WEBVTT

00:00:00.000 --> 00:00:29.440
上個月,GigHub上出現了一個中文項目,12天拿了4200個STAR,緊接著登頂GigHub趨勢榜全球第一。我第一反應是不信。提示詞合集能有什麼搞頭?GigHub上這類倉庫沒有1000也有800。結果它現在18000多STAR,還在漲。這期視頻我們就來拆這個項目。

00:00:30.000 --> 00:00:37.900
它叫Awesome GPT Image 2,作者网名仓和,自我定位是Prompt as Code,提示词及代码。

00:00:38.680 --> 00:00:50.600
官方描述写着,GPT Image 2工业级提示词引擎与模板库,530多个案例逆向工程,20多套工业级模板,还提炼出了Agent Skill。

00:00:50.600 --> 00:00:52.280
他做的买卖

00:00:52.280 --> 00:00:53.720
一句话能说清

00:00:53.720 --> 00:00:58.140
把散落在社区里的GPT Image 2出图咒语

00:00:58.140 --> 00:00:59.880
逆向拆解再改写

00:00:59.880 --> 00:01:02.160
重新打包成可以填空

00:01:02.160 --> 00:01:03.240
可以复用

00:01:03.240 --> 00:01:06.200
可以为给自动化流程的结构化资产

00:01:06.200 --> 00:01:09.000
要理解这东西为什么能活

00:01:09.000 --> 00:01:11.720
得先从他脚底下那个模型说起

00:01:11.720 --> 00:01:13.920
今年4月21号

00:01:13.920 --> 00:01:17.020
OpenAI发布了GPT Image 2

00:01:17.020 --> 00:01:20.660
官方叫法是ChatGPT Images 2.0

00:01:20.660 --> 00:01:23.080
这个模型有三个变化

00:01:23.080 --> 00:01:25.780
直接决定了提示词工程的价值

00:01:25.780 --> 00:01:27.000
第一个

00:01:27.000 --> 00:01:30.340
它是第一个内置原生推理的图像模型

00:01:30.340 --> 00:01:32.240
以前的图像模型

00:01:32.240 --> 00:01:33.660
你给它一段描述

00:01:33.660 --> 00:01:35.200
它边理解边画

00:01:35.200 --> 00:01:36.800
理解到哪画到哪

00:01:36.800 --> 00:01:38.460
GPT Image

00:01:38.460 --> 00:01:40.440
二把O系列的推理能力

00:01:40.440 --> 00:01:41.940
塞进了生成流程

00:01:41.940 --> 00:01:43.460
先想清楚布局

00:01:43.460 --> 00:01:44.240
层级

00:01:44.240 --> 00:01:45.300
逻辑关系

00:01:45.300 --> 00:01:46.220
再动笔

00:01:46.220 --> 00:01:48.720
复杂信息图直接受益

00:01:48.720 --> 00:01:49.760
这是质变

00:01:49.760 --> 00:01:51.000
第二个

00:01:51.000 --> 00:01:52.040
文字渲染

00:01:52.040 --> 00:01:55.700
以前AI画图最丢人的就是文字

00:01:55.700 --> 00:01:57.100
海报上的标语

00:01:57.100 --> 00:01:58.600
信息图里的数据

00:01:58.600 --> 00:02:00.520
溢出来就是鬼画符

00:02:00.520 --> 00:02:02.300
GPT Image

00:02:02.300 --> 00:02:04.040
二官方给出的数字

00:02:04.040 --> 00:02:08.600
文字准确率从90%提到99%

00:02:08.600 --> 00:02:09.880
还支持多语言

00:02:09.880 --> 00:02:11.280
中文海报

00:02:11.280 --> 00:02:13.100
标题正文都能写对

00:02:13.100 --> 00:02:15.660
我后来拿一张它的案例图

00:02:15.660 --> 00:02:17.420
为给实图模型检查

00:02:17.420 --> 00:02:19.540
对面回了句这张信息图

00:02:19.540 --> 00:02:22.180
应该是设计师用专业工具画的

00:02:22.180 --> 00:02:24.200
它是AI生成的

00:02:24.200 --> 00:02:25.540
第三个

00:02:25.540 --> 00:02:27.400
分辨率最高到2K

00:02:27.400 --> 00:02:30.380
物理仿真和人物一致性也有提升

00:02:30.380 --> 00:02:31.620
同一天

00:02:31.620 --> 00:02:33.120
OpenAI宣布

00:02:33.120 --> 00:02:34.540
Daw E全面退役

00:02:34.540 --> 00:02:36.460
这三条加起来

00:02:36.460 --> 00:02:39.060
意味着AI画图的应用边界

00:02:39.060 --> 00:02:40.520
突然扩了一大圈

00:02:40.520 --> 00:02:43.640
以前你只敢让AI画插画

00:02:43.640 --> 00:02:44.540
画风景

00:02:44.540 --> 00:02:46.100
错个细节无所谓

00:02:46.100 --> 00:02:48.140
现在从信息图

00:02:48.140 --> 00:02:48.860
海报

00:02:48.860 --> 00:02:50.300
到UI界面

00:02:50.300 --> 00:02:53.100
电商详情页和模拟社交截图

00:02:53.100 --> 00:02:54.680
这些文字密集

00:02:54.680 --> 00:02:56.340
结构严谨的场景

00:02:56.340 --> 00:02:57.940
AI都碰得动了

00:02:57.940 --> 00:02:59.760
新问题跟着来了

00:02:59.760 --> 00:03:01.420
能不能出图

00:03:01.420 --> 00:03:02.360
解决了

00:03:02.360 --> 00:03:04.020
能不能稳定出图

00:03:04.020 --> 00:03:05.080
可控出图

00:03:05.080 --> 00:03:06.400
可复用的出图

00:03:06.400 --> 00:03:07.420
没有解决

00:03:07.420 --> 00:03:08.940
同一个需求

00:03:08.940 --> 00:03:11.560
你今天写一段prompt出图不错

00:03:11.560 --> 00:03:13.460
明天换个说法就拉胯

00:03:13.460 --> 00:03:15.740
想复现别人的好图

00:03:15.740 --> 00:03:17.780
光看成品猜prompt

00:03:17.780 --> 00:03:18.980
能猜到天黑

00:03:18.980 --> 00:03:22.300
市场上所有人的痛感是一致的

00:03:22.300 --> 00:03:24.900
从抽卡玄学到填控工程

00:03:24.900 --> 00:03:27.060
中间缺一套东西

00:03:27.060 --> 00:03:28.580
仓盒的动作很快

00:03:28.580 --> 00:03:30.520
模型4月发布

00:03:30.520 --> 00:03:33.680
他连夜开始收集社区里的优质案例

00:03:33.680 --> 00:03:34.740
整理城库

00:03:34.740 --> 00:03:36.180
到8月底

00:03:36.180 --> 00:03:37.740
登顶趋势榜第一

00:03:37.740 --> 00:03:40.280
来看这个仓库长什么样

00:03:40.280 --> 00:03:42.680
它的内容主体分三层

00:03:42.680 --> 00:03:44.780
第一层,案例库

00:03:44.780 --> 00:03:49.060
526个案例,README里分成12个大类

00:03:49.060 --> 00:03:50.880
我给你报几个数

00:03:50.880 --> 00:03:53.720
UI和界面类73个

00:03:53.720 --> 00:03:55.820
海报排版类86个

00:03:55.820 --> 00:03:57.880
摄影写实类77个

00:03:57.880 --> 00:03:59.820
插画艺术类58个

00:03:59.820 --> 00:04:01.840
图表信息图53个

00:04:01.840 --> 00:04:03.680
产品电商41个

00:04:03.680 --> 00:04:05.520
品牌标志27个

00:04:05.520 --> 00:04:07.740
往下还有建筑空间

00:04:07.740 --> 00:04:08.840
人物角色

00:04:08.840 --> 00:04:09.980
场景叙事

00:04:09.980 --> 00:04:11.060
历史古风

00:04:11.060 --> 00:04:12.100
文档出版

00:04:12.100 --> 00:04:13.100
各有一批

00:04:13.100 --> 00:04:15.940
每个案例都是一张成品图

00:04:15.940 --> 00:04:18.180
配一段逆向出来的提示词

00:04:18.180 --> 00:04:19.420
第二层

00:04:19.420 --> 00:04:20.300
模板酷

00:04:20.300 --> 00:04:23.080
这是它跟其他提示词合集

00:04:23.080 --> 00:04:24.480
拉开差距的地方

00:04:24.480 --> 00:04:26.640
作者自己的说法是

00:04:26.640 --> 00:04:29.860
把390多个案例扒了个底朝天

00:04:29.860 --> 00:04:33.780
硬生生提炼出21套工业级提示词模板

00:04:33.780 --> 00:04:36.420
每套模板覆盖一个大类

00:04:36.420 --> 00:04:38.900
长在哪去就能填空的程度

00:04:38.900 --> 00:04:40.220
第三层

00:04:40.220 --> 00:04:41.260
Agent Scale

00:04:41.260 --> 00:04:46.900
整个风格库打包成一个叫GPT Image 2 Style Library的G

00:04:46.900 --> 00:04:55.000
能一行命令装进Clock Code Codex Cursor AI助手直接帮你选风格套模板

00:04:55.000 --> 00:04:59.680
这个三层结构就是Prompt as Code的实际含义

00:04:59.680 --> 00:05:03.880
提示词从一段散文变成一个结构化协议

00:05:03.880 --> 00:05:08.480
从主体 光照 到材质 布局 文案

00:05:08.480 --> 00:05:12.240
拆成一个个可以组合 可以替换的原子参数

00:05:12.240 --> 00:05:16.120
散文是给人读的 协议是给程序跑的

00:05:16.120 --> 00:05:18.320
面向的场景也变了

00:05:18.320 --> 00:05:22.960
不只是人肉复制粘贴 还有agent调用 批量脚本

00:05:22.960 --> 00:05:24.560
这些自动化工作流

00:05:24.560 --> 00:05:28.200
同一个赛道里 别加仓库再拼数量

00:05:28.200 --> 00:05:31.680
YouMind那个同类合集1100多条

00:05:31.680 --> 00:05:33.740
另一个1100条起步

00:05:33.740 --> 00:05:35.880
这个仓库500多条

00:05:35.880 --> 00:05:37.060
数量不占优

00:05:37.060 --> 00:05:40.220
但精华部分往模板和scale里沉淀

00:05:40.220 --> 00:05:43.340
堆料和做资产是两种玩法

00:05:43.340 --> 00:05:45.360
光说结构有点虚

00:05:45.360 --> 00:05:47.440
我们拿一个模板出来解剖

00:05:47.440 --> 00:05:49.720
就用UI界面类

00:05:49.720 --> 00:05:51.720
案例最多的场景之一

00:05:51.720 --> 00:05:53.960
它的模板分两种形态

00:05:53.960 --> 00:05:56.100
第一种常规模板

00:05:56.100 --> 00:05:57.720
中文的给人用

00:05:57.720 --> 00:05:59.560
你听这个格式

00:05:59.560 --> 00:06:02.320
为中括号产品类型

00:06:02.320 --> 00:06:03.260
生成一张

00:06:03.260 --> 00:06:04.640
中括号平台

00:06:04.640 --> 00:06:05.620
界面图

00:06:05.620 --> 00:06:07.080
核心功能

00:06:07.080 --> 00:06:08.760
中括号功能点A

00:06:08.760 --> 00:06:09.780
功能点B

00:06:09.780 --> 00:06:11.680
视觉风格极简

00:06:11.680 --> 00:06:13.700
主色中括号颜色

00:06:13.700 --> 00:06:15.440
布局顶部导航

00:06:15.440 --> 00:06:17.040
信息层级清晰

00:06:17.040 --> 00:06:18.160
留白充足

00:06:18.160 --> 00:06:20.900
输出高保真UI截图

00:06:20.900 --> 00:06:22.400
文字清晰可读

00:06:22.400 --> 00:06:23.860
比例9比16

00:06:23.860 --> 00:06:26.500
方括号就是你的填空位

00:06:26.500 --> 00:06:28.180
产品类型

00:06:28.180 --> 00:06:28.900
平台

00:06:28.900 --> 00:06:29.800
功能点

00:06:29.800 --> 00:06:30.560
主色

00:06:30.560 --> 00:06:31.220
布局

00:06:31.220 --> 00:06:31.920
比例

00:06:31.920 --> 00:06:32.900
六个变量

00:06:32.900 --> 00:06:35.720
填完就是一段能直接用的prompt

00:06:35.720 --> 00:06:38.860
第二种JSON进阶母版

00:06:38.860 --> 00:06:39.860
给程序用

00:06:39.860 --> 00:06:41.460
文档里明确写着

00:06:41.460 --> 00:06:43.900
推荐给agent调用

00:06:43.900 --> 00:06:46.840
同样一个健身app界面的需求

00:06:46.840 --> 00:06:49.400
变成一个JSON对象

00:06:49.400 --> 00:06:51.840
Type是UI Screenshot

00:06:51.840 --> 00:06:53.580
Platform是iOS

00:06:53.580 --> 00:06:56.940
Layout写卡片留加底部标签栏

00:06:56.940 --> 00:06:59.240
Style里嵌套主题

00:06:59.240 --> 00:07:00.680
主色字体

00:07:00.680 --> 00:07:02.060
Content里

00:07:02.060 --> 00:07:04.140
Header是Today's Activity

00:07:04.140 --> 00:07:06.380
Cards是一个数组

00:07:06.380 --> 00:07:08.060
每个元素代标题

00:07:08.060 --> 00:07:08.800
数据

00:07:08.800 --> 00:07:09.500
按钮

00:07:09.500 --> 00:07:10.640
最后

00:07:10.640 --> 00:07:12.280
Constraints写高保真

00:07:12.280 --> 00:07:13.360
文字可读

00:07:13.360 --> 00:07:14.340
9比16

00:07:14.340 --> 00:07:16.300
结构完全对得上

00:07:16.300 --> 00:07:19.480
但形态从自然语言变成了数据

00:07:19.480 --> 00:07:21.000
程序解析

00:07:21.000 --> 00:07:23.980
JSON比解析散文容易100倍

00:07:23.980 --> 00:07:27.100
这就是它能进自动化流程的原因

00:07:27.100 --> 00:07:29.140
每个大类除了母版

00:07:29.140 --> 00:07:30.960
还附一份B坑指南

00:07:30.960 --> 00:07:34.280
这部分是整个仓库最有味道的地方

00:07:34.280 --> 00:07:35.800
下一节单独讲

00:07:35.800 --> 00:07:38.940
使用路径文档里也写明白了

00:07:38.940 --> 00:07:39.700
三步

00:07:39.700 --> 00:07:40.840
第一步

00:07:40.840 --> 00:07:42.820
去精选案例区逛一圈

00:07:42.820 --> 00:07:44.520
找到你想要的方向

00:07:44.520 --> 00:07:45.700
第二步

00:07:45.700 --> 00:07:47.540
去画廊找相近案例

00:07:47.540 --> 00:07:48.640
先抄结构

00:07:48.640 --> 00:07:49.860
再抄风格词

00:07:49.860 --> 00:07:51.260
第三步

00:07:51.260 --> 00:07:52.380
打开模板页

00:07:52.380 --> 00:07:54.220
把业务变量填进去

00:07:54.220 --> 00:07:56.460
从灵感到成品prompt

00:07:56.460 --> 00:07:57.900
路径被铺平了

00:07:58.248 --> 00:08:07.248
B坑指南,值得单独讲一节,因为它是全仓库口语最声猛,传播力最强的部分。我挑几条给你听。

00:08:07.248 --> 00:08:09.068
UI类第一条

00:08:09.068 --> 00:08:10.468
不要给模糊指令

00:08:10.468 --> 00:08:12.348
原文这么写的

00:08:12.348 --> 00:08:13.468
明确平台

00:08:13.468 --> 00:08:14.168
比例

00:08:14.168 --> 00:08:14.828
布局

00:08:14.828 --> 00:08:18.008
否则模型会像个实习生一样乱排版

00:08:18.008 --> 00:08:19.968
这个比喻准得很

00:08:19.968 --> 00:08:25.108
GPT Image 2是个能力强弹需要明确攻单的实习生

00:08:25.108 --> 00:08:27.328
你给一句做个好看的界面

00:08:27.328 --> 00:08:28.728
它就自由发挥

00:08:28.728 --> 00:08:29.948
第二条

00:08:29.948 --> 00:08:31.348
强制文字锁定

00:08:31.348 --> 00:08:33.648
要求文字绝对可读

00:08:33.648 --> 00:08:35.628
必须显示指定的中文

00:08:35.628 --> 00:08:39.348
避免出现乱码按钮和毫无意义的火星文

00:08:39.348 --> 00:08:42.608
哪怕文字准确率到了99%

00:08:42.608 --> 00:08:43.648
你不锁

00:08:43.648 --> 00:08:47.608
剩下那1%就在你最显眼的地方等着你

00:08:47.608 --> 00:08:48.928
第三条

00:08:48.928 --> 00:08:50.048
我特别喜欢

00:08:50.048 --> 00:08:52.228
截图要区分平台特征

00:08:52.228 --> 00:08:54.108
X有蓝沟认证

00:08:54.108 --> 00:08:56.948
抖音有音乐叠片和点赞动画

00:08:56.948 --> 00:08:59.088
小红书是双列瀑布流

00:08:59.088 --> 00:09:01.688
生成前必须指定平台

00:09:01.688 --> 00:09:03.328
否则模型会混搭

00:09:03.328 --> 00:09:05.468
你想仿个抖音截图

00:09:05.468 --> 00:09:08.148
出来一个蓝钩加瀑布流的缝合怪

00:09:08.148 --> 00:09:09.968
摄影类有几条

00:09:09.968 --> 00:09:11.428
属于反直觉的

00:09:11.428 --> 00:09:13.548
AI画人为什么假

00:09:13.548 --> 00:09:15.348
因为太完美了

00:09:15.348 --> 00:09:17.788
所以要主动加瑕疵

00:09:17.788 --> 00:09:19.348
皮肤纹理

00:09:19.348 --> 00:09:20.728
雀斑这类细节

00:09:20.728 --> 00:09:22.648
再加轻微胶片颗粒

00:09:22.648 --> 00:09:24.508
真实感瞬间拉满

00:09:24.508 --> 00:09:25.868
还有一条

00:09:25.868 --> 00:09:27.108
用参数说话

00:09:27.108 --> 00:09:29.808
写F1.4代替前景深

00:09:29.808 --> 00:09:32.188
写50毫米代替半身照

00:09:32.188 --> 00:09:35.148
大模型吃相机参数这一套

00:09:35.148 --> 00:09:36.148
吃得很准

00:09:36.148 --> 00:09:37.748
历史古风类

00:09:37.748 --> 00:09:39.348
两条保命规则

00:09:39.348 --> 00:09:40.808
明确朝代

00:09:40.808 --> 00:09:43.768
不然大模型给你画出一个穿河服

00:09:43.768 --> 00:09:45.108
拿清朝折扇

00:09:45.108 --> 00:09:47.068
站在唐朝宫殿里的人

00:09:47.068 --> 00:09:48.968
还要强制排雷

00:09:48.968 --> 00:09:52.008
在prompt里写上禁用现代元素

00:09:52.008 --> 00:09:55.608
防止古风美女手里突然多出一杯星巴克

00:09:55.608 --> 00:09:57.388
建筑类就俩字

00:09:57.388 --> 00:09:58.148
视角

00:09:58.148 --> 00:10:00.648
建筑图最容易翻车的是

00:10:00.648 --> 00:10:01.728
透视变形

00:10:01.728 --> 00:10:04.008
写上人掩视角能压住它

00:10:04.008 --> 00:10:06.008
再送一个作弊码

00:10:06.008 --> 00:10:08.088
室外冷光配室内暖光

00:10:08.088 --> 00:10:10.168
空间高级赶立刻出来

00:10:10.168 --> 00:10:11.448
插画类

00:10:11.448 --> 00:10:12.708
慎用大师名

00:10:12.708 --> 00:10:14.388
直接写泛高

00:10:14.388 --> 00:10:16.948
模型容易照搬代表作的构图

00:10:16.948 --> 00:10:18.808
画出来一张假星空

00:10:18.808 --> 00:10:21.528
正确姿势是提取特征

00:10:21.528 --> 00:10:22.848
写旋转笔触

00:10:22.848 --> 00:10:24.268
而不是写名字

00:10:24.268 --> 00:10:25.728
你发现没有

00:10:25.728 --> 00:10:28.668
这些坑权都是实操里采出来的

00:10:28.668 --> 00:10:31.308
没有一条是坐在书房里想出来的

00:10:31.308 --> 00:10:33.048
这份指南本身

00:10:33.048 --> 00:10:35.588
就是社区几个月试错的结晶

00:10:35.588 --> 00:10:38.128
接下来说它的生态布局

00:10:38.128 --> 00:10:41.768
这部分是它登顶趋势榜的另一半答案

00:10:41.768 --> 00:10:43.368
光有内容不够

00:10:43.368 --> 00:10:46.108
这个项目把分发渠道铺满了

00:10:46.108 --> 00:10:46.968
第一

00:10:46.968 --> 00:10:49.108
AgentSkills上了NPM

00:10:49.108 --> 00:10:50.648
异形命令

00:10:50.648 --> 00:10:52.108
NPX Skills Add

00:10:52.108 --> 00:10:53.528
装进Cloud Code

00:10:53.528 --> 00:10:54.408
Codex

00:10:54.408 --> 00:10:55.128
Cursor

00:10:55.128 --> 00:10:56.028
第二

00:10:56.028 --> 00:10:58.708
Cloud Code插件市场也能搜到

00:10:58.708 --> 00:10:59.808
第三

00:10:59.808 --> 00:11:02.048
作者做了个可视化网站

00:11:02.048 --> 00:11:03.628
案例在线浏览

00:11:03.628 --> 00:11:06.048
登录之后还能测试生成

00:11:06.048 --> 00:11:07.148
第四

00:11:07.148 --> 00:11:09.828
REME有中英日多个语言版本

00:11:09.828 --> 00:11:10.848
第五

00:11:10.848 --> 00:11:13.548
投了软一封的科技爱好者周刊

00:11:13.548 --> 00:11:15.528
组合拳打下来

00:11:15.528 --> 00:11:19.048
加上Addi这些大V在X上转发

00:11:19.048 --> 00:11:21.448
才有12天4200STAR

00:11:21.448 --> 00:11:23.268
登顶趋势榜的爆发

00:11:23.268 --> 00:11:25.428
时间线值得复盘

00:11:25.428 --> 00:11:28.008
4月21号模型发布

00:11:28.008 --> 00:11:29.928
作者当天开始收集

00:11:29.928 --> 00:11:31.108
8月底登顶

00:11:31.108 --> 00:11:33.108
4个月18000STAR

00:11:33.108 --> 00:11:35.408
这个节奏说明一件事

00:11:35.408 --> 00:11:37.588
热点项目的窗口期极短

00:11:37.588 --> 00:11:39.288
GPP Image

00:11:39.288 --> 00:11:41.088
2全量开放的那一刻

00:11:41.088 --> 00:11:43.148
所有人都在找提示词

00:11:43.148 --> 00:11:45.508
谁先把散落的案例结构化

00:11:45.508 --> 00:11:47.868
谁就是那个应用百科全书

00:11:47.868 --> 00:11:49.788
这是时机的胜利

00:11:49.788 --> 00:11:51.548
也是执行力的胜利

00:11:51.548 --> 00:11:54.488
12天里保持高频更新

00:11:54.488 --> 00:11:57.128
案例数从300多滚到500多

00:11:57.128 --> 00:11:59.028
STAR就跟着滚上去

00:11:59.028 --> 00:12:00.508
讲完好话

00:12:00.508 --> 00:12:02.168
这部分泼点冷水

00:12:02.168 --> 00:12:03.688
有三件事

00:12:03.688 --> 00:12:05.268
你用之前最好知道

00:12:05.268 --> 00:12:06.528
第一件

00:12:06.528 --> 00:12:07.568
内容来源

00:12:07.568 --> 00:12:09.348
这个仓库的案例

00:12:09.348 --> 00:12:14.088
主要来自YouMind和OpenNANA两个社区的公开提示词

00:12:14.088 --> 00:12:17.928
作者做的是逆向工程家AI改写

00:12:17.928 --> 00:12:19.928
仓库里挂着一个徽章

00:12:19.928 --> 00:12:23.228
写着100%Original AI Rewritten

00:12:23.228 --> 00:12:25.728
ReadMe的声明也算坦诚

00:12:25.728 --> 00:12:27.708
参考公开提示词库

00:12:27.708 --> 00:12:29.268
做学习总结研究

00:12:29.268 --> 00:12:30.548
版权归原作

00:12:30.548 --> 00:12:32.868
有意义可以提医术删除

00:12:32.868 --> 00:12:34.528
披露是够了

00:12:34.528 --> 00:12:37.148
但你要明白自己拿到的是什么

00:12:37.148 --> 00:12:40.728
一份逆向家改写过的二手经验集

00:12:40.728 --> 00:12:42.888
原创实验产出谈不上

00:12:42.888 --> 00:12:44.108
第二件

00:12:44.108 --> 00:12:45.408
商业化浓度

00:12:45.408 --> 00:12:48.568
这个项目带着明显的商业设计

00:12:48.568 --> 00:12:52.808
仓库里有六家API中转站的赞助商位

00:12:52.808 --> 00:12:54.868
作者建了个付费社区

00:12:54.868 --> 00:12:57.088
9块91次性付费

00:12:57.088 --> 00:12:58.648
网站上还有会员制

00:12:58.648 --> 00:13:00.688
5美元300个credits

00:13:00.688 --> 00:13:02.408
我的看法是

00:13:02.408 --> 00:13:04.228
开源和赚钱不冲突

00:13:04.228 --> 00:13:06.948
这是开源可持续的一种路径

00:13:06.948 --> 00:13:09.588
内容本身MIT协议随便用

00:13:09.588 --> 00:13:12.448
但你看到工业集三个字的时候

00:13:12.448 --> 00:13:13.648
心里要清楚

00:13:13.648 --> 00:13:15.428
这是一份经验打包

00:13:15.428 --> 00:13:17.088
不是一份实验报告

00:13:17.088 --> 00:13:19.988
哪些模板真的稳定产出好图

00:13:19.988 --> 00:13:21.988
哪些只是看起来专业

00:13:21.988 --> 00:13:23.508
你得自己上手验

00:13:23.508 --> 00:13:25.228
顺带说一句

00:13:25.228 --> 00:13:28.068
我上一期视频刚拆过一个项目

00:13:28.068 --> 00:13:29.748
Le Flat Less Eye Tone

00:13:29.748 --> 00:13:34.148
人家拿283万字对照语料做统计检验

00:13:34.148 --> 00:13:36.828
26项特征只留11项

00:13:36.828 --> 00:13:39.428
跟那种知识生产方式比

00:13:39.428 --> 00:13:41.648
这个仓库是另一个流派

00:13:41.648 --> 00:13:42.528
快而操

00:13:42.528 --> 00:13:43.428
覆盖广

00:13:43.428 --> 00:13:45.548
靠社区共创滚动迭代

00:13:45.548 --> 00:13:47.608
两种流派没有高下

00:13:47.608 --> 00:13:50.368
但你得知道自己吃的是哪重粮

00:13:50.368 --> 00:13:51.708
第三件

00:13:51.708 --> 00:13:52.588
保质期

00:13:52.588 --> 00:13:56.368
提示词仓库的半衰期很短

00:13:56.368 --> 00:13:57.628
模型每次更新

00:13:57.628 --> 00:13:59.628
旧的技巧就贬值一批

00:13:59.628 --> 00:14:01.968
到13年就退役了

00:14:01.968 --> 00:14:05.048
GPT Image 2的技巧能吃多久

00:14:05.048 --> 00:14:05.968
没人知道

00:14:05.968 --> 00:14:08.748
作者的应对是持续更新

00:14:08.748 --> 00:14:10.468
仓库现在还在动

00:14:10.468 --> 00:14:12.708
这是它目前最大的护城河

00:14:12.708 --> 00:14:14.648
最后给你算笔账

00:14:14.648 --> 00:14:16.608
这东西对谁有用

00:14:16.608 --> 00:14:19.288
你要是经常做自媒体配图

00:14:19.288 --> 00:14:21.388
电商详情页这些需求

00:14:21.388 --> 00:14:22.768
或者活动海报

00:14:22.768 --> 00:14:24.328
或者做PPT

00:14:24.328 --> 00:14:25.568
缺失觉素材

00:14:25.568 --> 00:14:26.988
模板拿来填空

00:14:26.988 --> 00:14:29.848
出图质量的下限直接抬高一截

00:14:29.848 --> 00:14:31.728
何况它免费

00:14:31.728 --> 00:14:33.608
你要是在玩Agent

00:14:33.608 --> 00:14:36.808
想给自动化流程加一个出图环节

00:14:36.808 --> 00:14:40.788
JSON母版加Scale的设计就是给你准备的

00:14:40.788 --> 00:14:42.428
同类仓库里

00:14:42.428 --> 00:14:44.228
这块它做得最顺手

00:14:44.228 --> 00:14:47.088
想学图像Prompt怎么写的人

00:14:47.088 --> 00:14:48.588
把它当案例库翻

00:14:48.588 --> 00:14:50.608
500多个逆向样本

00:14:50.608 --> 00:14:52.928
看多了你自己就会写了

00:14:52.928 --> 00:14:54.548
装法很简单

00:14:54.548 --> 00:14:56.028
异形命令

00:14:56.028 --> 00:15:00.828
NPX Skills at Freestyle Fly-Awesome GPT Image 2

00:15:00.828 --> 00:15:04.828
Scale名叫GPT Image 2 Style Library

00:15:04.828 --> 00:15:08.288
或者Cloud Code插件市场直接搜

00:15:08.288 --> 00:15:11.048
最后说说我的整体判断

00:15:11.048 --> 00:15:14.268
这类项目的价值要分两层看

00:15:14.268 --> 00:15:17.928
案例和模板这层是当下好用的工具

00:15:17.928 --> 00:15:19.788
趁热用别供着

00:15:19.788 --> 00:15:21.648
真正值得带走的

00:15:21.648 --> 00:15:24.288
是Prompt as Code这个思路本身

00:15:24.288 --> 00:15:27.328
把模糊的描述拆成原子参数

00:15:27.328 --> 00:15:28.908
散文变成协议

00:15:28.908 --> 00:15:31.788
一次性抽卡变成可复用资产

00:15:31.788 --> 00:15:35.588
这个思路比任何一份具体模板都长寿

00:15:35.588 --> 00:15:37.148
模型换代之后

00:15:37.148 --> 00:15:38.328
它依然成立

00:15:38.328 --> 00:15:40.548
以后你写任何prompt

00:15:40.548 --> 00:15:41.488
图像的

00:15:41.488 --> 00:15:42.308
文本的

00:15:42.308 --> 00:15:43.848
都可以问自己一句

00:15:43.848 --> 00:15:45.828
我写的是一段散文

00:15:45.828 --> 00:15:47.128
还是一份协议

00:15:47.128 --> 00:15:50.128
我是每天跟AI打交道的人

00:15:50.128 --> 00:15:51.808
工具天天换

00:15:51.808 --> 00:15:53.408
思路可以一直攒

00:15:53.408 --> 00:15:55.088
我们下期见