1
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。

2
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。

3
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。

4
00:00:50,600 --> 00:00:52,280
他做的买卖

5
00:00:52,280 --> 00:00:53,720
一句话能说清

6
00:00:53,720 --> 00:00:58,140
把散落在社区里的GPT Image 2出图咒语

7
00:00:58,140 --> 00:00:59,880
逆向拆解再改写

8
00:00:59,880 --> 00:01:02,160
重新打包成可以填空

9
00:01:02,160 --> 00:01:03,240
可以复用

10
00:01:03,240 --> 00:01:06,200
可以为给自动化流程的结构化资产

11
00:01:06,200 --> 00:01:09,000
要理解这东西为什么能活

12
00:01:09,000 --> 00:01:11,720
得先从他脚底下那个模型说起

13
00:01:11,720 --> 00:01:13,920
今年4月21号

14
00:01:13,920 --> 00:01:17,020
OpenAI发布了GPT Image 2

15
00:01:17,020 --> 00:01:20,660
官方叫法是ChatGPT Images 2.0

16
00:01:20,660 --> 00:01:23,080
这个模型有三个变化

17
00:01:23,080 --> 00:01:25,780
直接决定了提示词工程的价值

18
00:01:25,780 --> 00:01:27,000
第一个

19
00:01:27,000 --> 00:01:30,340
它是第一个内置原生推理的图像模型

20
00:01:30,340 --> 00:01:32,240
以前的图像模型

21
00:01:32,240 --> 00:01:33,660
你给它一段描述

22
00:01:33,660 --> 00:01:35,200
它边理解边画

23
00:01:35,200 --> 00:01:36,800
理解到哪画到哪

24
00:01:36,800 --> 00:01:38,460
GPT Image

25
00:01:38,460 --> 00:01:40,440
二把O系列的推理能力

26
00:01:40,440 --> 00:01:41,940
塞进了生成流程

27
00:01:41,940 --> 00:01:43,460
先想清楚布局

28
00:01:43,460 --> 00:01:44,240
层级

29
00:01:44,240 --> 00:01:45,300
逻辑关系

30
00:01:45,300 --> 00:01:46,220
再动笔

31
00:01:46,220 --> 00:01:48,720
复杂信息图直接受益

32
00:01:48,720 --> 00:01:49,760
这是质变

33
00:01:49,760 --> 00:01:51,000
第二个

34
00:01:51,000 --> 00:01:52,040
文字渲染

35
00:01:52,040 --> 00:01:55,700
以前AI画图最丢人的就是文字

36
00:01:55,700 --> 00:01:57,100
海报上的标语

37
00:01:57,100 --> 00:01:58,600
信息图里的数据

38
00:01:58,600 --> 00:02:00,520
溢出来就是鬼画符

39
00:02:00,520 --> 00:02:02,300
GPT Image

40
00:02:02,300 --> 00:02:04,040
二官方给出的数字

41
00:02:04,040 --> 00:02:08,600
文字准确率从90%提到99%

42
00:02:08,600 --> 00:02:09,880
还支持多语言

43
00:02:09,880 --> 00:02:11,280
中文海报

44
00:02:11,280 --> 00:02:13,100
标题正文都能写对

45
00:02:13,100 --> 00:02:15,660
我后来拿一张它的案例图

46
00:02:15,660 --> 00:02:17,420
为给实图模型检查

47
00:02:17,420 --> 00:02:19,540
对面回了句这张信息图

48
00:02:19,540 --> 00:02:22,180
应该是设计师用专业工具画的

49
00:02:22,180 --> 00:02:24,200
它是AI生成的

50
00:02:24,200 --> 00:02:25,540
第三个

51
00:02:25,540 --> 00:02:27,400
分辨率最高到2K

52
00:02:27,400 --> 00:02:30,380
物理仿真和人物一致性也有提升

53
00:02:30,380 --> 00:02:31,620
同一天

54
00:02:31,620 --> 00:02:33,120
OpenAI宣布

55
00:02:33,120 --> 00:02:34,540
Daw E全面退役

56
00:02:34,540 --> 00:02:36,460
这三条加起来

57
00:02:36,460 --> 00:02:39,060
意味着AI画图的应用边界

58
00:02:39,060 --> 00:02:40,520
突然扩了一大圈

59
00:02:40,520 --> 00:02:43,640
以前你只敢让AI画插画

60
00:02:43,640 --> 00:02:44,540
画风景

61
00:02:44,540 --> 00:02:46,100
错个细节无所谓

62
00:02:46,100 --> 00:02:48,140
现在从信息图

63
00:02:48,140 --> 00:02:48,860
海报

64
00:02:48,860 --> 00:02:50,300
到UI界面

65
00:02:50,300 --> 00:02:53,100
电商详情页和模拟社交截图

66
00:02:53,100 --> 00:02:54,680
这些文字密集

67
00:02:54,680 --> 00:02:56,340
结构严谨的场景

68
00:02:56,340 --> 00:02:57,940
AI都碰得动了

69
00:02:57,940 --> 00:02:59,760
新问题跟着来了

70
00:02:59,760 --> 00:03:01,420
能不能出图

71
00:03:01,420 --> 00:03:02,360
解决了

72
00:03:02,360 --> 00:03:04,020
能不能稳定出图

73
00:03:04,020 --> 00:03:05,080
可控出图

74
00:03:05,080 --> 00:03:06,400
可复用的出图

75
00:03:06,400 --> 00:03:07,420
没有解决

76
00:03:07,420 --> 00:03:08,940
同一个需求

77
00:03:08,940 --> 00:03:11,560
你今天写一段prompt出图不错

78
00:03:11,560 --> 00:03:13,460
明天换个说法就拉胯

79
00:03:13,460 --> 00:03:15,740
想复现别人的好图

80
00:03:15,740 --> 00:03:17,780
光看成品猜prompt

81
00:03:17,780 --> 00:03:18,980
能猜到天黑

82
00:03:18,980 --> 00:03:22,300
市场上所有人的痛感是一致的

83
00:03:22,300 --> 00:03:24,900
从抽卡玄学到填控工程

84
00:03:24,900 --> 00:03:27,060
中间缺一套东西

85
00:03:27,060 --> 00:03:28,580
仓盒的动作很快

86
00:03:28,580 --> 00:03:30,520
模型4月发布

87
00:03:30,520 --> 00:03:33,680
他连夜开始收集社区里的优质案例

88
00:03:33,680 --> 00:03:34,740
整理城库

89
00:03:34,740 --> 00:03:36,180
到8月底

90
00:03:36,180 --> 00:03:37,740
登顶趋势榜第一

91
00:03:37,740 --> 00:03:40,280
来看这个仓库长什么样

92
00:03:40,280 --> 00:03:42,680
它的内容主体分三层

93
00:03:42,680 --> 00:03:44,780
第一层,案例库

94
00:03:44,780 --> 00:03:49,060
526个案例,README里分成12个大类

95
00:03:49,060 --> 00:03:50,880
我给你报几个数

96
00:03:50,880 --> 00:03:53,720
UI和界面类73个

97
00:03:53,720 --> 00:03:55,820
海报排版类86个

98
00:03:55,820 --> 00:03:57,880
摄影写实类77个

99
00:03:57,880 --> 00:03:59,820
插画艺术类58个

100
00:03:59,820 --> 00:04:01,840
图表信息图53个

101
00:04:01,840 --> 00:04:03,680
产品电商41个

102
00:04:03,680 --> 00:04:05,520
品牌标志27个

103
00:04:05,520 --> 00:04:07,740
往下还有建筑空间

104
00:04:07,740 --> 00:04:08,840
人物角色

105
00:04:08,840 --> 00:04:09,980
场景叙事

106
00:04:09,980 --> 00:04:11,060
历史古风

107
00:04:11,060 --> 00:04:12,100
文档出版

108
00:04:12,100 --> 00:04:13,100
各有一批

109
00:04:13,100 --> 00:04:15,940
每个案例都是一张成品图

110
00:04:15,940 --> 00:04:18,180
配一段逆向出来的提示词

111
00:04:18,180 --> 00:04:19,420
第二层

112
00:04:19,420 --> 00:04:20,300
模板酷

113
00:04:20,300 --> 00:04:23,080
这是它跟其他提示词合集

114
00:04:23,080 --> 00:04:24,480
拉开差距的地方

115
00:04:24,480 --> 00:04:26,640
作者自己的说法是

116
00:04:26,640 --> 00:04:29,860
把390多个案例扒了个底朝天

117
00:04:29,860 --> 00:04:33,780
硬生生提炼出21套工业级提示词模板

118
00:04:33,780 --> 00:04:36,420
每套模板覆盖一个大类

119
00:04:36,420 --> 00:04:38,900
长在哪去就能填空的程度

120
00:04:38,900 --> 00:04:40,220
第三层

121
00:04:40,220 --> 00:04:41,260
Agent Scale

122
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

123
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

124
00:04:55,000 --> 00:04:59,680
这个三层结构就是Prompt as Code的实际含义

125
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

126
00:05:03,880 --> 00:05:08,480
从主体 光照 到材质 布局 文案

127
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

128
00:05:12,240 --> 00:05:16,120
散文是给人读的 协议是给程序跑的

129
00:05:16,120 --> 00:05:18,320
面向的场景也变了

130
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

131
00:05:22,960 --> 00:05:24,560
这些自动化工作流

132
00:05:24,560 --> 00:05:28,200
同一个赛道里 别加仓库再拼数量

133
00:05:28,200 --> 00:05:31,680
YouMind那个同类合集1100多条

134
00:05:31,680 --> 00:05:33,740
另一个1100条起步

135
00:05:33,740 --> 00:05:35,880
这个仓库500多条

136
00:05:35,880 --> 00:05:37,060
数量不占优

137
00:05:37,060 --> 00:05:40,220
但精华部分往模板和scale里沉淀

138
00:05:40,220 --> 00:05:43,340
堆料和做资产是两种玩法

139
00:05:43,340 --> 00:05:45,360
光说结构有点虚

140
00:05:45,360 --> 00:05:47,440
我们拿一个模板出来解剖

141
00:05:47,440 --> 00:05:49,720
就用UI界面类

142
00:05:49,720 --> 00:05:51,720
案例最多的场景之一

143
00:05:51,720 --> 00:05:53,960
它的模板分两种形态

144
00:05:53,960 --> 00:05:56,100
第一种常规模板

145
00:05:56,100 --> 00:05:57,720
中文的给人用

146
00:05:57,720 --> 00:05:59,560
你听这个格式

147
00:05:59,560 --> 00:06:02,320
为中括号产品类型

148
00:06:02,320 --> 00:06:03,260
生成一张

149
00:06:03,260 --> 00:06:04,640
中括号平台

150
00:06:04,640 --> 00:06:05,620
界面图

151
00:06:05,620 --> 00:06:07,080
核心功能

152
00:06:07,080 --> 00:06:08,760
中括号功能点A

153
00:06:08,760 --> 00:06:09,780
功能点B

154
00:06:09,780 --> 00:06:11,680
视觉风格极简

155
00:06:11,680 --> 00:06:13,700
主色中括号颜色

156
00:06:13,700 --> 00:06:15,440
布局顶部导航

157
00:06:15,440 --> 00:06:17,040
信息层级清晰

158
00:06:17,040 --> 00:06:18,160
留白充足

159
00:06:18,160 --> 00:06:20,900
输出高保真UI截图

160
00:06:20,900 --> 00:06:22,400
文字清晰可读

161
00:06:22,400 --> 00:06:23,860
比例9比16

162
00:06:23,860 --> 00:06:26,500
方括号就是你的填空位

163
00:06:26,500 --> 00:06:28,180
产品类型

164
00:06:28,180 --> 00:06:28,900
平台

165
00:06:28,900 --> 00:06:29,800
功能点

166
00:06:29,800 --> 00:06:30,560
主色

167
00:06:30,560 --> 00:06:31,220
布局

168
00:06:31,220 --> 00:06:31,920
比例

169
00:06:31,920 --> 00:06:32,900
六个变量

170
00:06:32,900 --> 00:06:35,720
填完就是一段能直接用的prompt

171
00:06:35,720 --> 00:06:38,860
第二种JSON进阶母版

172
00:06:38,860 --> 00:06:39,860
给程序用

173
00:06:39,860 --> 00:06:41,460
文档里明确写着

174
00:06:41,460 --> 00:06:43,900
推荐给agent调用

175
00:06:43,900 --> 00:06:46,840
同样一个健身app界面的需求

176
00:06:46,840 --> 00:06:49,400
变成一个JSON对象

177
00:06:49,400 --> 00:06:51,840
Type是UI Screenshot

178
00:06:51,840 --> 00:06:53,580
Platform是iOS

179
00:06:53,580 --> 00:06:56,940
Layout写卡片留加底部标签栏

180
00:06:56,940 --> 00:06:59,240
Style里嵌套主题

181
00:06:59,240 --> 00:07:00,680
主色字体

182
00:07:00,680 --> 00:07:02,060
Content里

183
00:07:02,060 --> 00:07:04,140
Header是Today's Activity

184
00:07:04,140 --> 00:07:06,380
Cards是一个数组

185
00:07:06,380 --> 00:07:08,060
每个元素代标题

186
00:07:08,060 --> 00:07:08,800
数据

187
00:07:08,800 --> 00:07:09,500
按钮

188
00:07:09,500 --> 00:07:10,640
最后

189
00:07:10,640 --> 00:07:12,280
Constraints写高保真

190
00:07:12,280 --> 00:07:13,360
文字可读

191
00:07:13,360 --> 00:07:14,340
9比16

192
00:07:14,340 --> 00:07:16,300
结构完全对得上

193
00:07:16,300 --> 00:07:19,480
但形态从自然语言变成了数据

194
00:07:19,480 --> 00:07:21,000
程序解析

195
00:07:21,000 --> 00:07:23,980
JSON比解析散文容易100倍

196
00:07:23,980 --> 00:07:27,100
这就是它能进自动化流程的原因

197
00:07:27,100 --> 00:07:29,140
每个大类除了母版

198
00:07:29,140 --> 00:07:30,960
还附一份B坑指南

199
00:07:30,960 --> 00:07:34,280
这部分是整个仓库最有味道的地方

200
00:07:34,280 --> 00:07:35,800
下一节单独讲

201
00:07:35,800 --> 00:07:38,940
使用路径文档里也写明白了

202
00:07:38,940 --> 00:07:39,700
三步

203
00:07:39,700 --> 00:07:40,840
第一步

204
00:07:40,840 --> 00:07:42,820
去精选案例区逛一圈

205
00:07:42,820 --> 00:07:44,520
找到你想要的方向

206
00:07:44,520 --> 00:07:45,700
第二步

207
00:07:45,700 --> 00:07:47,540
去画廊找相近案例

208
00:07:47,540 --> 00:07:48,640
先抄结构

209
00:07:48,640 --> 00:07:49,860
再抄风格词

210
00:07:49,860 --> 00:07:51,260
第三步

211
00:07:51,260 --> 00:07:52,380
打开模板页

212
00:07:52,380 --> 00:07:54,220
把业务变量填进去

213
00:07:54,220 --> 00:07:56,460
从灵感到成品prompt

214
00:07:56,460 --> 00:07:57,900
路径被铺平了

215
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。

216
00:08:07,248 --> 00:08:09,068
UI类第一条

217
00:08:09,068 --> 00:08:10,468
不要给模糊指令

218
00:08:10,468 --> 00:08:12,348
原文这么写的

219
00:08:12,348 --> 00:08:13,468
明确平台

220
00:08:13,468 --> 00:08:14,168
比例

221
00:08:14,168 --> 00:08:14,828
布局

222
00:08:14,828 --> 00:08:18,008
否则模型会像个实习生一样乱排版

223
00:08:18,008 --> 00:08:19,968
这个比喻准得很

224
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

225
00:08:25,108 --> 00:08:27,328
你给一句做个好看的界面

226
00:08:27,328 --> 00:08:28,728
它就自由发挥

227
00:08:28,728 --> 00:08:29,948
第二条

228
00:08:29,948 --> 00:08:31,348
强制文字锁定

229
00:08:31,348 --> 00:08:33,648
要求文字绝对可读

230
00:08:33,648 --> 00:08:35,628
必须显示指定的中文

231
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

232
00:08:39,348 --> 00:08:42,608
哪怕文字准确率到了99%

233
00:08:42,608 --> 00:08:43,648
你不锁

234
00:08:43,648 --> 00:08:47,608
剩下那1%就在你最显眼的地方等着你

235
00:08:47,608 --> 00:08:48,928
第三条

236
00:08:48,928 --> 00:08:50,048
我特别喜欢

237
00:08:50,048 --> 00:08:52,228
截图要区分平台特征

238
00:08:52,228 --> 00:08:54,108
X有蓝沟认证

239
00:08:54,108 --> 00:08:56,948
抖音有音乐叠片和点赞动画

240
00:08:56,948 --> 00:08:59,088
小红书是双列瀑布流

241
00:08:59,088 --> 00:09:01,688
生成前必须指定平台

242
00:09:01,688 --> 00:09:03,328
否则模型会混搭

243
00:09:03,328 --> 00:09:05,468
你想仿个抖音截图

244
00:09:05,468 --> 00:09:08,148
出来一个蓝钩加瀑布流的缝合怪

245
00:09:08,148 --> 00:09:09,968
摄影类有几条

246
00:09:09,968 --> 00:09:11,428
属于反直觉的

247
00:09:11,428 --> 00:09:13,548
AI画人为什么假

248
00:09:13,548 --> 00:09:15,348
因为太完美了

249
00:09:15,348 --> 00:09:17,788
所以要主动加瑕疵

250
00:09:17,788 --> 00:09:19,348
皮肤纹理

251
00:09:19,348 --> 00:09:20,728
雀斑这类细节

252
00:09:20,728 --> 00:09:22,648
再加轻微胶片颗粒

253
00:09:22,648 --> 00:09:24,508
真实感瞬间拉满

254
00:09:24,508 --> 00:09:25,868
还有一条

255
00:09:25,868 --> 00:09:27,108
用参数说话

256
00:09:27,108 --> 00:09:29,808
写F1.4代替前景深

257
00:09:29,808 --> 00:09:32,188
写50毫米代替半身照

258
00:09:32,188 --> 00:09:35,148
大模型吃相机参数这一套

259
00:09:35,148 --> 00:09:36,148
吃得很准

260
00:09:36,148 --> 00:09:37,748
历史古风类

261
00:09:37,748 --> 00:09:39,348
两条保命规则

262
00:09:39,348 --> 00:09:40,808
明确朝代

263
00:09:40,808 --> 00:09:43,768
不然大模型给你画出一个穿河服

264
00:09:43,768 --> 00:09:45,108
拿清朝折扇

265
00:09:45,108 --> 00:09:47,068
站在唐朝宫殿里的人

266
00:09:47,068 --> 00:09:48,968
还要强制排雷

267
00:09:48,968 --> 00:09:52,008
在prompt里写上禁用现代元素

268
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

269
00:09:55,608 --> 00:09:57,388
建筑类就俩字

270
00:09:57,388 --> 00:09:58,148
视角

271
00:09:58,148 --> 00:10:00,648
建筑图最容易翻车的是

272
00:10:00,648 --> 00:10:01,728
透视变形

273
00:10:01,728 --> 00:10:04,008
写上人掩视角能压住它

274
00:10:04,008 --> 00:10:06,008
再送一个作弊码

275
00:10:06,008 --> 00:10:08,088
室外冷光配室内暖光

276
00:10:08,088 --> 00:10:10,168
空间高级赶立刻出来

277
00:10:10,168 --> 00:10:11,448
插画类

278
00:10:11,448 --> 00:10:12,708
慎用大师名

279
00:10:12,708 --> 00:10:14,388
直接写泛高

280
00:10:14,388 --> 00:10:16,948
模型容易照搬代表作的构图

281
00:10:16,948 --> 00:10:18,808
画出来一张假星空

282
00:10:18,808 --> 00:10:21,528
正确姿势是提取特征

283
00:10:21,528 --> 00:10:22,848
写旋转笔触

284
00:10:22,848 --> 00:10:24,268
而不是写名字

285
00:10:24,268 --> 00:10:25,728
你发现没有

286
00:10:25,728 --> 00:10:28,668
这些坑权都是实操里采出来的

287
00:10:28,668 --> 00:10:31,308
没有一条是坐在书房里想出来的

288
00:10:31,308 --> 00:10:33,048
这份指南本身

289
00:10:33,048 --> 00:10:35,588
就是社区几个月试错的结晶

290
00:10:35,588 --> 00:10:38,128
接下来说它的生态布局

291
00:10:38,128 --> 00:10:41,768
这部分是它登顶趋势榜的另一半答案

292
00:10:41,768 --> 00:10:43,368
光有内容不够

293
00:10:43,368 --> 00:10:46,108
这个项目把分发渠道铺满了

294
00:10:46,108 --> 00:10:46,968
第一

295
00:10:46,968 --> 00:10:49,108
AgentSkills上了NPM

296
00:10:49,108 --> 00:10:50,648
异形命令

297
00:10:50,648 --> 00:10:52,108
NPX Skills Add

298
00:10:52,108 --> 00:10:53,528
装进Cloud Code

299
00:10:53,528 --> 00:10:54,408
Codex

300
00:10:54,408 --> 00:10:55,128
Cursor

301
00:10:55,128 --> 00:10:56,028
第二

302
00:10:56,028 --> 00:10:58,708
Cloud Code插件市场也能搜到

303
00:10:58,708 --> 00:10:59,808
第三

304
00:10:59,808 --> 00:11:02,048
作者做了个可视化网站

305
00:11:02,048 --> 00:11:03,628
案例在线浏览

306
00:11:03,628 --> 00:11:06,048
登录之后还能测试生成

307
00:11:06,048 --> 00:11:07,148
第四

308
00:11:07,148 --> 00:11:09,828
REME有中英日多个语言版本

309
00:11:09,828 --> 00:11:10,848
第五

310
00:11:10,848 --> 00:11:13,548
投了软一封的科技爱好者周刊

311
00:11:13,548 --> 00:11:15,528
组合拳打下来

312
00:11:15,528 --> 00:11:19,048
加上Addi这些大V在X上转发

313
00:11:19,048 --> 00:11:21,448
才有12天4200STAR

314
00:11:21,448 --> 00:11:23,268
登顶趋势榜的爆发

315
00:11:23,268 --> 00:11:25,428
时间线值得复盘

316
00:11:25,428 --> 00:11:28,008
4月21号模型发布

317
00:11:28,008 --> 00:11:29,928
作者当天开始收集

318
00:11:29,928 --> 00:11:31,108
8月底登顶

319
00:11:31,108 --> 00:11:33,108
4个月18000STAR

320
00:11:33,108 --> 00:11:35,408
这个节奏说明一件事

321
00:11:35,408 --> 00:11:37,588
热点项目的窗口期极短

322
00:11:37,588 --> 00:11:39,288
GPP Image

323
00:11:39,288 --> 00:11:41,088
2全量开放的那一刻

324
00:11:41,088 --> 00:11:43,148
所有人都在找提示词

325
00:11:43,148 --> 00:11:45,508
谁先把散落的案例结构化

326
00:11:45,508 --> 00:11:47,868
谁就是那个应用百科全书

327
00:11:47,868 --> 00:11:49,788
这是时机的胜利

328
00:11:49,788 --> 00:11:51,548
也是执行力的胜利

329
00:11:51,548 --> 00:11:54,488
12天里保持高频更新

330
00:11:54,488 --> 00:11:57,128
案例数从300多滚到500多

331
00:11:57,128 --> 00:11:59,028
STAR就跟着滚上去

332
00:11:59,028 --> 00:12:00,508
讲完好话

333
00:12:00,508 --> 00:12:02,168
这部分泼点冷水

334
00:12:02,168 --> 00:12:03,688
有三件事

335
00:12:03,688 --> 00:12:05,268
你用之前最好知道

336
00:12:05,268 --> 00:12:06,528
第一件

337
00:12:06,528 --> 00:12:07,568
内容来源

338
00:12:07,568 --> 00:12:09,348
这个仓库的案例

339
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

340
00:12:14,088 --> 00:12:17,928
作者做的是逆向工程家AI改写

341
00:12:17,928 --> 00:12:19,928
仓库里挂着一个徽章

342
00:12:19,928 --> 00:12:23,228
写着100%Original AI Rewritten

343
00:12:23,228 --> 00:12:25,728
ReadMe的声明也算坦诚

344
00:12:25,728 --> 00:12:27,708
参考公开提示词库

345
00:12:27,708 --> 00:12:29,268
做学习总结研究

346
00:12:29,268 --> 00:12:30,548
版权归原作

347
00:12:30,548 --> 00:12:32,868
有意义可以提医术删除

348
00:12:32,868 --> 00:12:34,528
披露是够了

349
00:12:34,528 --> 00:12:37,148
但你要明白自己拿到的是什么

350
00:12:37,148 --> 00:12:40,728
一份逆向家改写过的二手经验集

351
00:12:40,728 --> 00:12:42,888
原创实验产出谈不上

352
00:12:42,888 --> 00:12:44,108
第二件

353
00:12:44,108 --> 00:12:45,408
商业化浓度

354
00:12:45,408 --> 00:12:48,568
这个项目带着明显的商业设计

355
00:12:48,568 --> 00:12:52,808
仓库里有六家API中转站的赞助商位

356
00:12:52,808 --> 00:12:54,868
作者建了个付费社区

357
00:12:54,868 --> 00:12:57,088
9块91次性付费

358
00:12:57,088 --> 00:12:58,648
网站上还有会员制

359
00:12:58,648 --> 00:13:00,688
5美元300个credits

360
00:13:00,688 --> 00:13:02,408
我的看法是

361
00:13:02,408 --> 00:13:04,228
开源和赚钱不冲突

362
00:13:04,228 --> 00:13:06,948
这是开源可持续的一种路径

363
00:13:06,948 --> 00:13:09,588
内容本身MIT协议随便用

364
00:13:09,588 --> 00:13:12,448
但你看到工业集三个字的时候

365
00:13:12,448 --> 00:13:13,648
心里要清楚

366
00:13:13,648 --> 00:13:15,428
这是一份经验打包

367
00:13:15,428 --> 00:13:17,088
不是一份实验报告

368
00:13:17,088 --> 00:13:19,988
哪些模板真的稳定产出好图

369
00:13:19,988 --> 00:13:21,988
哪些只是看起来专业

370
00:13:21,988 --> 00:13:23,508
你得自己上手验

371
00:13:23,508 --> 00:13:25,228
顺带说一句

372
00:13:25,228 --> 00:13:28,068
我上一期视频刚拆过一个项目

373
00:13:28,068 --> 00:13:29,748
Le Flat Less Eye Tone

374
00:13:29,748 --> 00:13:34,148
人家拿283万字对照语料做统计检验

375
00:13:34,148 --> 00:13:36,828
26项特征只留11项

376
00:13:36,828 --> 00:13:39,428
跟那种知识生产方式比

377
00:13:39,428 --> 00:13:41,648
这个仓库是另一个流派

378
00:13:41,648 --> 00:13:42,528
快而操

379
00:13:42,528 --> 00:13:43,428
覆盖广

380
00:13:43,428 --> 00:13:45,548
靠社区共创滚动迭代

381
00:13:45,548 --> 00:13:47,608
两种流派没有高下

382
00:13:47,608 --> 00:13:50,368
但你得知道自己吃的是哪重粮

383
00:13:50,368 --> 00:13:51,708
第三件

384
00:13:51,708 --> 00:13:52,588
保质期

385
00:13:52,588 --> 00:13:56,368
提示词仓库的半衰期很短

386
00:13:56,368 --> 00:13:57,628
模型每次更新

387
00:13:57,628 --> 00:13:59,628
旧的技巧就贬值一批

388
00:13:59,628 --> 00:14:01,968
到13年就退役了

389
00:14:01,968 --> 00:14:05,048
GPT Image 2的技巧能吃多久

390
00:14:05,048 --> 00:14:05,968
没人知道

391
00:14:05,968 --> 00:14:08,748
作者的应对是持续更新

392
00:14:08,748 --> 00:14:10,468
仓库现在还在动

393
00:14:10,468 --> 00:14:12,708
这是它目前最大的护城河

394
00:14:12,708 --> 00:14:14,648
最后给你算笔账

395
00:14:14,648 --> 00:14:16,608
这东西对谁有用

396
00:14:16,608 --> 00:14:19,288
你要是经常做自媒体配图

397
00:14:19,288 --> 00:14:21,388
电商详情页这些需求

398
00:14:21,388 --> 00:14:22,768
或者活动海报

399
00:14:22,768 --> 00:14:24,328
或者做PPT

400
00:14:24,328 --> 00:14:25,568
缺失觉素材

401
00:14:25,568 --> 00:14:26,988
模板拿来填空

402
00:14:26,988 --> 00:14:29,848
出图质量的下限直接抬高一截

403
00:14:29,848 --> 00:14:31,728
何况它免费

404
00:14:31,728 --> 00:14:33,608
你要是在玩Agent

405
00:14:33,608 --> 00:14:36,808
想给自动化流程加一个出图环节

406
00:14:36,808 --> 00:14:40,788
JSON母版加Scale的设计就是给你准备的

407
00:14:40,788 --> 00:14:42,428
同类仓库里

408
00:14:42,428 --> 00:14:44,228
这块它做得最顺手

409
00:14:44,228 --> 00:14:47,088
想学图像Prompt怎么写的人

410
00:14:47,088 --> 00:14:48,588
把它当案例库翻

411
00:14:48,588 --> 00:14:50,608
500多个逆向样本

412
00:14:50,608 --> 00:14:52,928
看多了你自己就会写了

413
00:14:52,928 --> 00:14:54,548
装法很简单

414
00:14:54,548 --> 00:14:56,028
异形命令

415
00:14:56,028 --> 00:15:00,828
NPX Skills at Freestyle Fly-Awesome GPT Image 2

416
00:15:00,828 --> 00:15:04,828
Scale名叫GPT Image 2 Style Library

417
00:15:04,828 --> 00:15:08,288
或者Cloud Code插件市场直接搜

418
00:15:08,288 --> 00:15:11,048
最后说说我的整体判断

419
00:15:11,048 --> 00:15:14,268
这类项目的价值要分两层看

420
00:15:14,268 --> 00:15:17,928
案例和模板这层是当下好用的工具

421
00:15:17,928 --> 00:15:19,788
趁热用别供着

422
00:15:19,788 --> 00:15:21,648
真正值得带走的

423
00:15:21,648 --> 00:15:24,288
是Prompt as Code这个思路本身

424
00:15:24,288 --> 00:15:27,328
把模糊的描述拆成原子参数

425
00:15:27,328 --> 00:15:28,908
散文变成协议

426
00:15:28,908 --> 00:15:31,788
一次性抽卡变成可复用资产

427
00:15:31,788 --> 00:15:35,588
这个思路比任何一份具体模板都长寿

428
00:15:35,588 --> 00:15:37,148
模型换代之后

429
00:15:37,148 --> 00:15:38,328
它依然成立

430
00:15:38,328 --> 00:15:40,548
以后你写任何prompt

431
00:15:40,548 --> 00:15:41,488
图像的

432
00:15:41,488 --> 00:15:42,308
文本的

433
00:15:42,308 --> 00:15:43,848
都可以问自己一句

434
00:15:43,848 --> 00:15:45,828
我写的是一段散文

435
00:15:45,828 --> 00:15:47,128
还是一份协议

436
00:15:47,128 --> 00:15:50,128
我是每天跟AI打交道的人

437
00:15:50,128 --> 00:15:51,808
工具天天换

438
00:15:51,808 --> 00:15:53,408
思路可以一直攒

439
00:15:53,408 --> 00:15:55,088
我们下期见