GoodCase.aiAI News · Radar↗AI Map↗Course · LearnPrompt↗
ProjectsGoodCase.aiAI News Radar ↗AI Map ↗LearnPrompt ↗
GoodCase.aiGoodCase.aiAI Prompt Case Collection
GoodCase.ai
RankingsCasesSkillsFavoritesSubmitDailyJoin us
中文EN
Agent Access →

Projects

AI News · Radar ↗AI Map ↗Course · LearnPrompt ↗
GoodCase.aiGoodCase.aiAI Prompt Case Collection
RankingsCasesSkillsFavoritesSubmitDailyJoin us
中文EN
Agent Access →

Loading…

GoodCase.ai · AI Prompt Case Collection
DSH PluginsChangelogFeedback
ModelsCreatorsSeedance 2.5Seedance 2.0GPT Image 2GeminiNano Banana ProClaudeClaude Opus 5.5Grok ImagineVeoKlingQwen ImageGPT Image 1.5MiniMax H3
← Director method Skills

Director method · unofficial research version

Spielberg: See the Face Before the Thing

Let the audience learn something is wrong from a face first, then see the thing in two or three steps — with the space always readable.

Audience investment and reveal · 7 rules · avg shot 2.5–5s · v1.2.0

Keyframe from the shared test brief

Cover reuses a generated keyframe · not a model input

Install ↓

GoodCase's own synthesis from public work and interviews. Not made or endorsed by the director or their company. Does not reproduce shots, lines, characters or plots from any real film. The full method text is in Chinese.

§ 01 · RESULT

See the result first.

Same original brief for all five directors. Keyframes below are single images generated from this Skill's storyboard; they are results, never model inputs.

Demo shotNot made
No demo video for this director yet. The keyframes below are single images generated from the storyboard, not video results.
  1. S01 一镜交代门/桌/窗三个地标的相对位置;热气只作画面边缘的局部信息

    S01 · 4.5sR5 R3 R7 R6

    一镜交代门/桌/窗三个地标的相对位置;热气只作画面边缘的局部信息

    △ partly 她的背和扶门的手在左前景、桌在中景、窗在右后最亮,三层景深成立;但桌上的茶杯清楚可见,分镜要的是只看到一缕几乎读不出的热气,R2 先藏后露被提前揭了。

  2. S02 先给反应:观众从她的脸知道出事了,但还看不到是什么

    S02 · 3.5sR1 R6

    先给反应:观众从她的脸知道出事了,但还看不到是什么

    ✓ matches the shot 100mm 中近景,左侧门框虚化做前景,四分之三侧脸朝右,视线落在画面中右偏下,看的东西不在画内。R1 先给反应成立。

  3. S03 揭示线索一:木桌上那杯还在冒热气的茶

    S03 · 4.5sR4 R2 R6

    揭示线索一:木桌上那杯还在冒热气的茶

    ✓ matches the shot 桌沿在前景、茶杯居中冒热气、窗在右后把热气打亮,杯盖斜靠托碟。

  4. S04 第二次反应:视线从桌面落到地面,为线索二指路

    S04 · 3sR1 R2

    第二次反应:视线从桌面落到地面,为线索二指路

    ✓ matches the shot 同侧轴的中近景,左下角有一缕热气和一点杯沿,视线已经从杯子落到右侧地面,手没有抬。

  5. S05 揭示线索二:一道水痕从桌边通向窗口,并把人、痕迹、窗连成一条线

    S05 · 4.5sR3 R7 R5 R6

    揭示线索二:一道水痕从桌边通向窗口,并把人、痕迹、窗连成一条线

    ✓ matches the shot 低角度,左侧只见鞋、小腿和攥着钥匙的手,一道湿痕反着窗光伸向窗下墙根。机位比分镜写的 20cm 略高,不影响读法。

  6. S06 拐点:她主动走向画面里最亮的地方,用推镜代替切镜

    S06 · 5sR4 R5 R6

    拐点:她主动走向画面里最亮的地方,用推镜代替切镜

    △ partly 从她身后看出去,湿痕伸向右侧,窗是画面最亮处也是唯一出口,构图成立;但桌面是空的,茶杯不见了。

  7. S07 结局不变:窗外只有一个空院子,三条线索一条都不解释

    S07 · 5sR3 R6 R2

    结局不变:窗外只有一个空院子,三条线索一条都不解释

    ✓ matches the shot 越过她的肩,镜头垂直玻璃,脏玻璃后是一个空院子:石板地、围墙、木门、一棵秃树,没有人。

Generated with:codex exec · image_generation · gpt-6.1-sol / high · codex-cli 0.163.0 · each shot gets only the references it cites; used across the film: C01 / L01 / P01 · one run per shot, no retouching

§ 02 · METHOD

How this director decides.

Visual thesis from the test run:她每靠近一条线索一步,画面里最亮的那块光(窗)就在构图里更大一点;等她终于站到最亮的地方,那里什么都没有。

R3 · S01teaching · not a model input

Three layers in one frame

三层纵深同时说话

宽景里前景、中景、背景各放一层信息:前景是人物或她手里的东西,中景是线索,背景是出口、窗或光源。用纵深调度代替切镜。

S01:门内胸高,固定不动

R1 · S02teaching · not a model input

Her face first, then the tea

  1. S02 迈步中、重心前移 → 落脚、停住、视线锁在画面中右偏下 → 完全不动,屏住一口气
  2. S04 眼睛仍在杯上、下巴平 → 低头、视线沿地面向右移动、肩膀下沉、始终不伸手 → 盯住右侧地面,手空垂在身侧

先拍看的人,再拍被看的东西

关键事物出现之前,先给人物的反应:视线停住、动作停在半途、身体后撤。观众从她的脸知道有事,再切到她看见的东西。

R7 · S01teaching · not a model input

The clue enters as sound

1234567
  • ↔ S01→S02:画外水滴声跨切点延续 0.3 秒,画面已换成她的脸
  • ↔ S02 出点前 0.5 秒:环境声整体掏空一拍,制造揭示前的真空
  • ↔ S04→S05:低弦早于画面 0.5 秒进入,带观众过切点
  • ↔ S06→S07:窗外风声在 S06 末 0.5 秒提前进入,其余环境声同时掐空

线索先以声音或边缘出现

关键线索第一次出现时只给声音或画面边缘的一部分,后面的镜头再把它放进画面中心。

  1. R1

    先拍看的人,再拍被看的东西 · Reaction before reveal

    关键事物出现之前,先给人物的反应:视线停住、动作停在半途、身体后撤。观众从她的脸知道有事,再切到她看见的东西。

    观众是通过人物进入事件的。先看到反应,观众会主动想知道她看见了什么。

    Check:每个揭示镜头之前紧挨着一个反应镜头,分镜里写出视线方向。

  2. R2

    管住观众知道什么 · Control what the audience knows

    每个节拍写明观众此刻知道什么、不知道什么。关键信息分两到三步给:先是声音、影子或局部,再是全貌。

    信息释放的顺序就是悬念本身。一次给全,紧张感就没了。

    Check:节拍表的「观众知道」逐拍递进,完整揭示出现在局部线索之后。

  3. R3

    三层纵深同时说话 · Three planes at once

    宽景里前景、中景、背景各放一层信息:前景是人物或她手里的东西,中景是线索,背景是出口、窗或光源。用纵深调度代替切镜。

    一个画面里同时看到人和线索的关系,观众不需要剪辑解释就能读懂。

    Check:至少一个宽景写出三层分别是什么。

  4. R4

    拐点用推镜,不用切 · Push in at the turn

    人物意识到事情不对的那一刻,用一个缓慢推镜推到中近景,推的速度跟她察觉的速度一致;推完停住。

    推镜把观众拉进人物的意识里,比切到特写更连续、更像亲历。

    Check:情绪拐点所在的镜头写了 slow push-in,并写出推镜的起止景别。

  5. R5

    空间方位始终清楚 · Keep the geography readable

    开场用一个镜头交代主要地标(入口、家具、出口)的相对位置;人物移动方向在全片保持一致,不越轴。每个宽景里能看见一个出口。

    观众知道她在哪、能往哪走,才会替她担心下一步。

    Check:有建立空间的镜头;相邻镜头的运动方向一致或写明为什么越轴。

  6. R6

    光从看得见的地方来 · Motivated light

    每一镜的主光来自画面里能指认的光源(窗、门缝、灯)。最亮的光源同时是人物视线的目标或要去的方向。

    光把观众的眼睛引向下一条信息,而不是装饰。

    Check:每一镜写出光源名字和方向。

  7. R7

    线索先以声音或边缘出现 · Clues arrive as sound or edge first

    关键线索第一次出现时只给声音或画面边缘的一部分,后面的镜头再把它放进画面中心。

    声音比画面更早抵达,观众会跟着人物一起转头。

    Check:声音方案里至少一条线索先于画面出现,并写明在哪一镜补上画面。

Use it for

  • · 发现、揭晓、悬念:观众需要一步步知道发生了什么
  • · 普通人面对超出日常的事件,需要观众代入人物
  • · 需要清楚空间关系的段落:地标、通道、出口在哪里

Not for

  • · 刻意让观众迷失方位的心理段落
  • · 以情绪氛围为主、没有信息要释放的抒情片段
  • · 纯产品展示或无人物的空镜

This is not it

  • · 只加逆光、丁达尔光和仰拍不算这套方法。必须写清观众先知道什么、何时看见关键事物。
  • · 揭示不是一次性大全景。没有反应镜头铺垫的揭示,等于把答案直接递给观众。
  • · 不要让配乐替画面说话,音乐只在人物意识到之后进来。
  • · 不要借用真实电影里的道具、怪物、台词或标志性镜头。

§ 03 · RECIPE

Follow it shot by shot.

Everything below is the Skill's actual output on the shared brief, unedited except for layout.

Input used in the test

用斯皮尔伯格的拍法,把下面这个故事做成一份 AI 视频的完整拍摄方案:30 秒,16:9。

故事:一位四十岁左右的女人回到空置多年的老屋。她推门进屋,发现原本空着的木桌上多了一杯还在冒热气的茶。她看见地上有一道水痕从桌边通向窗口。她走到窗前,窗外只有一个空院子。

要求:人物、三条线索的顺序和结局都不要改,没有对白。我已经有三张参考图:C01 是她的定妆照,L01 是老屋室内,P01 是桌上那杯茶。生成工具还没定,按只支持首帧图生视频来写。

Model inputs (generation_input)

  • C01C01 · identity · Ready
    她的脸、短发、深灰旧呢子大衣、手里的一串旧钥匙
    → sent to the model
  • L01L01 · location · Ready
    老屋室内:门、木桌、窗的位置关系,墙面与地面的材质,窗外院子
    → sent to the model
  • P01P01 · prop · Ready
    桌上那杯茶:白瓷带盖茶杯、热气、桌面木纹
    → sent to the model

Covers, teaching cards and the keyframes above are never passed to the model.

  1. B1 她推门进入空置多年的老屋Audience knows:知道这是久无人住的屋子、门在左桌在中窗在右;不知道屋里有人来过
  2. B2 木桌上多了一杯还在冒热气的茶Audience knows:知道有人几分钟前还在这儿;不知道是谁、现在在哪
  3. B3 地上一道水痕从桌边通向窗口Audience knows:知道有东西从桌边移动到了窗口,方向明确;不知道那东西还在不在屋里
  4. B4 她走到窗前Audience knows:知道水痕的终点就是这扇窗;不知道窗外有什么
  5. B5 窗外只有一个空院子Audience knows:知道窗外没有人,三条线索一条都没被解释;不知道的比进门时更多
ShotTaskSize / cameraAction (start → end)SoundCutRulesAssets
S01
4.5s
一镜交代门/桌/窗三个地标的相对位置;热气只作画面边缘的局部信息广景 24mm
门内胸高,固定不动
手扶门站在门槛 → 向右跨进一步、手离开门、门在身后微微继续敞开 → 站在屋内,重心在前脚,面朝屋里的窗门轴吱呀、鞋底踩灰、空屋底噪;最后 0.4 秒一颗画外水滴声在她停住的瞬间切,把『她看见了什么』留给下一镜R5 R3 R7 R6C01 L01 P01 K01
S02
3.5s
先给反应:观众从她的脸知道出事了,但还看不到是什么中近景 100mm
她的眼高,固定,与 S01 同侧轴
迈步中、重心前移 → 落脚、停住、视线锁在画面中右偏下 → 完全不动,屏住一口气脚步声停;底噪在切点前掏空一拍反应已经给足,观众正在追问,切到揭示R1 R6C01 L01 K02
S03
4.5s
揭示线索一:木桌上那杯还在冒热气的茶中景推至中近景 50mm
桌面高度,沿轴缓慢推镜,减速停住
桌面静止、热气上升 → 镜头缓慢推近,热气持续卷动 → 杯子占据画面中心,推镜完全停住杯中热水极细的 tick、空屋底噪,无音乐推镜停住即切,不多给一帧解释R4 R2 R6P01 L01 K03
S04
3s
第二次反应:视线从桌面落到地面,为线索二指路中近景 100mm
她的眼高,固定,不越轴
眼睛仍在杯上、下巴平 → 低头、视线沿地面向右移动、肩膀下沉、始终不伸手 → 盯住右侧地面,手空垂在身侧衣料、一次吞咽;单音低弦从此刻进入她的视线已经指向地面,必须让观众跟着看过去R1 R2C01 L01 K04
S05
4.5s
揭示线索二:一道水痕从桌边通向窗口,并把人、痕迹、窗连成一条线广景低角度 21mm
离地 20cm,缓慢右摇,减速停住
她的鞋在画左、水痕横贯画面 → 镜头沿水痕向右摇,一颗水从桌沿落入痕中 → 停在窗下墙根,水痕终点在画面里一颗水滴落地(S01 的画外声在此补上画面)、玻璃外闷响的风声、低弦延续摇镜已经把人和窗连成一条线,切到她的行动R3 R7 R5 R6C01 L01 K05
S06
5s
拐点:她主动走向画面里最亮的地方,用推镜代替切镜中景推至中近景 35mm
右肩后方、肩高,随她行走推镜,与她同速停住
背对镜头将要起步 → 沿水痕从左走到右、到玻璃前停住、肩膀一次起伏 → 站定不动,头没有转,镜头同时停住脚步、衣料;最后 0.5 秒窗外风声提前进入,其余环境声掐空推镜停住即切,把观众交给她正看着的东西R4 R5 R6C01 L01 K06
S07
5s
结局不变:窗外只有一个空院子,三条线索一条都不解释广景越肩 50mm
她眼高,垂直于玻璃,完全固定
空院子静止 → 全程无一物进出画面,只有她的肩部剪影随呼吸微动、玻璃上浮尘轻颤 → 与第一帧完全一致:院子还是空的风声、远处一声鸟;音乐在第一帧掐掉;最后 1 秒只剩房间底噪片尾,不收不解释,停在『没有答案』上R3 R6 R2C01 L01 K07
S01 · keyframe + motion prompts

Keyframe prompt

24mm-equivalent wide shot, camera at chest height just inside the room on the door-to-window axis. Three planes: FOREGROUND — the woman's back and shoulder at frame left, one hand still resting on the edge of the opened entrance door, a wedge of pale daylight and floating dust coming in behind her; MIDGROUND — the bare wooden table at frame centre, reading as almost empty, with only one thin thread of vapour rising above it, backlit and barely legible, not an obvious cloud; BACKGROUND — the window at frame right, the brightest rectangle in the frame, the room's only exit, the courtyard beyond it blown out and unreadable. Her eyeline points forward into the room toward the window light. LOCKED IDENTITY: a woman around forty years old; her face, hairstyle and clothing exactly as in reference image C01 — do not redesign her. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust on every surface; do not add furniture or modern objects. LOCKED PROP: the tea cup on the wooden table exactly as in reference image P01 — same cup, same tea colour, same position on the table, but here read only as a small dark shape, not the subject. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right; she only ever moves left to right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, warm door light against cool window light, shadow detail retained. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no plastic buckets, no modern lamps or LED fixtures, no second person, no visible steam cloud, no fresh paint.

Motion prompt (image-to-video)

Locked camera, no push, no shake. She starts standing on the threshold with one hand on the opened door, takes a single step into the room from left to right and her hand leaves the door; the door swings a few centimetres further open behind her and stops. Dust drifts slowly through the door-light wedge. The thin thread of vapour above the table keeps rising but stays thin and stays at the edge of legibility — do not turn it into a visible cloud. Ends with her standing just inside the room, weight on the front foot, still facing into the room toward the window. Maintain LOCKED IDENTITY (a woman around forty, face, hairstyle and clothing exactly as reference image C01), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no lip movement, no text, no watermark. Avoid: no second person entering frame, no door slamming, no curtain blowing, no camera orbit, no handheld shake.
S02 · keyframe + motion prompts

Keyframe prompt

100mm-equivalent medium close-up, camera at her eye height, same side of the axis as the previous wide shot. Three planes: FOREGROUND — the out-of-focus edge of the door frame at frame left; MIDGROUND — her face and shoulders, turned three-quarters toward frame right, sharp; BACKGROUND — the soft out-of-focus bright rectangle of the window at frame right. Her eyeline is locked on something at frame centre-right, slightly below her own eye level; what she is looking at is not in the frame. Motivated backlight from the opened door behind frame left catches the left edge of her jaw; the daylight from the window at frame right fills the front of her face. LOCKED IDENTITY: a woman around forty years old; her face, hairstyle and clothing exactly as in reference image C01 — do not redesign her. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right; she only ever moves left to right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, shadow detail retained. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no tears, no exaggerated fear expression, no mouth open, no second person, no modern jewellery.

Motion prompt (image-to-video)

Locked camera, no push, no shake. She starts mid-step with her weight still shifting forward; her foot comes down and she stops walking, her head settles, and her eyeline locks on something at frame centre-right just below her eye level. One short held breath, no blink on the last beat. Ends completely motionless, lips closed, eyes fixed on that off-screen point. Do not reveal what she is looking at; keep it outside the frame for the whole shot. Maintain LOCKED IDENTITY (a woman around forty, face, hairstyle and clothing exactly as reference image C01), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no lip movement, no text, no watermark. Avoid: no gasping with open mouth, no hand rising to her face, no head shake, no camera push, no rack focus to the background.
S03 · keyframe + motion prompts

Keyframe prompt

50mm-equivalent medium shot of the wooden table, camera at seated eye height, level, on the door-to-window axis. She is not in this frame. Three planes: FOREGROUND — the near edge of the wooden table, its dust layer sharp and unbroken; MIDGROUND — the tea cup at frame centre, steam rising and curling off the surface of the tea, the dust around the cup visibly disturbed in a small clean patch; BACKGROUND — the window at frame right throwing the steam into bright relief, the courtyard beyond it blown out and unreadable. The brightest area of the frame is the window, directly behind the steam. LOCKED PROP: the tea cup on the wooden table exactly as in reference image P01 — same cup, same tea colour, same position on the table. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, shadow detail retained. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no hand in frame, no saucer or spoon that is not in reference image P01, no second cup, no teapot, no modern mug, no printed packaging.

Motion prompt (image-to-video)

Slow push-in straight along the axis, from medium shot to medium close-up on the tea cup, moving at a crawl and decelerating to a complete stop before the shot ends. Nothing in the room moves except the steam, which keeps rising and curling continuously from the first frame to the last. Ends with the cup filling the centre of the frame, steam still clearly visible against the window light, camera stopped and static. Maintain LOCKED PROP (the tea cup exactly as reference image P01, same cup, same tea colour, same position), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no text, no watermark. Avoid: no hand entering frame, nobody touching the cup, no face reflected in the tea, no cup sliding or tipping, no steam fading out, no zoom snap, no handheld shake.
S04 · keyframe + motion prompts

Keyframe prompt

100mm-equivalent medium close-up of the woman, camera at her eye height, same side of the axis as the earlier reaction shot, no axis crossing. She now stands closer to the table than before. Three planes: FOREGROUND — a thin curl of steam crossing the lower-left corner of the frame, slightly out of focus; MIDGROUND — her face and shoulders, sharp, turned three-quarters toward frame right; BACKGROUND — the out-of-focus bright window rectangle at frame right. Her eyeline has dropped: she is looking down and toward frame right, at the floor. Motivated light: daylight from the window at frame right fills her face from the front-right; the weaker door light from behind frame left edges her left shoulder. LOCKED IDENTITY: a woman around forty years old; her face, hairstyle and clothing exactly as in reference image C01 — do not redesign her. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right; she only ever moves left to right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, shadow detail retained. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no tears, no hand raised to her mouth, no second person, no phone, no modern eyewear.

Motion prompt (image-to-video)

Locked camera, no push, no shake. She starts with her eyes still on the cup and her chin level; her head tilts down, her eyeline travels down and across to frame right along the floor, her shoulders drop a few centimetres as her breathing changes, and she does not reach toward the cup at any point. Ends with her looking down at the floor at frame right, head still, hands empty at her sides. The curl of steam in the lower-left foreground keeps moving gently throughout. Maintain LOCKED IDENTITY (a woman around forty, face, hairstyle and clothing exactly as reference image C01), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no lip movement, no text, no watermark. Avoid: no hand reaching for the cup, no mouth opening to speak, no rapid blinking, no turning to look at camera, no camera push.
S05 · keyframe + motion prompts

Keyframe prompt

21mm-equivalent wide low-angle shot, camera 20cm above the floorboards, lens level with the floor, on the table-to-window axis. Three planes: FOREGROUND — her shoes and lower legs at frame left, sharp, standing still in the dust; MIDGROUND — one continuous wet trail on the dusty floorboards running from the foot of the wooden table at frame centre toward frame right, dark against the pale dust and breaking the window light into glinting segments; BACKGROUND — the window at frame right, the brightest area of the frame, with the wet trail ending at the wall directly beneath it, and the blown-out courtyard beyond the glass. LOCKED IDENTITY: a woman around forty years old; her clothing exactly as in reference image C01 — do not redesign it. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust except where the wet trail breaks it; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, shadow detail retained. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no footprints in the dust, no human face or figure reflected in the water, no puddle shaped like a body, no plastic buckets, no modern skirting boards, no floor drain.

Motion prompt (image-to-video)

Slow pan to the right along the floor at walking pace, starting on her shoes at frame left, travelling along the wet trail, and decelerating to a complete stop with the far end of the trail and the base of the window wall held in frame. One single drop of water falls from the edge of the table into the trail early in the shot. The wet trail itself stays exactly as it is: it does not spread, grow, flow or move. Her shoes stay planted and do not leave the frame on their own. Ends with the camera stopped on the end of the trail beneath the window. Maintain LOCKED IDENTITY (her clothing exactly as reference image C01), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no text, no watermark. Avoid: no feet walking into frame, no water flowing like a stream, no new puddles forming, no reflection of a person in the water, no whip pan, no handheld shake.
S06 · keyframe + motion prompts

Keyframe prompt

35mm-equivalent medium shot from behind her, camera just off her right shoulder at her shoulder height, aimed down the axis toward the window. Three planes: FOREGROUND — her back, shoulder and the back of her head, left of frame centre, dark against the light; MIDGROUND — the wet trail on the floorboards leading away from her toward frame right; BACKGROUND — the window at frame right-of-centre, the brightest rectangle in the frame and the room's only exit, the courtyard beyond it still blown out and unreadable. Her eyeline, read from her head position, points straight at the window. LOCKED IDENTITY: a woman around forty years old; her hair and clothing exactly as in reference image C01 — do not redesign them. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same layout, same wall and floor materials, undisturbed dust except where the wet trail breaks it; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right; she only ever moves left to right. LOCKED LIGHT: daylight through the window at frame right is the brightest source; a weaker wedge of light comes from the opened door behind frame left. Natural colours, cool window light against the warm room, shadow detail retained in her back. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no second person, no figure visible through the glass, no curtains blocking the window, no modern window frame, no air conditioner.

Motion prompt (image-to-video)

Slow push-in from behind her, from medium shot to medium close-up on the back of her head and shoulders, the camera moving at exactly her walking speed and stopping when she stops. She starts with her back to camera about to take a step, walks from left to right along the wet trail to the window, and halts at the glass; her shoulders rise once on an inhale and settle. Ends with her standing still at the window, back to camera, head not turned, camera stopped. Keep the courtyard beyond the glass over-exposed and unreadable for this whole shot — do not reveal what is outside. Maintain LOCKED IDENTITY (a woman around forty, hair and clothing exactly as reference image C01), LOCKED LOCATION (the room of reference image L01, undisturbed dust), LOCKED LIGHT (brightest daylight from the window at frame right, weaker door light from behind frame left). No dialogue, no lip movement, no text, no watermark. Avoid: no camera orbiting around her, no handheld shake, no turning to look at camera, no figure appearing outside the window, no hand raised to the glass, no reflection of a face in the glass.
S07 · keyframe + motion prompts

Keyframe prompt

50mm-equivalent wide shot through the window from just behind her, camera at her eye height, level, lens axis perpendicular to the glass. Three planes: FOREGROUND — the dark silhouetted edge of her shoulder and the side of her head at frame left, plus the dusty window frame at the left and bottom of frame, both out of focus; MIDGROUND — the old window glass itself, streaked and dust-filmed, sharp; BACKGROUND — an empty courtyard in flat daylight, bare swept ground, a low enclosing wall and a shut gate, completely empty, nobody there, nothing moving. The courtyard is the brightest part of the frame and the target of her eyeline. LOCKED IDENTITY: a woman around forty years old; her hair and clothing exactly as in reference image C01 — do not redesign them, here read only as a dark silhouette at the frame edge. LOCKED LOCATION: the interior of a long-vacant old house, exactly the room in reference image L01 — same window, same wall and floor materials; do not add furniture or modern objects. LOCKED GEOGRAPHY: entrance door at frame left, wooden table at frame centre, window at frame right; this shot looks out through that window. LOCKED LIGHT: daylight in the courtyard is the brightest source in the frame; the room side stays dark with shadow detail retained. Natural colours, cool exterior daylight. No text, no watermark, no dutch angle, no unmotivated rim light. Avoid: no person in the courtyard, no animal, no bicycle, no car, no human shadow on the ground, no footprints in the courtyard, no laundry, no modern satellite dish, no open gate.

Motion prompt (image-to-video)

Locked camera, no push, no pan, no zoom. The courtyard stays completely empty and completely still from the first frame to the last; nothing enters the frame and nothing leaves it. The only movement in the shot is her shoulder silhouette at the left edge settling very slightly as she breathes, and a faint shimmer of dust on the glass. Ends exactly as it began: the same empty courtyard, nobody there, camera static. Maintain LOCKED IDENTITY (a woman around forty, hair and clothing exactly as reference image C01, a dark silhouette at the frame edge), LOCKED LOCATION (the window and room of reference image L01), LOCKED LIGHT (bright daylight outside, dark room side with shadow detail). No dialogue, no lip movement, no text, no watermark. Avoid: no figure appearing or walking past, no gate opening, no bird flying through frame, no face reflected in the glass, no curtain movement, no reveal of any kind, no camera move.

Sound and edit

  • · 房间底噪:空屋低频嗡鸣 + 极细高频静默感,S01 第一帧到片尾连续
  • · 环境声:院中风声、远处一声鸟;S01–S05 隔玻璃低通闷掉,S06 末打开,S07 最清晰
  • · 拟音:门轴吱呀(S01)、鞋底踩灰与木地板碎响(S01 S02 S04 S06)、衣料摩擦(S02 S04 S06)
  • · 点状音效:一颗水滴落地(S01 末画外先入,S05 补上画面);杯中热水极细 tick(S03)
  • · 音乐:单音持续低弦,S04 她低头时进入,S06 末渐强,S07 第一帧掐掉只留风
  • ↔ S01→S02:画外水滴声跨切点延续 0.3 秒,画面已换成她的脸
  • ↔ S02 出点前 0.5 秒:环境声整体掏空一拍,制造揭示前的真空
  • ↔ S04→S05:低弦早于画面 0.5 秒进入,带观众过切点
  • ↔ S06→S07:窗外风声在 S06 末 0.5 秒提前进入,其余环境声同时掐空
  • ↔ S07 末 1 秒:所有层收掉,只剩房间底噪到黑

硬切 6 个,无转场特效。反应/揭示三对必须相邻且不可换序:S02↔S03、S04↔S05、S06↔S07。每镜入点取动作开始后 0.2 秒,出点取结束状态稳定后 0.3 秒,S07 留满 5 秒。全片两处运动拐点:S03 推、S06 推,其余靠切与摇。可靠剪辑修的问题:S01 热气过显→缩到 3.5 秒并推后入点;S05 摇速不稳→10% 变速平滑;S07 末出现异动→出点提前;推镜不够慢→后期加 2% 缓动缩放。

Tool adapter and fallbacks

unknown · 首帧图生视频

  • · 不支持首尾帧:每镜只用首帧 K01–K07,结束状态全部写在运动提示词里
  • · 视频模型不支持多图参考:先用生图模型吃 C01/L01/P01 产出 K01–K07,视频模型只吃单张首帧
  • · 不支持原生音频:全部声音交给后期,按 sound 分层制作
  • · 不支持相机指令:S03 S06 生成静态镜头后期做 3–5% 匀速缓慢放大,S05 生成静态广景后期做横向位移;为此 7 镜统一出 2K 以上留裁切余量
  • · 单次时长仅 4 秒:S06 S07 各减 0.5 秒,总长 29 秒,不要拉慢 S07 凑时长
  • · 未使用视频运动参考,也未使用原生音频,不声称已使用
  • · 人物连贯性风险集中在 S02 与 S04 两个正面反应镜:优先多抽版本,或用同一张 K02 轻微裁切派生 K04
  • · 同一镜失败 3 次改设计:S06 拆为 S06a 起步 + S06b 到窗前停住(新首帧);S05 改为只拍结果的固定广景;不堆提示词

§ 04 · INSTALL

director-spielberg · v1.2.0

Install into your agent.

curl -LO https://goodcase.ai/director-skills/packages/director-spielberg.skill
unzip -o director-spielberg.skill
npx skills add ./director-spielberg --global --copy --yes --agent claude-code

The one-line GitHub install is not available yet: this package has not been mirrored to the public repo. The download path above is the one we verified.

Download .skill ↓

24538 B · zip sha256 80333ffe9de3ddb2cad0ad1453aef1731c856f8ac48f16d2c6a977e0295c18be

What has actually been verified

  • Package structurePassed2026-10-25 · 4 checks
  • Real installPassed2026-10-25 · skills 1.5.22 · claude-code · 7 files byte-identical
  • Triggered on the shared briefPassed2026-10-25 · claude-opus-5[1m] · Claude Code 2.1.268
  • Demo shotNot made

Installing the Skill does not give your agent a video API. Without a generation tool it delivers copyable prompts and an asset list.

§ 05 · EVIDENCE

Sources, runs, and what failed.

  • DGA Quarterly:斯皮尔伯格访谈(2006 冬) ↗研究起点。本模块没有引用其中原话,规则是 GoodCase 的归纳。
  • DirectorSKILL · 01_spielberg.md(MIT,固定 commit c65ae0d) ↗MIT · Copyright (c) 2026 wangzhang-wu · 借鉴了三层纵深、反应先于揭示的结构;文字为 GoodCase 重写,并补上观众信息递进的验收。

Trigger run · shared brief

Passed

2026-10-25T06:11:02.124Z · 494s · claude-opus-5[1m] · 2.1.268 (Claude Code)

Skill invoked:✓ · references:references/workflow.md, references/director-spielberg.md, references/asset-contract.md, references/output-contract.md, references/acceptance.md, references/sources.md

临时项目 .claude/skills + --setting-sources project + --strict-mcp-config

  • ✓ 输出末尾有可解析的分镜 JSON
  • ✓ director 字段与模块一致(写的是 spielberg)
  • ✓ skillVersion 与包一致(写的是 1.2.0)
  • ✓ 封面、教学图、示范结果没有被列为模型输入
  • ✓ 素材 status 只用 missing / ready / uploaded
  • ✓ 片长与要求相符(±20%)(合计 30 秒)
  • ✓ 每一镜都引用了本模块存在的规则编号
  • ✓ 镜头引用的素材 ID 都在素材清单里
  • ✓ 用户给的 C01 / L01 / P01 都进了素材清单
  • ✓ 每一镜的关键帧与运动提示词都是完整正文(不是引用)
  • ✓ 末尾有逐条验收(17 条)

Fallback run · no assets, no tools

Passed

2026-10-25T06:49:08.756Z · 447s · claude-opus-5[1m] · 2.1.268 (Claude Code)

Skill invoked:✓ · references:references/workflow.md, references/director-spielberg.md, references/asset-contract.md, references/output-contract.md, references/acceptance.md, references/sources.md

临时项目 .claude/skills + --setting-sources project + --strict-mcp-config

  • ✓ 输出末尾有可解析的分镜 JSON
  • ✓ director 字段与模块一致(写的是 spielberg)
  • ✓ skillVersion 与包一致(写的是 1.2.0)
  • ✓ 封面、教学图、示范结果没有被列为模型输入
  • ✓ 素材 status 只用 missing / ready / uploaded
  • ✓ 无素材时所有 generation_input 都标 missing(13 件)
  • ✓ 缺失素材都给了制作提示词
  • ✓ 写出了工具降级方式(7 条)
  • ✓ 分镜仍然完整给出,提示词是完整正文(8 镜)

Specialty brief · text-only storyboard

Passed

2026-10-25T06:24:54.088Z · 547s · claude-opus-5[1m] · 2.1.268 (Claude Code)

Skill invoked:✓ · references:references/workflow.md, references/director-spielberg.md, references/asset-contract.md, references/output-contract.md, references/acceptance.md, references/sources.md

临时项目 .claude/skills + --setting-sources project + --strict-mcp-config

  • ✓ 输出末尾有可解析的分镜 JSON
  • ✓ director 字段与模块一致(写的是 spielberg)
  • ✓ skillVersion 与包一致(写的是 1.2.0)
  • ✓ 封面、教学图、示范结果没有被列为模型输入
  • ✓ 素材 status 只用 missing / ready / uploaded
  • ✓ 片长与要求相符(±20%)(合计 30 秒)
  • ✓ 每一镜都引用了本模块存在的规则编号
  • ✓ 镜头引用的素材 ID 都在素材清单里
  • ✓ 每一镜的关键帧与运动提示词都是完整正文(不是引用)
  • ✓ 末尾有逐条验收(11 条)

Blind review across all five

Passed

Pre-registered:每个配对必须引用样本里的调度、信息释放顺序、镜长或切点、声音安排作为理由;只引用题材或色彩词的配对,即使猜对也记为依据不合格;pass line:两位审阅者各自至少 4/5 配对正确

  • codex exec · 只读 · 纯文本 · gpt-6.1-sol:5/5This director's sample:✓ 开场先交代门、桌、窗的相对位置,热气仅在边缘露出,水滴先以画外声出现;随后严格按人物停步屏息再揭示茶杯、低头反应再揭示水痕、窗前停住再揭示院子的顺序释放信息;人物持续从左向右移动且不越轴,行动拐点用随行缓推完成,停住后才切向她所看的窗外。
  • Claude 子代理 · 全新上下文 · claude-opus-5-5(子代理继承本会话模型):5/5This director's sample:✓ beats逐条写观众知道什么与不知道什么;S01一镜交代门在左桌在中窗在右三个地标且不越轴;反应镜先于揭示镜,三对相邻不可换序(S02→S03、S04→S05、S06→S07);水滴先在S01以画外声出现、到S05才进入画面;拐点S06用随行推镜代替切镜并与她同速停住;镜长3–5秒。

只有一个室内题材、每位只有一次生成,结论只能说明这一次输出可区分,不能推出方法稳定可区分;维伦纽瓦的大尺度能力没有被这个题材检验到

Failures and fixes

  1. 2026-10-25
    生图链路
    现有复测管线的 codex-runner 注释写明参考图参数「v1 刻意不传、从未实跑」,不确定 codex 内置 image_generation 会不会真的用 --image 传入的图锁定身份。正式生成前先做一次 spike:传入 C01 与 L01 生成一镜,人工核对房间结构、墙面剥落、窗与树、人物大衣发型与钥匙均与参考一致,确认参考图进入了生成;spike 产物不发布,记录在 image-log.json。skills/directors/runs/image-log.json(group=spike)
  2. 2026-10-25
    触发复测隔离
    按 deep-reasoner 的审查意见用隔离 HOME 跑 claude -p,结果读不到登录态,直接 api_error,复测无法进行。改为临时项目目录 + --setting-sources project + --strict-mcp-config。init 事件核对:只加载内置 Skill 与被测的那一个,本机 ~/.claude/skills 下的其他 Skill 与 MCP 都没加载。scripts/skills/run-director-trigger.mjs 头注释;各 run.json 的 loadedSkills
  3. 2026-10-25
    同题复测污染
    第一版公共流程与导演规则里的举例和同题剧本重叠:导演意图示例写的是「她离那杯茶越近」,素材合约示例写的是「深灰旧呢子大衣」(正是 C01),赖特 R7 用「圆形杯口与圆形窗」举匹配剪辑,斯皮尔伯格 R5 写「门、桌、窗」。带着这些示例去跑同题复测,等于提前给了答案。把示例全部换成与剧本无关的题材(男孩与铁门、井盖与钟面、入口家具出口),内容 sha 随之变化;已开始的王家卫第一次触发复测中止作废、不留输出,用修正后的包重跑。单测加了一条:包内不得出现同题剧本原句。scripts/skills/lib/director-skills.test.mjs「同题剧本不能进包」
  4. 2026-10-25
    输出合约 v1.0.0
    v1.0.0 复测里斯皮尔伯格与维伦纽瓦的输出把逐镜 keyframePrompt 写成「见 K01 prompt」,完整文字放进了 assets[].prompt。检查器只查非空,判了通过;生图脚本把这句引用原样发给图片模型,出来的是 C01 灰底定妆照的翻版、一张海边、几张通用房间,和分镜无关。王家卫那次输出是完整正文,没踩到。这是合约缺陷:公共流程第 5 步、输出合约、验收清单都写明逐镜提示词必须是能直接粘贴的完整正文,不许引用。检查器改用 isFullPrompt(至少 60 字、不以 见/参见/同上/see/refer 开头),生图脚本遇到同样情况直接拒绝。公共流程与五个模块升到 1.1.0,五位全部按新包重跑;v1.0.0 的输出、生图日志与图片整份留档,不上页面。skills/directors/runs/_history/v1.0.0/(三位导演的 run 与 24 条生图记录);scripts/skills/lib/director-skills.test.mjs「见 K01 prompt」测试
  5. 2026-10-25
    输出合约 v1.1.0
    v1.1.0 复测里四位导演(斯皮尔伯格、维伦纽瓦、王家卫、韦斯·安德森)都替用户手里的参考图编了外观。用户只说有 C01 定妆照、L01 老屋、P01 茶杯,Skill 看不到图,却在锁定句里写了「肩长深棕发」「头发后梳」「低发髻、石板灰大衣」「鼠尾草绿墙」「赭红釉茶杯」,并标成 reference C01 / L01 / P01。真实使用时这些文字会和用户的图打架。生图脚本另外拼接了三件素材真实的锁定句,所以图片大多跟着参考图走,韦斯那份「桌上唯一一件不对的东西是红杯子」也就落空了。合约只要求锁定句每镜原样重复,没说内容从哪来,是合约的缺口。公共流程第 5 步与素材合约加了一条:用户给了图但没描述时,锁定句只写「与参考图 C01 完全一致」加上用户文字里能确认的事实,不许替参考图编发型、服装、颜色、材质;导演方法需要改造型或颜色的,写进假设作为建议。验收清单加第 4 条。公共流程与五个模块升到 1.2.0,五位重跑;v1.1.0 的四份输出、三份关键帧核对、32 条生图记录留档,不上页面。这一条没法用正则自动判,靠逐份人工核对并写进每位的关键帧核对记录。skills/directors/runs/_history/v1.1.0/(四份 output.md 的锁定句;spielberg / villeneuve / wong-kar-wai 的 keyframe-review.json)
  6. 2026-10-25
    关键帧连戏(未修,记为已知问题)
    v1.2.0 五位的关键帧里,后半段的宽镜头普遍出现桌上的茶杯消失:斯皮尔伯格 S06、维伦纽瓦 S05、王家卫 S10、韦斯·安德森 S06 / S07、赖特 S24。原因有两层:Skill 只在前几镜把 P01 列进 assetRefs,后面桌子仍在画面里却不再引用茶杯;同时复测脚本给 L01 写的锁定句是「a bare square wooden table」,房间参考图里本来就是一张空桌。另一个反复出现的偏差是维伦纽瓦的小比例人物:提示词写 8% 到 12% 画高,出来是 35% 到 65%。本轮没有修,照实记进每位的关键帧核对。下一版合约候选:道具在后续镜头里仍在画面中,就继续列进 assetRefs 并带上它的锁定句;复测脚本的 L01 锁定句改成不预设桌面状态。小比例人物可以考虑先生成空镜再把人物合成进去。改合约会让五个包的内容 sha 再变一次,需要再跑一轮,留给下一轮计划。skills/directors/runs/<slug>/keyframe-review.json 里标 partial 的各镜;skills/directors/source/common-brief.mjs 的 L01 lockSentence

Keyframe generation log

Done

7/7 shots, 7 attempts, 0 failed.

Manual review(2026-10-25,Claude(本会话执行者)人工逐张对照分镜,未请第三方复核):7 镜:pass 5、partial 2、fail 0。共尝试 7 次,没有重生成,全部是单次示范,不代表稳定复现。先反应后揭示、三层景深这类靠摆位的要求做得稳;偏差一处是「只露一缕热气」被画成整只杯子,一处是后段宽镜里茶杯消失。出现人物的 6 张短发、深灰呢子大衣、深绿毛衣、铜钥匙与 C01 一致;7 张房间与 L01 一致;S01 / S03 茶杯杯盖斜靠托碟,与 P01 一致。假设里直接写了「C01/L01/P01 外观不可见,锁定句只写与参考图一致」,门、桌、窗的左右关系写成可整体镜像的假设。

One success per shot is a single demonstration, not proof of stable reproduction.

Package, installation and generation are verified separately. Anything not run is marked as not run.