个看得见的变化：血管、白眼、脖子僵硬地偏过去
4. 触发镜：袭击本身，写成慢动作加冲击
5. 传染：被咬的人走同一套递进，时间压得更短
6. 人群恐慌和封门：门、行李、从玻璃后抓过来的手
7. 收尾镜：封住的门还在震，或者一个外部大景，什么都没解决

要点:

• 每一镜给时间码，而且只放一个变化。卧铺列车那条把 26 个镜头排进 30 秒，Shot 2 只有 dark veins emerging beneath the skin，Shot 3 只有 Her eyes cloud milky white。
• 第一镜就把脸锁死，别等事情发生。卧铺列车开头写 <<<image_1>>>, face and outfit matching reference；丧尸列车那条写 Character A, matching reference face/outfit；韩屋仪式那条写 Keep the Word character's face and outfit consistent throughout。
• 感染要传下去，第二轮把时间压短。丧尸列车那条第一个人从出症状到变完用了十二秒，被他咬的人只用七秒，就是 Shot 13 到 16。
• 恐怖的落点放在别人的反应上。卧铺列车切到 A sleeping passenger stirs as another blood drop lands on his forehead，天台那条是 friends fall silent, chairs scrape back。
• 结尾不给解决。卧铺列车收在夜行的列车外景，窗里还在乱；韩屋仪式那条收在 One intact talisman emits faint dark smoke。

常见坑:

• 一镜里塞完整个变身。拆成四五镜：白眼、血管爬开、抽搐、非人的嘶吼、头猛地甩回来。
• 靠血浆量买恐怖。高热度那几条最狠的镜头是一滴血落在额头上，用极近景加慢动作拖住。
• 中段塌成一团乱打，看不清谁咬了谁。连混乱镜头也要点名主语和对象，像 The infected turns and lunges at nearby passengers。
• 把一致性锁放到结尾的风格段里。放那么后面脸早就飘了，锁定句要写在 Shot 1，而且感染之后仍然得是同一张脸。

Copy-ready lead-in / 可复制引导语

For users who would rather paste into a chat model than run this Skill, hand them this line above the structure:

I want a horror-suspense clip. [The setting is an underground car park late at night, a young woman looking for her car.] [The first to turn is the security guard; I am sending you his reference photo.] Using the prompt template below, rewrite it into one ready-to-use Seedance video prompt for me:

我要做一条恐怖悬疑短片，【场景是深夜的地下车库，一个女生在找自己的车】，【第一个出事的是那位保安，我把他的参考图发给你】。请根据下面这个提示语模板，帮我改写成一条可以直接用的 Seedance 视频提示语：

Notes

• references/cases.md is generated from data/ in LearnPrompt/awesome-seedance and refreshed daily; do not hand-edit an installed copy.
• Prompts in the reference file belong to their creators (linked on every card). Transfer the structure; do not present a close copy as original work.
• For the full library across every model, install the goodcase Skill from LearnPrompt/goodcase-lite.