AI News

A Yi AI Notes shares the seven-layer prompt controlled-variable method to reduce rerolling and rewriting in AI video

A Yi AI NotesUpdated Nov 16 21:00

English text is being prepared.

AI summary

A Yi AI Notes recommends giving Miller's seven-layer prompt to Claude and, when working against the material, changing only one layer at a time instead of rewriting the whole thing after one bad reroll. The article cites practical data from Miller's "Wake Up, I'm Hungry": 43 shots were drawn only 52 times, with only 9 extra draws. The approach is to first use a lock card to fix the art style, lighting, and character naming as a baseline, then break down the shot with the seven-layer architecture, and when a draw fails, change only the layer that has the problem.

Why read this

AI score 55

The original text uses data from 43 shots drawn only 52 times to show that replacing full rewrites with layer-by-layer constraint changes can reduce wasted rerolls.

Topics

Tags

Quoted post

Miller

@TuyetloanP81840

https://t.co/z3DU1tUdaa

View original

https://t.co/z3DU1tUdaa

View original post

Body

看完这篇我彻底悟了,甚至忍不住想说个暴论,90%的人用AI视频抽卡抽到崩溃,根本原因就两个字,手贱, 铁汁们,别再一抽废就换模型了,把这篇的七层提示词丢给 Claude,让它对着你的素材只改一层,做 AI 短剧的建议直接收藏。 大家是不是也经历过每次一抽废就把整段提示词删了重写,改完之后抽好了不知道是哪句起作用,抽废了也不知道该退回哪一步, 讲真这和蒙着眼睛调参数没任何区别哈哈,

想起我之前做AIGC工作流也老犯这个毛病,一跑偏就全盘推翻重写,后来逼自己每次只改一个约束条件,整个效率才真正跑顺,

咱们看下Miller的《醒醒,我饿了》实操数据:43段镜头只抽了52次,多抽的只有9次,他干的事说白了就是把玄学抽卡改造成了软件工程里的控制变量法,

先用锁定卡把画风光线和角色称呼彻底写死, 等于给系统定死基准线,再用七层架构拆解镜头, 每次抽废了只动出问题的那一层,其他一个字不改,

我一直觉得不管是做AI短剧,还是平时写代码写文案,只要你还在靠反复重写整段提示词碰运气,

你就永远只是个给模型打工的抽卡员,真正能把AI变成生产力的人,都在干同一件事,

把模型可以瞎猜的自由度一点点关进笼子里,把概率游戏变成确定性流水线。

Source

A Yi AI Notes