⏱ 章節
00:00 開場：7 行的檔案，21 萬顆星
00:57 什麼是 Agent Skill：一個資料夾，三層載入
03:57 先給答案：五種類型，六條規律
06:04 規律一：description 決定它會不會被叫出來
09:04 規律二：正文寫短，細節和腳本放外面
12:42 規律三：先寫清楚不做什麼，再堵住捷徑
14:58 行為型 skill 怎麼寫（grill-me、TDD）
20:11 品味型 skill：用角色和反例教審美
21:17 實測統計：38 支知名 skill 量出來的結果
23:24 趨勢：這一年 skill 的寫法怎麼變
24:38 怎麼測一支 skill 到底有沒有用

大家好，我是 Jim，這裡是 AI Notebook。今天這支影片只回答一個問題，一支好的 skill 到底要怎麼設計。你電腦裡的 AI，其實是可以被你教的。教的方法不是打一長串指令，是給它一個資料夾，資料夾裡放一份說明書，AI 覺得這件事用得上，才自己翻開來看。這種資料夾有一個名字，叫 skill。現在全世界最紅的一支 skill，整份檔案只有七行。七行，收到了二十一萬顆星星。那是 GitHub 上的收藏數。同一時間，另一種做法叫 MCP，做同一件事，花掉的字數是它的三十二倍，而字數在這裡是要付錢的。所以問題來了，同樣是教 AI 做事，為什麼有人七行就成，有人寫了五百行還是沒被理睬。為了回答

📎 資料來源
• Agent Skills overview — Claude Docs https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview
• Skill authoring best practices — Claude Docs https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices
• Equipping agents for the real world with Agent Skills — Anthropic Engineering https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills
• anthropics/skills — the official skill repository (pdf, skill-creator, webapp-testing, frontend-design) https://github.com/anthropics/skills
• Agent Skills — the open specification https://agentskills.io
• Simon Willison — Claude Skills are awesome, maybe a bigger deal than MCP https://simonwillison.net/2025/Oct/16/claude-skills/
• Jesse Vincent — Superpowers (announcement post) https://blog.fsck.com/2025/10/09/superpowers/
• obra/superpowers — TDD, systematic-debugging, verification-before-completion https://github.com/obra/superpowers
• mattpocock/skills — grill-me and grilling https://github.com/mattpocock/skills
• kepano/obsidian-skills — obsidian-markdown and obsidian-bases https://github.com/kepano/obsidian-skills
• andrej-karpathy-skills — the behavioural guidelines file https://github.com/multica-ai/andrej-karpathy-skills
• OthmanAdi/planning-with-files — hook-enforced planning skill https://github.com/OthmanAdi/planning-with-files
• SkillsBench — benchmarking agent skills (arXiv) https://arxiv.org/pdf/2606.11435
• Evaluating Agent Skills with MLflow https://mlflow.org/blog/evaluating-skills-mlflow/
• anthropics/skills #1447 — the 2026-07-16 rewrite of the docx / pptx / xlsx skills https://github.com/anthropics/skills/pull/1447
• obra/superpowers 9c9547c (2025-10-16) — 「Now that skills are a first-class thing in Claude Code, restore them to the primary plugin」 https://github.com/obra/superpowers/commit/9c9547cc042d05906d0eb26506f275ff6c82f679

---
Jim AI Notebook | Note the Future
每日深度解析 AI 最新進展

#AI #AISkill #AgentSkills #ClaudeCode #MCP #提示詞工程 #AINotebook

1
00:00:00,000 --> 00:00:06,020
大家好,我是Jim

2
00:00:06,020 --> 00:00:07,260
这里是AI Notebook

3
00:00:07,260 --> 00:00:09,160
今天这支影片只回答一个问题

4
00:00:09,160 --> 00:00:11,280
一支好的SKILL到底要怎么设计

5
00:00:11,280 --> 00:00:13,780
你电脑里的AI其实是可以被你教的

6
00:00:13,780 --> 00:00:15,820
教的方法不是打一长串指令

7
00:00:15,820 --> 00:00:17,760
是给他一个资料夹资料夹里

8
00:00:17,760 --> 00:00:18,860
放一份说明书

9
00:00:18,860 --> 00:00:21,200
AI觉得这件事用得上才自己翻开来

10
00:00:21,200 --> 00:00:23,300
看这种资料夹有一个名字叫SKILL

11
00:00:23,300 --> 00:00:24,960
现在全世界最红的一支

12
00:00:24,960 --> 00:00:25,640
SKILL

13
00:00:25,640 --> 00:00:27,400
整份档案只有七行

14
00:00:27,400 --> 00:00:29,280
七行收到了21万颗星星

15
00:00:29,280 --> 00:00:30,980
那是Github上的收藏

16
00:00:30,980 --> 00:00:33,600
同一时间另一种做法叫MCP

17
00:00:33,600 --> 00:00:34,600
做同一件事

18
00:00:34,600 --> 00:00:36,280
花掉的字数是它的32倍

19
00:00:36,280 --> 00:00:37,920
而字数在这里是要付钱的

20
00:00:37,920 --> 00:00:38,680
所以问题来了

21
00:00:38,680 --> 00:00:40,200
同样是教AI做事

22
00:00:40,200 --> 00:00:41,840
为什么有人七行就成

23
00:00:41,840 --> 00:00:42,960
有人写了500行

24
00:00:42,960 --> 00:00:44,040
还是没被理睬

25
00:00:44,040 --> 00:00:45,140
为了回答这件事

26
00:00:45,140 --> 00:00:46,880
我把十几个最有名的资料夹

27
00:00:46,880 --> 00:00:48,000
一个一个打开来读

28
00:00:48,000 --> 00:00:50,200
还把38支有名的SKILL

29
00:00:50,200 --> 00:00:51,700
全部量了一遍行数

30
00:00:51,700 --> 00:00:52,480
描述长度

31
00:00:52,480 --> 00:00:53,500
附带的程式

32
00:00:53,500 --> 00:00:55,120
一支一支数出来

33
00:00:55,120 --> 00:00:57,280
所以我把它们一条一条整理出来了

34
00:00:57,280 --> 00:00:59,380
我们先看那支七行的档案

35
00:00:59,380 --> 00:01:00,560
它到底长什么样子

36
00:01:00,560 --> 00:01:01,380
第一行写名字

37
00:01:01,380 --> 00:01:02,740
第二行写一句描述

38
00:01:02,740 --> 00:01:04,540
说明这支SKILL是干嘛的

39
00:01:04,540 --> 00:01:06,500
写的是一场毫不留情的访谈

40
00:01:06,500 --> 00:01:07,960
用来把一个计划问清楚

41
00:01:07,960 --> 00:01:09,880
第三行是一行设定

42
00:01:09,880 --> 00:01:11,080
意思是禁止模型

43
00:01:11,080 --> 00:01:12,380
自己教用它

44
00:01:12,380 --> 00:01:13,880
剩下的正文只有一句话

45
00:01:13,880 --> 00:01:14,700
去跑一次拷问

46
00:01:14,700 --> 00:01:15,760
就这样没了

47
00:01:15,760 --> 00:01:16,780
没有流程图

48
00:01:16,780 --> 00:01:17,380
没有范例

49
00:01:17,380 --> 00:01:18,380
没有半行程式

50
00:01:18,380 --> 00:01:20,200
一份七行的纯文字档

51
00:01:20,200 --> 00:01:21,460
被21万个人收藏

52
00:01:21,460 --> 00:01:22,540
而我在同一个星期

53
00:01:22,540 --> 00:01:25,240
打开另一支官方的SKILL光说明书

54
00:01:25,240 --> 00:01:26,740
就有485行

55
00:01:26,740 --> 00:01:28,700
旁边还附了10支小程式

56
00:01:28,700 --> 00:01:30,320
两支都是很成功的作品

57
00:01:30,320 --> 00:01:32,180
这就是这件事最有意思的地方

58
00:01:32,180 --> 00:01:33,100
教AI做事

59
00:01:33,100 --> 00:01:34,280
没有一个标准答案

60
00:01:34,280 --> 00:01:35,280
可是它有规律

61
00:01:35,280 --> 00:01:36,740
这一年多一大批人

62
00:01:36,740 --> 00:01:37,900
在同一个题目上

63
00:01:37,900 --> 00:01:40,260
试错答案都写在公开的档案里

64
00:01:40,260 --> 00:01:41,500
所以我做了两件事

65
00:01:41,500 --> 00:01:43,220
我把最有名的十几支

66
00:01:43,220 --> 00:01:44,460
一支一支打开来读

67
00:01:44,460 --> 00:01:46,960
另外我把38支全部量了一遍

68
00:01:46,960 --> 00:01:48,240
这件事没有那么悬

69
00:01:48,240 --> 00:01:49,180
它其实有几条

70
00:01:49,180 --> 00:01:50,840
可以直接照做的规矩

71
00:01:50,840 --> 00:01:51,880
不过在那之前

72
00:01:51,880 --> 00:01:52,920
我们得先讲清楚

73
00:01:52,920 --> 00:01:54,240
SKILL到底是什么

74
00:01:54,240 --> 00:01:55,600
SKILL这个字

75
00:01:55,600 --> 00:01:57,440
白话讲就是一个资料夹

76
00:01:57,440 --> 00:01:58,720
里面放着你教AI

77
00:01:58,720 --> 00:01:59,760
做一件事的方法

78
00:01:59,760 --> 00:02:01,560
资料夹里一定有一份主档案

79
00:02:01,560 --> 00:02:02,440
那就是说明书

80
00:02:02,440 --> 00:02:04,480
本文旁边可以再放参考文件

81
00:02:04,480 --> 00:02:06,340
也可以放写好的小程式

82
00:02:06,340 --> 00:02:07,580
AI平常不会看它

83
00:02:07,580 --> 00:02:09,340
等到他觉得这件事用得上

84
00:02:09,340 --> 00:02:10,460
才自己翻开来读

85
00:02:10,460 --> 00:02:12,020
官方自己用的比喻

86
00:02:12,020 --> 00:02:14,120
是帮新同事准备的入职手册

87
00:02:14,120 --> 00:02:15,240
新人第一天上班

88
00:02:15,240 --> 00:02:17,660
你不会把整间公司的规章念给他听

89
00:02:17,660 --> 00:02:20,120
你给他一本手册放在柜子上

90
00:02:20,120 --> 00:02:22,140
他要报仗的时候翻报仗那一张

91
00:02:22,140 --> 00:02:24,320
要请假的时候翻请假那一张

92
00:02:24,320 --> 00:02:25,520
SKILL就是那本手册

93
00:02:25,520 --> 00:02:26,960
这样讲你可能会问

94
00:02:26,960 --> 00:02:28,700
那跟我平常直接打字

95
00:02:28,700 --> 00:02:30,060
叫他做事差在哪里

96
00:02:30,060 --> 00:02:30,680
差别是这样

97
00:02:30,680 --> 00:02:32,420
你在对话框里打的指令

98
00:02:32,420 --> 00:02:33,160
讲完就没了

99
00:02:33,160 --> 00:02:34,960
下次还要总打系统提示词

100
00:02:34,960 --> 00:02:36,680
是贴在门口那边规章墙

101
00:02:36,680 --> 00:02:37,600
永远都在

102
00:02:37,600 --> 00:02:39,400
可是贴太多就没有人看

103
00:02:39,400 --> 00:02:41,520
SKILL是柜子里那本作业手册

104
00:02:41,520 --> 00:02:42,840
外面只贴一张标签

105
00:02:42,840 --> 00:02:45,780
用到才翻开MCP是门禁卡跟外线电话

106
00:02:45,780 --> 00:02:48,460
他负责连到外面的系统去把资料抓回来

107
00:02:48,460 --> 00:02:51,040
最后一种是把模型送回学校重读

108
00:02:51,040 --> 00:02:52,020
那最贵也最慢

109
00:02:52,020 --> 00:02:53,400
所以四种方法里面

110
00:02:53,400 --> 00:02:53,720
SKILL

111
00:02:53,720 --> 00:02:55,240
卡在一个很好的位置

112
00:02:55,240 --> 00:02:56,060
他可以带着走

113
00:02:56,060 --> 00:02:57,440
可以存进版本控制

114
00:02:57,440 --> 00:02:58,600
随时想改就改

115
00:02:58,600 --> 00:03:01,400
我们把那本手册分成三层来看

116
00:03:01,400 --> 00:03:02,580
这是整件事的灵魂

117
00:03:02,580 --> 00:03:03,880
先讲Token这个词

118
00:03:03,880 --> 00:03:05,860
Token就是模型的字数配额

119
00:03:05,860 --> 00:03:08,080
你塞给它的每一个字都要算钱

120
00:03:08,080 --> 00:03:09,720
也会占掉它的记忆空间

121
00:03:09,720 --> 00:03:11,800
第一层是每一支SKILL的名牌

122
00:03:11,800 --> 00:03:13,780
名牌上只有名字跟一句描述

123
00:03:13,780 --> 00:03:14,940
大概100个Token

124
00:03:14,940 --> 00:03:17,300
这一层永远都在模型随时看得到

125
00:03:17,300 --> 00:03:18,780
就像柜子上那一排标签

126
00:03:18,780 --> 00:03:20,560
第二层是说明书本文

127
00:03:20,560 --> 00:03:22,540
只有当模型判断这件事用得上

128
00:03:22,540 --> 00:03:24,380
它才把整份说明书书读进来

129
00:03:24,380 --> 00:03:26,860
官方建议控制在5000个Token以内

130
00:03:26,860 --> 00:03:29,160
第三层是资料夹里其他的东西

131
00:03:29,160 --> 00:03:30,700
参考文件跟小程式

132
00:03:30,700 --> 00:03:32,240
用不到就是零

133
00:03:32,240 --> 00:03:33,840
这一层里面最妙的是程式

134
00:03:33,840 --> 00:03:35,120
程式是拿去执行的

135
00:03:35,120 --> 00:03:36,500
不是拿去读的模型

136
00:03:36,500 --> 00:03:38,360
不会把几百行程式读进脑袋

137
00:03:38,360 --> 00:03:39,860
它只是跑一次看结果

138
00:03:39,860 --> 00:03:41,900
所以你可以在一支SKILL里

139
00:03:41,900 --> 00:03:42,780
放一整套工具

140
00:03:42,780 --> 00:03:44,040
成本几乎是零

141
00:03:44,040 --> 00:03:45,760
这三层有一个正式名称叫

142
00:03:45,760 --> 00:03:46,600
渐进式揭露

143
00:03:46,600 --> 00:03:48,220
翻成白话就是先看标签

144
00:03:48,220 --> 00:03:49,040
需要才翻书

145
00:03:49,040 --> 00:03:50,240
真的动手才拿工具

146
00:03:50,240 --> 00:03:52,020
这三层是后面每一条

147
00:03:52,020 --> 00:03:54,040
规则的地基底值

148
00:03:54,040 --> 00:03:55,420
讲到这里我们看第一件

149
00:03:55,420 --> 00:03:56,540
真正要决定的是

150
00:03:56,540 --> 00:03:58,480
你要写一支SKILL之前

151
00:03:58,480 --> 00:03:59,560
先问自己一个问题

152
00:03:59,560 --> 00:04:01,240
你要改的是模型的哪一层

153
00:04:01,240 --> 00:04:02,620
这个问题最关键

154
00:04:02,620 --> 00:04:03,900
因为答案不一样

155
00:04:03,900 --> 00:04:06,000
你的档案就长得完全不一样

156
00:04:06,000 --> 00:04:07,140
答案总共有五种

157
00:04:07,140 --> 00:04:09,020
你要补的是他不知道的东西

158
00:04:09,020 --> 00:04:10,040
这叫知识型

159
00:04:10,040 --> 00:04:12,360
一款笔记软体公开的那五支

160
00:04:12,360 --> 00:04:13,520
就是这样把模型

161
00:04:13,520 --> 00:04:15,280
没学过的规格塞进他脑袋

162
00:04:15,280 --> 00:04:16,980
你要挡的是他的坏习惯

163
00:04:16,980 --> 00:04:17,880
这叫禁令型

164
00:04:17,880 --> 00:04:19,460
有一位很有名的研究者

165
00:04:19,460 --> 00:04:20,360
整理的那四条就是

166
00:04:20,360 --> 00:04:21,620
他不给新知识

167
00:04:21,620 --> 00:04:23,720
他用否定句在模型四周立强

168
00:04:23,720 --> 00:04:25,520
你要定的是先后顺序

169
00:04:25,520 --> 00:04:26,460
这叫流程型

170
00:04:26,460 --> 00:04:27,700
开头那支拷问就是

171
00:04:27,700 --> 00:04:29,300
他规定先做什么再做什么

172
00:04:29,300 --> 00:04:31,020
用状态决定什么时候结束

173
00:04:31,020 --> 00:04:32,220
你要调的是审美

174
00:04:32,220 --> 00:04:33,180
这叫品位型

175
00:04:33,180 --> 00:04:34,920
官方那支管界面设计的

176
00:04:34,920 --> 00:04:36,260
就是他给角色

177
00:04:36,260 --> 00:04:37,520
给反例不给检查表

178
00:04:37,520 --> 00:04:39,760
你要防的是他偷懒跟说谎

179
00:04:39,760 --> 00:04:40,520
这叫纪律型

180
00:04:40,520 --> 00:04:42,020
那一整套工作纪律就是

181
00:04:42,020 --> 00:04:43,020
他预先列出

182
00:04:43,020 --> 00:04:44,460
你会找的每一个借口

183
00:04:44,460 --> 00:04:45,420
一条一条挡掉

184
00:04:45,420 --> 00:04:47,120
这五种写法完全不一样

185
00:04:47,120 --> 00:04:48,640
可是他们都算好作品

186
00:04:48,640 --> 00:04:50,180
因为他们动的层不同

187
00:04:50,180 --> 00:04:52,240
所以这张图不是分类游戏

188
00:04:52,240 --> 00:04:54,380
它是你动手前的第一个岔路口

189
00:04:54,380 --> 00:04:55,940
你选错陈后面写

190
00:04:55,940 --> 00:04:57,280
再多都是白费力气

191
00:04:57,280 --> 00:04:58,280
陈决定的篇幅

192
00:04:58,280 --> 00:04:59,200
也决定的语气

193
00:04:59,200 --> 00:04:59,980
再见

194
00:04:59,980 --> 00:05:00,000
再见

195
00:05:00,000 --> 00:05:02,300
好了 接下來是共通的部分

196
00:05:02,300 --> 00:05:04,360
我讀完十幾支最有名的作品

197
00:05:04,360 --> 00:05:05,700
整理出六條鐵律

198
00:05:05,700 --> 00:05:07,160
不管你改的是哪一層

199
00:05:07,160 --> 00:05:08,140
這六條都成立

200
00:05:08,140 --> 00:05:09,400
我現在全部給你後面

201
00:05:09,400 --> 00:05:11,140
再一條一條帶你看證據

202
00:05:11,140 --> 00:05:13,000
第一條那句描述是一支掉溝

203
00:05:13,000 --> 00:05:14,340
不是一段摘要

204
00:05:14,340 --> 00:05:16,360
你不是在寫這支skill做了什麼

205
00:05:16,360 --> 00:05:17,800
你是在寫使用者講出

206
00:05:17,800 --> 00:05:19,700
哪一種話的時候應該來找我

207
00:05:19,700 --> 00:05:20,240
第二條

208
00:05:20,240 --> 00:05:22,300
本文只放地圖說明書的本體

209
00:05:22,300 --> 00:05:23,200
越短越好

210
00:05:23,200 --> 00:05:25,040
細節推到旁邊的參考檔

211
00:05:25,040 --> 00:05:26,800
模型真的需要的時候才去翻

212
00:05:26,800 --> 00:05:27,700
第三條

213
00:05:27,700 --> 00:05:29,000
確定的事情交給腳本

214
00:05:29,000 --> 00:05:29,760
步驟固定

215
00:05:29,760 --> 00:05:30,820
答案只有一個的是

216
00:05:30,820 --> 00:05:32,700
你寫成一支小城市讓他去跑

217
00:05:32,700 --> 00:05:34,520
不要用文字拜託他自己算

218
00:05:34,520 --> 00:05:35,120
第四條

219
00:05:35,120 --> 00:05:36,120
解釋為什麼

220
00:05:36,120 --> 00:05:37,520
勝過用大寫壓人

221
00:05:37,520 --> 00:05:40,120
你告訴模型這件事做錯會發生什麼

222
00:05:40,120 --> 00:05:42,420
比連寫30次不准有效的多

223
00:05:42,420 --> 00:05:43,000
第五條

224
00:05:43,000 --> 00:05:44,320
模型會抄得捷徑

225
00:05:44,320 --> 00:05:45,660
你要先寫使堵住

226
00:05:45,660 --> 00:05:47,260
你不要只寫正確做法

227
00:05:47,260 --> 00:05:49,500
你要把它會找的每個藉口先寫出來

228
00:05:49,500 --> 00:05:50,120
第六條

229
00:05:50,120 --> 00:05:51,160
寫完一定要測

230
00:05:51,160 --> 00:05:52,660
你先讓模型裸著做一次

231
00:05:52,660 --> 00:05:53,760
看它卡在哪裡

232
00:05:53,760 --> 00:05:55,260
再回頭改你的檔案

233
00:05:55,260 --> 00:05:56,060
這六條

234
00:05:56,060 --> 00:05:58,200
就像考卷的答案先寫在黑板上

235
00:05:58,200 --> 00:05:59,760
接下來我們一條一條回去

236
00:05:59,760 --> 00:06:01,520
看真正做得好的那些人

237
00:06:01,520 --> 00:06:03,820
是怎麼把這幾條寫進檔案裡的

238
00:06:03,820 --> 00:06:04,920
第一條

239
00:06:04,920 --> 00:06:06,520
鐵律說描述是掉鉤

240
00:06:06,520 --> 00:06:08,700
我們就從官方最早那批裡面

241
00:06:08,700 --> 00:06:10,120
最經典的一支開始

242
00:06:10,120 --> 00:06:12,120
看處理PDF檔的那一支

243
00:06:12,120 --> 00:06:13,160
你打開他的說明書

244
00:06:13,160 --> 00:06:14,120
第一眼要看的不是本文

245
00:06:14,120 --> 00:06:15,420
是檔案最上面那幾行

246
00:06:15,420 --> 00:06:16,300
那幾行是

247
00:06:16,300 --> 00:06:17,600
這支SKILL的名牌

248
00:06:17,600 --> 00:06:19,400
白話就是檔案開頭的名牌

249
00:06:19,400 --> 00:06:21,160
你想像一下模型的處境

250
00:06:21,160 --> 00:06:22,840
它面前有100張標籤

251
00:06:22,840 --> 00:06:23,740
你問了一句話

252
00:06:23,740 --> 00:06:26,440
它要在0點幾秒內決定翻哪一本

253
00:06:26,440 --> 00:06:29,100
它唯一能看的就是每一張標籤上的那一句

254
00:06:29,100 --> 00:06:30,900
所以社群後來有一個共識

255
00:06:30,900 --> 00:06:32,160
描述才是班機

256
00:06:32,160 --> 00:06:32,900
本文不是

257
00:06:32,900 --> 00:06:34,160
你本文寫的再漂亮

258
00:06:34,160 --> 00:06:34,840
描述模糊

259
00:06:34,840 --> 00:06:36,700
這支SKILL一輩子不會被翻開

260
00:06:36,700 --> 00:06:38,340
那官方這一支怎麼寫

261
00:06:38,340 --> 00:06:39,840
它把那一欄分成三段

262
00:06:39,840 --> 00:06:40,980
開頭是一句命令

263
00:06:40,980 --> 00:06:43,380
只要使用者想對PDF檔做任何事

264
00:06:43,380 --> 00:06:45,540
就用這支SKILL範圍開到最大

265
00:06:45,540 --> 00:06:46,980
接著是一長串動詞

266
00:06:46,980 --> 00:06:47,700
讀取合併

267
00:06:47,700 --> 00:06:48,500
切開旋轉

268
00:06:48,500 --> 00:06:49,340
填表單

269
00:06:49,340 --> 00:06:50,740
把使用者可能會講的詞

270
00:06:50,740 --> 00:06:52,200
全部鋪進去

271
00:06:52,200 --> 00:06:53,300
最後補一句兜頂

272
00:06:53,300 --> 00:06:54,940
只要使用者提到一個PDF檔

273
00:06:54,940 --> 00:06:56,300
或者要你做一個出來

274
00:06:56,300 --> 00:06:56,840
就用這一支

275
00:06:56,840 --> 00:06:58,640
你看得出來這三段的方向

276
00:06:58,640 --> 00:07:00,740
它不是在介紹自己

277
00:07:00,740 --> 00:07:02,080
它是在猜你會怎麼開

278
00:07:02,680 --> 00:07:03,580
名牌

279
00:07:03,580 --> 00:07:05,580
講完我們翻進同一支的本文

280
00:07:05,580 --> 00:07:07,900
這裡剛好看得到第四條鐵律

281
00:07:07,900 --> 00:07:09,440
這一支的本文有300多行

282
00:07:09,440 --> 00:07:10,600
可是它的文字非常少

283
00:07:10,600 --> 00:07:11,900
因為裡面塞了17段

284
00:07:11,900 --> 00:07:13,640
可以直接執行的程式碼

285
00:07:13,640 --> 00:07:15,640
它的第一句話先講自己是誰

286
00:07:15,640 --> 00:07:17,200
第二句就開始分錄

287
00:07:17,200 --> 00:07:19,280
要填表單的去看表單那一份

288
00:07:19,280 --> 00:07:21,840
要進階用法的去看參考那一份

289
00:07:21,840 --> 00:07:23,300
等於這本手冊的第一頁就是

290
00:07:23,300 --> 00:07:24,840
目錄不是長篇大論

291
00:07:24,840 --> 00:07:26,440
我最喜歡的是它的客製

292
00:07:26,440 --> 00:07:27,900
整份文件裡大寫的禁止

293
00:07:27,900 --> 00:07:29,100
只出現兩次

294
00:07:29,100 --> 00:07:31,280
而且那兩次講的是同一件事

295
00:07:31,280 --> 00:07:31,980
他說

296
00:07:31,980 --> 00:07:33,740
永遠不要在PDF裡使用那種

297
00:07:33,740 --> 00:07:35,340
上標下標的特殊符號

298
00:07:35,340 --> 00:07:36,280
為什麼

299
00:07:36,280 --> 00:07:37,800
因為那些字元印出來會變成

300
00:07:37,800 --> 00:07:38,780
一整塊黑色的方塊

301
00:07:38,780 --> 00:07:39,840
整份文件就毀了

302
00:07:39,840 --> 00:07:42,120
而且你在螢幕上看不出來

303
00:07:42,120 --> 00:07:42,940
你注意它的寫法

304
00:07:42,940 --> 00:07:44,280
它不是只寫一句不準

305
00:07:44,280 --> 00:07:45,380
它把後果寫在後面

306
00:07:45,380 --> 00:07:47,620
這就是那條鐵律的長相

307
00:07:47,620 --> 00:07:49,780
大寫的命令是稀缺資源

308
00:07:49,780 --> 00:07:52,020
同一份文件裡寫30次不準模型

309
00:07:52,020 --> 00:07:53,040
就當你是背景音樂

310
00:07:53,040 --> 00:07:54,320
你只寫一次

311
00:07:54,320 --> 00:07:57,040
而且那一次是真的會出事的一壘

312
00:07:57,040 --> 00:07:57,740
它才記得

313
00:07:57,740 --> 00:07:59,720
描述要強

314
00:07:59,720 --> 00:08:00,740
正文要軟

315
00:08:00,740 --> 00:08:02,520
這兩件事官方自己有一支東西

316
00:08:02,520 --> 00:08:03,320
同時示範了

317
00:08:03,320 --> 00:08:05,680
它放了一支專門用來寫Skill的

318
00:08:05,680 --> 00:08:07,340
Skill你可以叫它原技能

319
00:08:07,340 --> 00:08:10,400
它裡面有兩句話我覺得是整個題目最重要的兩句

320
00:08:10,400 --> 00:08:11,180
第一句是

321
00:08:11,180 --> 00:08:13,940
請把Skill的描述寫的稍微強勢一點

322
00:08:13,940 --> 00:08:14,800
為什麼要強勢

323
00:08:14,800 --> 00:08:17,780
因為官方觀察到模型天生有一個傾向

324
00:08:17,780 --> 00:08:20,940
他會覺得這件事我自己好像也會就不去翻手冊了

325
00:08:20,940 --> 00:08:22,880
這個傾向叫做觸發不足

326
00:08:22,880 --> 00:08:24,400
所以描述要主動一點

327
00:08:24,400 --> 00:08:27,340
要像一個站在門口說這件事交給我的人

328
00:08:27,340 --> 00:08:28,440
第二句剛好相反

329
00:08:28,440 --> 00:08:32,140
他說如果你發現自己一直在寫全大寫的永遠要跟絕對不準

330
00:08:32,140 --> 00:08:33,280
那是一支黃旗

331
00:08:33,280 --> 00:08:34,380
黃旗的意思是

332
00:08:34,380 --> 00:08:35,700
你先停下來想一下

333
00:08:35,700 --> 00:08:36,800
他接著解釋

334
00:08:36,800 --> 00:08:38,340
今天的模型很聰明

335
00:08:38,340 --> 00:08:39,780
他會揣摩你的意思

336
00:08:39,780 --> 00:08:40,900
與其大聲命令

337
00:08:40,900 --> 00:08:42,600
不如告訴他為什麼這件事重要

338
00:08:42,600 --> 00:08:45,040
這兩句放在一起就是這門手藝的性格

339
00:08:45,040 --> 00:08:46,200
對外要強勢

340
00:08:46,200 --> 00:08:48,200
讓人找得到你對那要講道理

341
00:08:48,200 --> 00:08:49,200
不要用大寫壓人

342
00:08:49,200 --> 00:08:51,200
他還順便示範了最後一條鐵律

343
00:08:51,200 --> 00:08:53,700
他教的方法是你開兩個AI

344
00:08:53,700 --> 00:08:55,800
一個負責寫這支Skill

345
00:08:55,800 --> 00:08:57,140
另一個完全不知情

346
00:08:57,140 --> 00:08:58,140
直接拿去用

347
00:08:58,140 --> 00:08:59,500
你看第二個卡在哪裡

348
00:08:59,500 --> 00:09:01,860
回頭改改的依據是他真的做了什麼

349
00:09:01,860 --> 00:09:04,260
不是你以為他會怎麼做

350
00:09:04,260 --> 00:09:05,500
第二條鐵律說

351
00:09:05,500 --> 00:09:07,040
董倫指放地圖第三條說

352
00:09:07,040 --> 00:09:09,200
能寫成程式的就不要用嘴巴教

353
00:09:09,200 --> 00:09:10,900
這兩條聽起來像潔癖

354
00:09:10,900 --> 00:09:12,800
其實他們是被錢逼出來的

355
00:09:12,800 --> 00:09:15,760
我們把時間拉回2025年的10月你就懂了

356
00:09:15,760 --> 00:09:17,360
那時候大家再接MCP

357
00:09:17,360 --> 00:09:20,260
MCP你可以想成AI的門禁卡跟外線電話

358
00:09:20,260 --> 00:09:22,140
他讓模型連得到你的信箱

359
00:09:22,140 --> 00:09:22,960
你的資料庫

360
00:09:22,960 --> 00:09:24,160
你的專案很好用

361
00:09:24,160 --> 00:09:25,300
可是有一個副作用

362
00:09:25,300 --> 00:09:25,960
每接一套

363
00:09:25,960 --> 00:09:28,400
他就要先把這套系統能做的每件事

364
00:09:28,400 --> 00:09:30,040
完整寫給模型看一遍

365
00:09:30,040 --> 00:09:30,760
接五套

366
00:09:30,760 --> 00:09:32,440
你還沒開始交代工作

367
00:09:32,440 --> 00:09:34,700
模型的記憶就先被塞掉一大半

368
00:09:34,700 --> 00:09:36,940
後來有人拿同一個任務實測

369
00:09:36,940 --> 00:09:38,800
同樣是去Github上做事

370
00:09:38,800 --> 00:09:39,700
跑75次

371
00:09:39,700 --> 00:09:41,760
一邊走MCP邊走命令列

372
00:09:41,760 --> 00:09:42,700
加上一支Skill

373
00:09:42,700 --> 00:09:45,560
走MCP那一邊花掉的字數是另一邊的32倍

374
00:09:45,560 --> 00:09:47,700
32倍不是慢一點的差別是

375
00:09:47,700 --> 00:09:49,460
同一件是你付了32次的錢

376
00:09:49,460 --> 00:09:50,540
就在這個氣氛裡

377
00:09:50,540 --> 00:09:52,840
10月16號官方一次丟出4支

378
00:09:52,840 --> 00:09:55,940
Skill教模型處理最無聊的4種檔案PDF

379
00:09:55,940 --> 00:09:57,040
文書檔簡報

380
00:09:57,040 --> 00:09:58,680
試算表他們沒有協定

381
00:09:58,680 --> 00:09:59,800
沒有伺服器

382
00:09:59,800 --> 00:10:00,040
就是

383
00:10:00,000 --> 00:10:01,400
是四个资料夹

384
00:10:01,400 --> 00:10:03,640
你现在知道为什么所有人都在比谁写的短了

385
00:10:03,640 --> 00:10:04,680
指数

386
00:10:04,680 --> 00:10:05,760
这件事被看见之后

387
00:10:05,760 --> 00:10:08,200
同一天有一位很资深的独立开发者

388
00:10:08,200 --> 00:10:10,800
写了一篇短文把整个气氛点燃了

389
00:10:10,800 --> 00:10:12,400
标题是Claude的Skill

390
00:10:12,400 --> 00:10:14,360
很厉害可能比MTP更重要

391
00:10:14,360 --> 00:10:15,920
他的理由不是效能是简单

392
00:10:15,920 --> 00:10:17,800
他说MCP是一整套协定

393
00:10:17,800 --> 00:10:19,080
你要跑伺服器要处理

394
00:10:19,080 --> 00:10:20,880
连线Skill只是一份纯文字档

395
00:10:20,880 --> 00:10:21,960
加上开头几行设定

396
00:10:21,960 --> 00:10:22,480
如此而已

397
00:10:22,480 --> 00:10:24,160
他讲了一句我很喜欢的话

398
00:10:24,160 --> 00:10:26,240
他说这个做法更贴近语言模型

399
00:10:26,240 --> 00:10:28,680
本来的精神就是丢一段文字进去

400
00:10:28,680 --> 00:10:30,080
让模型自己想办法

401
00:10:30,080 --> 00:10:32,680
两个月之后这句话应验了

402
00:10:32,680 --> 00:10:33,960
12月18号

403
00:10:33,960 --> 00:10:35,960
这套格式被公开成一份规格

404
00:10:35,960 --> 00:10:37,580
捐给了一个中立的基金会

405
00:10:37,580 --> 00:10:38,780
48小时之内

406
00:10:38,780 --> 00:10:41,440
微软把它接近自己的程式编辑器

407
00:10:41,440 --> 00:10:42,100
OpenAI

408
00:10:42,100 --> 00:10:44,520
也在自家产品里用上一个格式

409
00:10:44,520 --> 00:10:45,980
简单到所有人都懒得吵

410
00:10:45,980 --> 00:10:47,860
它就变成大家的共同语言

411
00:10:47,860 --> 00:10:50,460
这件事对你写档案有一个很直接的含义

412
00:10:50,460 --> 00:10:51,900
这个格式从头到尾

413
00:10:51,900 --> 00:10:53,480
没有给你任何华丽的功能

414
00:10:53,480 --> 00:10:54,840
它只给你一份纯文字

415
00:10:54,840 --> 00:10:57,320
所以你能赢的地方从来不是技术

416
00:10:57,320 --> 00:10:59,180
是你怎么把话写清楚

417
00:10:59,180 --> 00:11:00,700
把话写清楚

418
00:11:00,700 --> 00:11:02,080
可以清楚到什么程度

419
00:11:02,080 --> 00:11:03,600
官方同一个资料夹里

420
00:11:03,600 --> 00:11:04,700
有一只走到了极端

421
00:11:04,700 --> 00:11:05,900
它负责测试网页

422
00:11:05,900 --> 00:11:07,900
说明书只有95行

423
00:11:07,900 --> 00:11:09,540
它第一句话没有任何寒暄

424
00:11:09,540 --> 00:11:11,500
直接说要测试本机的网页

425
00:11:11,500 --> 00:11:13,420
就写一只原生的测试脚本

426
00:11:13,420 --> 00:11:15,460
它真正的灵魂在第14行

427
00:11:15,460 --> 00:11:17,140
它说先跑一次说明指令

428
00:11:17,140 --> 00:11:18,680
不要去读脚本的原始码

429
00:11:18,680 --> 00:11:19,960
理由写在后面

430
00:11:19,960 --> 00:11:20,960
这些脚本很大

431
00:11:20,960 --> 00:11:22,860
读进来会占掉你的记忆空间

432
00:11:22,860 --> 00:11:23,720
它们的存在

433
00:11:23,720 --> 00:11:25,140
就是要被当成黑盒子

434
00:11:25,140 --> 00:11:25,760
直接呼叫

435
00:11:25,760 --> 00:11:26,920
不是被读进去的

436
00:11:27,320 --> 00:11:28,280
你看这句话等于

437
00:11:28,280 --> 00:11:30,060
把刚刚那个成本观念

438
00:11:30,060 --> 00:11:32,200
直接写进了给模型的指令里

439
00:11:32,200 --> 00:11:33,100
它甚至没有用

440
00:11:33,100 --> 00:11:34,320
长篇文字说明流程

441
00:11:34,320 --> 00:11:35,320
它画了一颗决策树

442
00:11:35,320 --> 00:11:36,800
用纯文字画的网页

443
00:11:36,800 --> 00:11:37,860
有没有跑起来

444
00:11:37,860 --> 00:11:38,760
有就往这边

445
00:11:38,760 --> 00:11:39,860
没有就往那边

446
00:11:39,860 --> 00:11:41,700
每一个分支都给一条退路

447
00:11:41,700 --> 00:11:42,480
95行

448
00:11:42,480 --> 00:11:43,240
一棵树

449
00:11:43,240 --> 00:11:43,940
一条铁律

450
00:11:43,940 --> 00:11:46,480
这一只后来常常被拿来当饭本

451
00:11:46,480 --> 00:11:48,380
因为它同时证明了两条说明书

452
00:11:48,380 --> 00:11:49,500
短不代表交得少

453
00:11:49,500 --> 00:11:51,660
难的部分你可以封进城市里

454
00:11:51,660 --> 00:11:53,880
把功夫封进城市这条路

455
00:11:53,880 --> 00:11:55,400
社群有一只玩得更远

456
00:11:55,400 --> 00:11:57,340
它接着是一个AI笔记工具

457
00:11:57,340 --> 00:11:58,160
你可以叫它

458
00:11:58,160 --> 00:12:00,040
自动去读一整堆资料

459
00:12:00,040 --> 00:12:00,940
再回答这只

460
00:12:00,940 --> 00:12:01,620
skill

461
00:12:01,620 --> 00:12:03,160
做了一件我第一次看到的事

462
00:12:03,160 --> 00:12:04,640
它的脚本每次回答完

463
00:12:04,640 --> 00:12:06,300
都会在结尾自动加上一句话

464
00:12:06,300 --> 00:12:08,240
这句话不是给你看的

465
00:12:08,240 --> 00:12:09,380
是给模型看的写着

466
00:12:09,380 --> 00:12:10,040
这样就是

467
00:12:10,040 --> 00:12:11,320
你需要知道的

468
00:12:11,320 --> 00:12:12,120
全部了吗

469
00:12:12,120 --> 00:12:13,220
然后说明书里规定

470
00:12:13,220 --> 00:12:14,560
模型只要看到这句话

471
00:12:14,560 --> 00:12:15,520
就必须停下来

472
00:12:15,520 --> 00:12:17,140
重新检查自己还缺什么

473
00:12:17,140 --> 00:12:18,360
缺就再问一次

474
00:12:18,360 --> 00:12:19,600
问到其为止

475
00:12:19,600 --> 00:12:21,020
你想想这件事的结构

476
00:12:21,020 --> 00:12:22,280
一般的skill

477
00:12:22,280 --> 00:12:24,040
只有在被翻开的那一刻说话

478
00:12:24,040 --> 00:12:24,920
说完就闭嘴

479
00:12:24,920 --> 00:12:26,200
这一只不是

480
00:12:26,200 --> 00:12:27,400
它让工具在每一次

481
00:12:27,400 --> 00:12:28,360
吐出结果的时候

482
00:12:28,360 --> 00:12:30,040
继续对模型下指令

483
00:12:30,040 --> 00:12:31,160
等于在流程中间

484
00:12:31,160 --> 00:12:32,760
开了第二条指令通道

485
00:12:32,760 --> 00:12:34,120
这也提醒我们一件事

486
00:12:34,120 --> 00:12:36,200
你的skill不只有那份说明书

487
00:12:36,200 --> 00:12:37,760
任何模型会看到的东西

488
00:12:37,760 --> 00:12:38,920
包括程式的输出

489
00:12:38,920 --> 00:12:39,800
错误讯息

490
00:12:39,800 --> 00:12:41,080
甚至档案名称

491
00:12:41,080 --> 00:12:43,080
都是你在对它说话

492
00:12:43,080 --> 00:12:44,440
本文要瘦下来最难的

493
00:12:44,440 --> 00:12:45,120
不是三字

494
00:12:45,120 --> 00:12:46,280
是决定什么不写

495
00:12:46,280 --> 00:12:48,040
这件事有一个人做的特别好

496
00:12:48,040 --> 00:12:49,440
而且他不是工程师

497
00:12:49,440 --> 00:12:52,080
有一款很多人在用的笔记软体

498
00:12:52,080 --> 00:12:54,320
他的创办人把自己私下在用的skill

499
00:12:54,320 --> 00:12:55,040
直接公开

500
00:12:55,040 --> 00:12:56,040
五个纯文字档

501
00:12:56,040 --> 00:12:57,360
收到四万颗星

502
00:12:57,360 --> 00:12:58,880
我把其中一只打开来读

503
00:12:58,880 --> 00:13:00,240
他管的是笔记的写法

504
00:13:00,240 --> 00:13:01,760
他的第一句话就把我说服了

505
00:13:01,760 --> 00:13:03,040
他说这一只只涵盖

506
00:13:03,040 --> 00:13:04,800
我们这家特有的写法

507
00:13:04,800 --> 00:13:06,160
标准的部分

508
00:13:06,160 --> 00:13:07,040
视为模型

509
00:13:07,040 --> 00:13:07,680
已经会了

510
00:13:07,680 --> 00:13:09,280
一句话省掉一半篇幅

511
00:13:09,280 --> 00:13:11,120
这是很多人写说明书的时候

512
00:13:11,120 --> 00:13:12,320
不敢做的事

513
00:13:12,320 --> 00:13:13,040
你会怕漏

514
00:13:13,040 --> 00:13:14,040
所以什么都写

515
00:13:14,040 --> 00:13:15,440
写到最后又臭又长

516
00:13:15,440 --> 00:13:17,040
重点被埋在中间

517
00:13:17,040 --> 00:13:19,280
他反过来先讲清楚我不教什么

518
00:13:19,280 --> 00:13:21,840
剩下的空间才装得下真正要教的东西

519
00:13:21,840 --> 00:13:23,600
同一组里面还有一只更聪明

520
00:13:23,600 --> 00:13:26,160
他管的是一种特殊的档案格式

521
00:13:26,160 --> 00:13:27,880
所以他的触发条件不写情境

522
00:13:27,880 --> 00:13:29,120
直接写副档名

523
00:13:29,120 --> 00:13:30,960
只要碰到这种副档名的档案

524
00:13:30,960 --> 00:13:31,600
就用

525
00:13:31,600 --> 00:13:32,960
这叫硬触发

526
00:13:32,960 --> 00:13:34,480
比讲情境更不会误判

527
00:13:34,480 --> 00:13:36,400
这也是掉勾那一条的另一种写法

528
00:13:36,400 --> 00:13:38,160
而且他的流程第五步是验证

529
00:13:38,160 --> 00:13:40,880
把最容易写错的三种情况直接列出来

530
00:13:40,880 --> 00:13:42,720
叫模型自己回头检查

531
00:13:42,720 --> 00:13:45,520
你把最常踩的那三个坑写进流程里

532
00:13:45,520 --> 00:13:48,320
就等于替他先挡掉一半的错

533
00:13:48,320 --> 00:13:50,400
替模型先挡掉错误这件事

534
00:13:50,400 --> 00:13:52,960
2026年1月有一个更极端的版本

535
00:13:52,960 --> 00:13:55,440
卡帕西就是那位很有名的AI研究者

536
00:13:55,440 --> 00:13:58,640
在网路上抱怨AI帮他写程式的几个毛病

537
00:13:58,640 --> 00:13:59,920
他说他不确认前提

538
00:13:59,920 --> 00:14:02,000
就直接动手50行能解决的事

539
00:14:02,000 --> 00:14:03,200
他写500行

540
00:14:03,200 --> 00:14:05,680
他还会顺手改掉你根本没有叫他改的地方

541
00:14:05,680 --> 00:14:07,360
这些话没有一句是新的

542
00:14:07,360 --> 00:14:08,480
特别的是隔天

543
00:14:08,480 --> 00:14:11,360
隔天就有人把这几句话整理成一个档案

544
00:14:11,360 --> 00:14:12,320
67行

545
00:14:12,320 --> 00:14:14,000
4条原则放上网

546
00:14:14,000 --> 00:14:15,920
后来这个档案累积到20万颗星

547
00:14:15,920 --> 00:14:16,960
他凭什么

548
00:14:16,960 --> 00:14:18,320
因为他的写法很准

549
00:14:18,320 --> 00:14:20,160
每一条原则的主体不是教

550
00:14:20,160 --> 00:14:20,960
你要做什么

551
00:14:20,960 --> 00:14:22,320
是一串不要做什么

552
00:14:22,320 --> 00:14:22,960
不要假设

553
00:14:22,960 --> 00:14:24,080
不要藏起你的困惑

554
00:14:24,080 --> 00:14:25,840
不要偷偷扩大范围行为

555
00:14:25,840 --> 00:14:27,680
这种东西用否定句写笔

556
00:14:27,680 --> 00:14:29,120
用正面句写有效

557
00:14:29,120 --> 00:14:31,040
因为模型不要知道界限在哪里

558
00:14:31,040 --> 00:14:31,920
这就是捷径

559
00:14:31,920 --> 00:14:33,440
那一条铁绿的圆形

560
00:14:33,440 --> 00:14:36,160
你要先把它会走的每一条歪路写死

561
00:14:36,160 --> 00:14:39,200
更妙的是每一条后面都挂一个自我测试的问题

562
00:14:39,200 --> 00:14:40,160
你写了200行

563
00:14:40,160 --> 00:14:41,680
如果50行就能写完

564
00:14:41,680 --> 00:14:42,480
那就重写

565
00:14:42,480 --> 00:14:43,520
你改了每一行

566
00:14:43,520 --> 00:14:45,760
能不能直接对回使用者的要求

567
00:14:45,760 --> 00:14:47,440
形容词是没办法执行的

568
00:14:47,440 --> 00:14:48,400
你叫他简洁

569
00:14:48,400 --> 00:14:49,520
他不知道什么叫简洁

570
00:14:49,520 --> 00:14:51,640
你给他一个可以自己回答的问题

571
00:14:51,640 --> 00:14:52,960
他就能自己检查

572
00:14:52,960 --> 00:14:54,480
这一只开头还很诚实

573
00:14:54,480 --> 00:14:58,160
他先说明这套原则是拿速度换谨慎

574
00:14:58,160 --> 00:14:58,800
到这里为止

575
00:14:58,800 --> 00:15:00,000
我们看的都是交工具

576
00:15:00,000 --> 00:15:00,760
Skill

577
00:15:00,760 --> 00:15:01,920
接下來這一派更難

578
00:15:01,920 --> 00:15:03,580
他們要改的不是模型

579
00:15:03,580 --> 00:15:05,640
會什麼是模型的習慣

580
00:15:05,640 --> 00:15:06,920
我們要回到影片開頭

581
00:15:06,920 --> 00:15:08,700
那七航這隻Skill的作者

582
00:15:08,700 --> 00:15:10,680
是一位很有名的前端工程師

583
00:15:10,680 --> 00:15:11,700
他做的東西叫

584
00:15:11,700 --> 00:15:12,380
拷問我

585
00:15:12,380 --> 00:15:13,200
你打一個指令

586
00:15:13,200 --> 00:15:14,840
他就開始一輪一輪的問你

587
00:15:14,840 --> 00:15:15,560
把你的計畫

588
00:15:15,560 --> 00:15:17,660
問到你自己都想清楚為止

589
00:15:17,660 --> 00:15:19,460
這就是21萬顆星那一隻

590
00:15:19,460 --> 00:15:21,500
他最反直覺的地方在哪裡

591
00:15:21,500 --> 00:15:24,200
這一年所有工具都在讓AI動作更快

592
00:15:24,200 --> 00:15:25,160
他偏偏反過來

593
00:15:25,160 --> 00:15:27,640
在你動工之前先加一道磨茶

594
00:15:27,640 --> 00:15:29,180
他的結構是兩層

595
00:15:29,180 --> 00:15:31,480
前面那七航是門上寫著一航設定

596
00:15:31,480 --> 00:15:33,020
禁止模型自己叫用

597
00:15:33,020 --> 00:15:35,840
也就是說AI不可以自己決定要來質問你

598
00:15:35,840 --> 00:15:37,880
只有你親手打那個指令

599
00:15:37,880 --> 00:15:39,160
這扇門才會開

600
00:15:39,160 --> 00:15:40,960
真正的引擎放在旁邊

601
00:15:40,960 --> 00:15:42,740
另一支23航

602
00:15:42,740 --> 00:15:44,020
你想想這個安排的意思

603
00:15:44,020 --> 00:15:47,100
全場行為改造做的最徹底的一支

604
00:15:47,100 --> 00:15:48,640
反而最不信任自動觸發

605
00:15:48,640 --> 00:15:50,180
因為什麼時候該被拷問

606
00:15:50,180 --> 00:15:51,700
這是人的決定不是模型的

607
00:15:51,700 --> 00:15:53,240
他的文件裡有一句話

608
00:15:53,240 --> 00:15:54,260
我很喜歡一場

609
00:15:54,260 --> 00:15:56,580
你從頭到尾都沒有反駁過的質問

610
00:15:56,580 --> 00:15:58,100
就是一場你本來不需要的質問

611
00:15:58,100 --> 00:16:01,180
那23航的引擎寫的是一台狀態機

612
00:16:01,180 --> 00:16:03,740
他把一場訪談化成一顆決策樹

613
00:16:03,740 --> 00:16:06,040
你的計畫裡每一件還沒決定的事

614
00:16:06,040 --> 00:16:07,320
都是樹上的一個節點

615
00:16:07,320 --> 00:16:09,100
有些節點要等別的先定下來

616
00:16:09,100 --> 00:16:11,920
然後他規定每一輪只准問前沿的問題

617
00:16:11,920 --> 00:16:13,200
前沿的意思是

618
00:16:13,200 --> 00:16:14,740
前面的條件都已經確定了

619
00:16:14,740 --> 00:16:15,760
不用靠瞎猜

620
00:16:15,760 --> 00:16:17,040
就能回答的那幾題

621
00:16:17,040 --> 00:16:19,860
這一輪的前沿要一次全部問完每題編號

622
00:16:19,860 --> 00:16:22,940
而且每一題後面要附上他自己建議的答案

623
00:16:22,940 --> 00:16:23,960
這一點很重要

624
00:16:23,960 --> 00:16:26,260
他不是丟一張空白問卷給你

625
00:16:26,260 --> 00:16:28,060
他是帶著立場來問

626
00:16:28,060 --> 00:16:29,840
你只要說我同意或者說不對

627
00:16:29,840 --> 00:16:32,400
他還把責任切開寫實在文件裡

628
00:16:32,400 --> 00:16:33,940
查得到的事實是

629
00:16:33,940 --> 00:16:35,980
你的工作永遠不是使用者的工作

630
00:16:35,980 --> 00:16:37,520
環境裡查得到的自己去查

631
00:16:37,520 --> 00:16:39,060
可是決策是使用者的

632
00:16:39,060 --> 00:16:40,340
不准替他決定

633
00:16:40,340 --> 00:16:41,620
最後是出口條件

634
00:16:41,620 --> 00:16:43,660
我覺得這是整支最值得學的一句

635
00:16:43,660 --> 00:16:44,440
他不寫問

636
00:16:44,440 --> 00:16:45,720
滿幾題就結束

637
00:16:45,720 --> 00:16:47,000
他寫前沿清空

638
00:16:47,000 --> 00:16:49,040
而且沒有任何東西被默默假設

639
00:16:49,040 --> 00:16:50,320
這一場才算結束

640
00:16:50,320 --> 00:16:53,140
這件事還有後續有人抱怨另一個工具

641
00:16:53,140 --> 00:16:54,680
用這套問了200題

642
00:16:54,680 --> 00:16:56,720
要求作者加上提數上限

643
00:16:56,720 --> 00:16:57,740
作者公開拒絕了

644
00:16:57,740 --> 00:16:59,280
他說用自然語言引導他

645
00:16:59,280 --> 00:17:01,340
本來就是這套東西的控制界面

646
00:17:01,580 --> 00:17:02,620
不是一個數字

647
00:17:02,620 --> 00:17:04,400
你覺得問太多你就叫他停

648
00:17:04,400 --> 00:17:06,960
因為提數上限會把問太多跟問得爛

649
00:17:06,960 --> 00:17:08,240
混成同一件事

650
00:17:08,760 --> 00:17:09,780
同樣是要改習慣

651
00:17:10,040 --> 00:17:11,320
另外一條線走的是紀律

652
00:17:11,320 --> 00:17:14,380
這是一位工程師公開的一整套工作紀律

653
00:17:14,640 --> 00:17:17,720
Skill好幾萬個開發者在用我挑其中最有名的一隻

654
00:17:17,720 --> 00:17:19,260
先寫測試那一隻

655
00:17:19,260 --> 00:17:20,780
他的開場三句話很漂亮

656
00:17:20,780 --> 00:17:23,080
先寫測試看著他失敗再寫

657
00:17:23,080 --> 00:17:24,620
剛好能讓他通過的程式

658
00:17:24,620 --> 00:17:26,160
這件事本身不新

659
00:17:26,160 --> 00:17:27,180
新的是他怎麼寫

660
00:17:27,180 --> 00:17:28,720
這份文件有一半的篇幅

661
00:17:28,720 --> 00:17:30,260
不是在教你正確的做法

662
00:17:30,260 --> 00:17:32,300
他在列你會想偷懶的每句話

663
00:17:32,300 --> 00:17:33,320
他做了一張表

664
00:17:33,320 --> 00:17:34,860
左邊是藉口右邊是現實

665
00:17:35,120 --> 00:17:35,880
一共11條

666
00:17:36,140 --> 00:17:37,420
這件事很簡單

667
00:17:37,420 --> 00:17:38,180
不用測試

668
00:17:38,180 --> 00:17:40,240
右邊回你簡單的事情才最容易錯

669
00:17:40,500 --> 00:17:41,520
我等一下補測試

670
00:17:41,520 --> 00:17:42,280
右邊回你

671
00:17:42,280 --> 00:17:43,060
等一下不會來

672
00:17:43,300 --> 00:17:44,340
我先留著當參考

673
00:17:44,580 --> 00:17:46,380
右邊回你刪除就是刪除他

674
00:17:46,640 --> 00:17:48,180
甚至預測了模型的內心系

675
00:17:48,420 --> 00:17:50,220
只要你心裡冒出這一次先跳過

676
00:17:50,220 --> 00:17:52,020
這句話本身就是合理化

677
00:17:52,260 --> 00:17:53,540
後面還有是三條紅旗

678
00:17:54,060 --> 00:17:56,360
每一條的處置方式都一樣刪掉重寫

679
00:17:56,620 --> 00:17:57,140
你發現了嗎

680
00:17:57,380 --> 00:17:59,440
他把說明書寫成了一道心理防線

681
00:17:59,700 --> 00:18:02,500
作者的假設是模型一定會想超捷徑

682
00:18:02,760 --> 00:18:04,560
那我就把每一條捷徑先寫出來

683
00:18:04,560 --> 00:18:05,580
一條一條堵住

684
00:18:06,100 --> 00:18:07,380
同一套裡面還有一隻

685
00:18:07,620 --> 00:18:09,420
專制AI嘴上說做完了

686
00:18:09,420 --> 00:18:10,960
他的觸發時機寫得很妙

687
00:18:11,220 --> 00:18:12,240
不是寫在哪一種任務

688
00:18:12,500 --> 00:18:14,800
是寫在你正要宣稱這件事完成修好

689
00:18:15,060 --> 00:18:15,820
通過的那一刻

690
00:18:16,080 --> 00:18:18,120
等於把觸發器裝在模型的嘴上

691
00:18:18,380 --> 00:18:19,400
他的鐵律只有一句

692
00:18:19,400 --> 00:18:20,940
沒有新鮮的驗證證據

693
00:18:21,200 --> 00:18:21,960
不准宣稱完成

694
00:18:22,220 --> 00:18:23,500
後面一樣掛一張表

695
00:18:23,760 --> 00:18:24,780
竹條處理內心性

696
00:18:25,040 --> 00:18:25,540
應該可以了

697
00:18:25,800 --> 00:18:26,560
回你去跑一次

698
00:18:26,820 --> 00:18:27,600
我很有信心

699
00:18:28,100 --> 00:18:29,120
回你信心不等於證據

700
00:18:29,380 --> 00:18:29,900
我累了

701
00:18:30,160 --> 00:18:31,180
回你疲勞不是藉口

702
00:18:31,680 --> 00:18:33,220
連說一句太好了都被列進紅旗

703
00:18:33,480 --> 00:18:35,520
因為那通常出現在還沒檢查的時候

704
00:18:35,780 --> 00:18:36,300
最後一句

705
00:18:36,560 --> 00:18:37,580
最重跳過任何一步

706
00:18:37,840 --> 00:18:39,120
那叫說謊不叫驗證

707
00:18:39,620 --> 00:18:41,420
作者還做了一件很少人做的事

708
00:18:41,680 --> 00:18:43,980
他把壓力測試放進公開的專案裡

709
00:18:44,240 --> 00:18:46,020
測試腳本模擬線上全掛錢

710
00:18:46,280 --> 00:18:47,040
每一秒都在燒

711
00:18:47,300 --> 00:18:48,080
主管在旁邊咆哮

712
00:18:48,080 --> 00:18:50,640
看這支skill撐不撐得住撐住了才算數

713
00:18:50,900 --> 00:18:52,940
這套做法背後其實有實證

714
00:18:53,200 --> 00:18:54,480
有一份2025年的研究

715
00:18:54,740 --> 00:18:56,020
看了將近三萬次對話

716
00:18:56,280 --> 00:18:58,060
結果是加上權威跟童賊壓力

717
00:18:58,320 --> 00:18:59,080
這兩種說法之後

718
00:18:59,340 --> 00:19:00,880
模型的服從3x3

719
00:19:01,400 --> 00:19:02,420
升到7x2人身上

720
00:19:02,680 --> 00:19:04,980
有效的說服在模型身上一樣有效

721
00:19:05,740 --> 00:19:06,520
這一整類skill

722
00:19:07,020 --> 00:19:08,820
我讀完整理出五招共通的做法

723
00:19:09,080 --> 00:19:11,880
他們是前面那六條鐵律在行為這一層的

724
00:19:12,140 --> 00:19:12,660
具體版本

725
00:19:12,920 --> 00:19:14,960
第一招實際是要卡在工作開始之前

726
00:19:15,220 --> 00:19:15,980
教工具的skill

727
00:19:16,240 --> 00:19:17,000
寫的是我會做什麼

728
00:19:17,000 --> 00:19:19,560
改行為的skill寫的是我要在哪一刻攔住你

729
00:19:19,820 --> 00:19:20,320
有一支

730
00:19:20,580 --> 00:19:22,380
寫的是在你決定一個計劃之前

731
00:19:22,640 --> 00:19:23,660
第二招出口條件

732
00:19:23,920 --> 00:19:24,420
用狀態

733
00:19:24,680 --> 00:19:26,480
不用數量前沿清空就是一個例子

734
00:19:26,720 --> 00:19:27,500
第三招

735
00:19:27,760 --> 00:19:29,040
把壞習慣寫成對照表

736
00:19:29,280 --> 00:19:30,060
模型對這樣錯

737
00:19:30,320 --> 00:19:30,820
改成這樣

738
00:19:31,080 --> 00:19:31,600
比對抽象的美德

739
00:19:31,840 --> 00:19:32,620
敏感的多

740
00:19:32,880 --> 00:19:33,380
第四招

741
00:19:33,640 --> 00:19:33,900
形容詞

742
00:19:34,160 --> 00:19:34,400
換成字冊題

743
00:19:34,660 --> 00:19:35,680
第五招最重要

744
00:19:35,940 --> 00:19:37,220
說服不夠的時候就上機制

745
00:19:37,480 --> 00:19:39,020
這中間是一個光譜

746
00:19:39,280 --> 00:19:40,040
最輕的一端

747
00:19:40,300 --> 00:19:40,800
有一支skill

748
00:19:41,060 --> 00:19:41,840
全文只有一句話

749
00:19:42,080 --> 00:19:43,620
停你剛剛那段沒有說服我

750
00:19:43,880 --> 00:19:44,400
重講一次

751
00:19:44,400 --> 00:19:46,700
二十個字就是一顆打斷按鈕

752
00:19:46,960 --> 00:19:50,280
中間是剛剛那套拷問用演算法規定他要怎麼問

753
00:19:50,540 --> 00:19:51,560
最重的一端有一支

754
00:19:51,820 --> 00:19:53,620
要你動工之前先寫計劃檔

755
00:19:53,880 --> 00:19:54,640
他不靠說服

756
00:19:54,900 --> 00:19:55,920
他掛了四個hook

757
00:19:56,180 --> 00:19:57,720
hook就是系統層級的關卡

758
00:19:58,220 --> 00:19:59,760
你每送出一句話系統自己去執行

759
00:19:59,760 --> 00:20:27,520
理容 gear

760
00:20:00,000 --> 00:20:01,080
不管模型记不记得

761
00:20:01,080 --> 00:20:04,360
他每一回合都把计划党重新塞回模型眼前

762
00:20:04,360 --> 00:20:08,120
还规定连续失败三次就必须停下来回报人类

763
00:20:08,120 --> 00:20:09,520
你要一件事百分之百发生

764
00:20:09,520 --> 00:20:11,680
就不要拜托模型记得

765
00:20:11,680 --> 00:20:13,400
行为还有规则可以写

766
00:20:13,400 --> 00:20:14,720
可是有一层更难

767
00:20:14,720 --> 00:20:17,160
那就是品味官方有一支很不一样的作品

768
00:20:17,160 --> 00:20:18,280
只有55行

769
00:20:18,280 --> 00:20:19,560
管的是界面设计

770
00:20:19,560 --> 00:20:20,920
他的描述没有写功能

771
00:20:20,920 --> 00:20:21,960
写的是一种病

772
00:20:21,960 --> 00:20:24,600
要做出不像模板预设值的选择

773
00:20:24,600 --> 00:20:25,440
你有没有发现

774
00:20:25,440 --> 00:20:27,560
AI生出来的网页常常长得很像

775
00:20:27,560 --> 00:20:29,600
这支skill直接把那个像点名出来

776
00:20:29,600 --> 00:20:30,720
他说有三种长相

777
00:20:30,720 --> 00:20:32,720
一种是米白底色配针线大标题

778
00:20:32,720 --> 00:20:33,640
再配一个赤桃局

779
00:20:33,640 --> 00:20:35,440
一种是近乎黑的底色

780
00:20:35,440 --> 00:20:36,800
配荧光绿的重点色

781
00:20:36,800 --> 00:20:39,040
一种是报纸感的细线

782
00:20:39,040 --> 00:20:40,480
加上完全没有圆角

783
00:20:40,480 --> 00:20:42,960
他接着说这三种都很好看他们合法

784
00:20:42,960 --> 00:20:45,120
可是他们是预设值不是选择

785
00:20:45,720 --> 00:20:47,200
他做法不是给检查表

786
00:20:47,200 --> 00:20:48,080
是先派角色

787
00:20:48,080 --> 00:20:51,120
他说你现在是一间小工作室的设计总监

788
00:20:51,120 --> 00:20:54,040
你的客户已经把所有有模板感的提案

789
00:20:54,040 --> 00:20:54,840
都退回来过了

790
00:20:54,840 --> 00:20:56,000
然后他规定两段是

791
00:20:56,000 --> 00:20:57,640
先定一套自己的设计语言

792
00:20:57,640 --> 00:20:58,960
颜色字体版型

793
00:20:58,960 --> 00:21:00,080
加一个记忆点

794
00:21:00,080 --> 00:21:01,200
定完先自我批判

795
00:21:01,200 --> 00:21:02,480
一轮问自己一句

796
00:21:02,480 --> 00:21:05,240
这是不是我对任何类似题目都会给的答案

797
00:21:05,240 --> 00:21:07,440
过了这一关才准开始写程式

798
00:21:07,440 --> 00:21:09,680
收尾他引了香奈儿那一句话

799
00:21:09,680 --> 00:21:11,040
出门之前照一次镜子

800
00:21:11,040 --> 00:21:12,520
拿掉一件配饰

801
00:21:12,520 --> 00:21:13,840
你要改的是品位的时候

802
00:21:13,840 --> 00:21:15,400
你给的不是规则

803
00:21:15,400 --> 00:21:17,120
角色跟反例

804
00:21:17,120 --> 00:21:18,320
到这里六条铁律

805
00:21:18,320 --> 00:21:19,840
你都看过长什么样子了

806
00:21:19,840 --> 00:21:21,880
可是只读十几只样本太小

807
00:21:21,880 --> 00:21:22,880
我不想只靠感觉

808
00:21:22,880 --> 00:21:23,880
所以我做了一件事

809
00:21:23,880 --> 00:21:25,920
我把38支有名的skill抓下来

810
00:21:25,920 --> 00:21:26,800
一支一支亮

811
00:21:26,800 --> 00:21:28,240
四组官方17支

812
00:21:28,240 --> 00:21:30,800
那一整套纪律型的14支

813
00:21:30,800 --> 00:21:32,240
两支社群名作

814
00:21:32,240 --> 00:21:33,440
再加上笔记软体

815
00:21:33,440 --> 00:21:34,480
那五支每一支亮

816
00:21:34,480 --> 00:21:35,920
六件事说明书多少行

817
00:21:35,920 --> 00:21:36,960
描述多少字圆

818
00:21:36,960 --> 00:21:38,160
用了几次大写命令

819
00:21:38,160 --> 00:21:39,280
附了几支程式

820
00:21:39,280 --> 00:21:40,800
连出去几个参考档

821
00:21:40,800 --> 00:21:42,120
有几段範例程式码

822
00:21:42,120 --> 00:21:43,920
先讲最反直觉的一个结果

823
00:21:43,920 --> 00:21:44,800
它们都很短

824
00:21:44,800 --> 00:21:47,360
全部38支的中位数是182行

825
00:21:47,360 --> 00:21:50,400
而官方那17支中位数只有130行

826
00:21:50,400 --> 00:21:54,800
最短的那一支管内部沟通写作的整份说明书只有32行

827
00:21:54,800 --> 00:21:57,520
32行官方拿它当正式的范例放出来

828
00:21:57,520 --> 00:21:59,840
38支里面超过300行

829
00:21:59,840 --> 00:22:01,600
只有9支不到四分之一

830
00:22:01,600 --> 00:22:02,720
顺带一提

831
00:22:02,720 --> 00:22:04,560
全场最长的是哪一支

832
00:22:04,560 --> 00:22:07,200
是教你怎么写skill的那一支679行

833
00:22:07,200 --> 00:22:09,200
教别人精简的那个人自己写最长

834
00:22:09,200 --> 00:22:10,560
这件事有点好笑

835
00:22:10,560 --> 00:22:11,360
可是也合理

836
00:22:11,360 --> 00:22:12,880
因为那是一本教科书

837
00:22:12,880 --> 00:22:14,160
不是一份作业手册

838
00:22:14,160 --> 00:22:16,040
所以本文只放地图这一条

839
00:22:16,040 --> 00:22:17,040
不是我的偏好

840
00:22:17,040 --> 00:22:18,280
是数字

841
00:22:18,280 --> 00:22:19,600
第二个发现更有意思

842
00:22:19,600 --> 00:22:20,800
同样是描述那一栏

843
00:22:20,800 --> 00:22:22,360
两派人的写法差很多

844
00:22:22,360 --> 00:22:26,000
官方那17支描述的中位数是319个字元

845
00:22:26,000 --> 00:22:27,600
最长的一支写到1000多

846
00:22:27,600 --> 00:22:29,320
几乎顶到上限纪律型

847
00:22:29,320 --> 00:22:31,720
那14支中位数只有105个字元

848
00:22:31,720 --> 00:22:32,520
差了三倍

849
00:22:32,520 --> 00:22:34,200
再看大写命令词

850
00:22:34,200 --> 00:22:35,000
纪律型

851
00:22:35,000 --> 00:22:38,040
平均每一支用的次数是官方的2.4倍

852
00:22:38,040 --> 00:22:40,200
而笔记软体那5支全部是0

853
00:22:40,200 --> 00:22:41,240
一次都没有用过

854
00:22:41,240 --> 00:22:43,400
我本来的猜测是爱用大写的skill

855
00:22:43,400 --> 00:22:45,640
就是没有程式可以用的那一种

856
00:22:45,640 --> 00:22:46,720
我算了一下

857
00:22:46,720 --> 00:22:49,040
这两件事的相关系数只有0.18

858
00:22:49,040 --> 00:22:50,000
几乎没有关系

859
00:22:50,000 --> 00:22:51,160
我猜错了

860
00:22:51,160 --> 00:22:52,320
可是我算到另外一组

861
00:22:52,320 --> 00:22:54,080
相关系数是0.571

862
00:22:54,080 --> 00:22:56,880
这在这种样本里算是很强的讯号

863
00:22:56,880 --> 00:22:59,160
那一组是描述写的越长的skill

864
00:22:59,160 --> 00:23:00,280
负的程式越多

865
00:23:00,280 --> 00:23:02,480
工具型的skill把功夫写成程式码

866
00:23:02,480 --> 00:23:03,920
所以说明书可以短

867
00:23:03,920 --> 00:23:06,360
可是他必须把使用者可能会讲的

868
00:23:06,360 --> 00:23:08,120
每一种情况都写进描述里

869
00:23:08,120 --> 00:23:09,280
不然没有人翻得到他

870
00:23:09,280 --> 00:23:10,400
纪律型刚好

871
00:23:10,400 --> 00:23:11,880
相反他没有程式码

872
00:23:11,880 --> 00:23:14,200
可以依靠他要改的是模型的行为

873
00:23:14,200 --> 00:23:15,720
所以描述一句话就够

874
00:23:15,720 --> 00:23:17,120
力气全部花在正文

875
00:23:17,120 --> 00:23:18,920
把道理讲到你不会走偏

876
00:23:18,920 --> 00:23:19,840
这正好对

877
00:23:19,840 --> 00:23:20,920
回一开始那张图

878
00:23:20,920 --> 00:23:22,200
你要改的层不一样

879
00:23:22,200 --> 00:23:24,680
连描述该写多长都不一样

880
00:23:24,680 --> 00:23:27,000
不过我刚刚量的都是他们今天的样子

881
00:23:27,000 --> 00:23:28,960
这些档案不是写完就摆着

882
00:23:28,960 --> 00:23:30,240
每个月都在被改

883
00:23:30,240 --> 00:23:33,000
所以我把10个GitHub专案的完整历史抓下来

884
00:23:33,000 --> 00:23:35,280
每个月底重新量一次同一棵树

885
00:23:35,280 --> 00:23:37,320
同一批作者同一批技能

886
00:23:37,320 --> 00:23:38,600
只有时间在动

887
00:23:38,600 --> 00:23:39,520
方向很清楚

888
00:23:39,520 --> 00:23:42,240
官方那支处理文书档的skill描述栏

889
00:23:42,240 --> 00:23:44,880
从362个字原厂到835

890
00:23:44,880 --> 00:23:47,920
本文反而从197行砍到91行

891
00:23:47,920 --> 00:23:49,400
门口的招牌越写越大

892
00:23:49,400 --> 00:23:51,040
屋子里的东西越搬越少

893
00:23:51,040 --> 00:23:52,440
描述写的内容也换了

894
00:23:52,440 --> 00:23:53,560
以前写这是什么

895
00:23:53,560 --> 00:23:55,760
现在写什么时候用第一版里

896
00:23:55,760 --> 00:23:58,040
那四个大写命令现在一个都不剩

897
00:23:58,040 --> 00:23:59,800
还有一件更反直觉的事

898
00:23:59,800 --> 00:24:02,840
大家以为长文会被搬去旁边的参考档

899
00:24:02,840 --> 00:24:03,760
官方的做法是

900
00:24:03,760 --> 00:24:05,240
把两个参考档直接删掉

901
00:24:05,240 --> 00:24:06,680
将近1000行整个不见

902
00:24:06,680 --> 00:24:07,480
只留一张表

903
00:24:07,480 --> 00:24:08,960
跟一份地雷清单

904
00:24:08,960 --> 00:24:11,000
废工的部分推给脚本

905
00:24:11,000 --> 00:24:13,080
管试算表的那一支更清楚

906
00:24:13,080 --> 00:24:14,040
第一版写的是

907
00:24:14,040 --> 00:24:16,000
每一份试算表都必须零公式

908
00:24:16,000 --> 00:24:16,440
错误

909
00:24:16,440 --> 00:24:17,800
现在同一条规则写成

910
00:24:17,800 --> 00:24:19,920
只要那支检查程式还在报错

911
00:24:19,920 --> 00:24:20,880
就不要交出去

912
00:24:20,880 --> 00:24:22,160
规则一个字都没变

913
00:24:22,160 --> 00:24:23,680
可是它从一句口号

914
00:24:23,680 --> 00:24:25,160
变成一支跑得动的程式

915
00:24:25,160 --> 00:24:27,440
还有一个很干脆的日子

916
00:24:27,440 --> 00:24:28,840
2025年16号

917
00:24:28,840 --> 00:24:30,480
官方把格式定下来的那一天

918
00:24:30,480 --> 00:24:33,120
社群那套纪律型SKILL的作者

919
00:24:33,120 --> 00:24:34,840
把自己发明的两个栏位拿掉

920
00:24:34,840 --> 00:24:36,200
收回成一个描述栏

921
00:24:36,200 --> 00:24:38,520
从那天起没有人再用过

922
00:24:38,520 --> 00:24:39,640
规格归规格

923
00:24:39,640 --> 00:24:41,320
外面的市场又是另外一回事

924
00:24:41,320 --> 00:24:42,280
今年上半年

925
00:24:42,280 --> 00:24:43,680
这件事的气氛整个变了

926
00:24:43,680 --> 00:24:45,040
公开的SKILL数量

927
00:24:45,040 --> 00:24:46,360
冲到47000只

928
00:24:46,360 --> 00:24:47,920
有人做了一次品质抽查

929
00:24:47,920 --> 00:24:49,000
满分12分

930
00:24:49,000 --> 00:24:50,360
平均只有6.2分

931
00:24:50,360 --> 00:24:52,640
你可以想成一个大市场摊位

932
00:24:52,640 --> 00:24:54,760
从10个变成47000个

933
00:24:54,760 --> 00:24:55,880
可是有一半的摊位

934
00:24:55,880 --> 00:24:57,360
卖的是同一包东西

935
00:24:57,360 --> 00:25:00,000
包装还是自动生成的痛点就转弯了

936
00:25:00,000 --> 00:25:01,960
以前大家问的是有没有人写过这个

937
00:25:01,960 --> 00:25:03,760
现在大家问的是这一支能不能信

938
00:25:03,760 --> 00:25:05,940
所以三件事同时在发生

939
00:25:05,940 --> 00:25:07,760
第一 大公司开始自己进驻

940
00:25:07,760 --> 00:25:09,260
设计工具 笔记软体

941
00:25:09,260 --> 00:25:11,960
经流服务都把官方认证的skill放进目录

942
00:25:11,960 --> 00:25:15,460
第二 企业版本开始提供整个组织统一管理跟集合

943
00:25:15,460 --> 00:25:17,980
因为装一支skill等于装一套软体

944
00:25:17,980 --> 00:25:19,200
它里面可以复程式

945
00:25:19,200 --> 00:25:21,800
公司必须知道员工到底装了什么

946
00:25:21,800 --> 00:25:23,440
第三红的定义变了

947
00:25:23,440 --> 00:25:25,780
2025年红是讨论区的头条

948
00:25:25,780 --> 00:25:26,940
是一个晚上被转报

949
00:25:26,940 --> 00:25:29,180
2026年红是进了官方目录

950
00:25:29,180 --> 00:25:31,000
是安装数字 是评测分数

951
00:25:31,000 --> 00:25:32,580
这条路你应该觉得眼熟

952
00:25:32,580 --> 00:25:35,080
任何一个应用程式商店都走过同一段

953
00:25:35,080 --> 00:25:37,000
skill一多就会有考试

954
00:25:37,000 --> 00:25:39,080
这刚好带我们回到最后一条铁律

955
00:25:39,080 --> 00:25:39,840
写完要测

956
00:25:39,840 --> 00:25:43,240
今年出现一份专门考skill的评测86个任务

957
00:25:43,240 --> 00:25:45,120
横跨11个领域做法很简单

958
00:25:45,120 --> 00:25:46,880
同一个模型 同一批任务

959
00:25:46,880 --> 00:25:48,420
一边给它精选过的skill

960
00:25:48,420 --> 00:25:50,520
一边什么都不给比通过率

961
00:25:50,520 --> 00:25:54,440
结果是有skill的那一边平均高出16.2个百分点

962
00:25:54,440 --> 00:25:56,380
这个数字同时说了两件事

963
00:25:56,380 --> 00:25:58,560
第一 一份写得好的说明书

964
00:25:58,560 --> 00:25:59,740
效果是实打实的

965
00:25:59,740 --> 00:26:01,520
第二 那也只是十几个百分点

966
00:26:01,520 --> 00:26:02,880
不是换一颗脑袋

967
00:26:02,880 --> 00:26:05,420
skill是把模型本来就会的事情

968
00:26:05,420 --> 00:26:06,720
导到正确的做法上

969
00:26:06,720 --> 00:26:08,480
不是给它一个新的能力

970
00:26:08,480 --> 00:26:10,380
所以官方最早的那个建议

971
00:26:10,380 --> 00:26:11,780
现在看起来特别有分量

972
00:26:11,780 --> 00:26:13,500
他说先做评测再写文件

973
00:26:13,500 --> 00:26:14,480
顺序是这样

974
00:26:14,480 --> 00:26:16,360
你先让模型裸着做一次

975
00:26:16,360 --> 00:26:18,080
把它失败的地方记下来

976
00:26:18,080 --> 00:26:19,580
写成三个测试情境

977
00:26:19,580 --> 00:26:20,600
亮出一个基准分

978
00:26:20,600 --> 00:26:22,080
然后才开始写指令

979
00:26:22,080 --> 00:26:23,100
先写最少的那一版

980
00:26:23,100 --> 00:26:24,600
再一轮一轮加那套

981
00:26:24,600 --> 00:26:26,740
纪律型skill的作者更严格

982
00:26:26,740 --> 00:26:28,000
他规定任何人要改

983
00:26:28,000 --> 00:26:29,640
他的skill都必须附上

984
00:26:29,640 --> 00:26:31,520
改之前跟改之后的评测对照

985
00:26:31,520 --> 00:26:33,100
这一条最容易被跳过

986
00:26:33,100 --> 00:26:34,940
可是他是唯一能告诉你

987
00:26:34,940 --> 00:26:36,540
你刚刚写的那一堆字

988
00:26:36,540 --> 00:26:37,740
到底有没有用的东西

989
00:26:38,460 --> 00:26:40,380
我们今天先给答案再回头看证据

990
00:26:40,380 --> 00:26:42,060
如果你只带走一句话

991
00:26:42,060 --> 00:26:43,240
我希望是这一句

992
00:26:43,240 --> 00:26:45,280
skill不是给AI的新功能

993
00:26:45,280 --> 00:26:46,620
是给AI的说明书

994
00:26:46,620 --> 00:26:47,620
而说明书的价值

995
00:26:47,620 --> 00:26:48,660
不在你写了多少

996
00:26:48,660 --> 00:26:50,500
在他被翻开的那一刻

997
00:26:50,500 --> 00:26:52,200
刚好写着他需要的那一句

998
00:26:52,200 --> 00:26:53,240
六条铁律

999
00:26:53,240 --> 00:26:54,860
我再念一次描述是掉沟

1000
00:26:54,860 --> 00:26:56,580
本文只放地图确定的事情

1001
00:26:56,580 --> 00:26:57,300
交给脚本

1002
00:26:57,300 --> 00:26:59,240
解释为什么胜过用大写压人

1003
00:26:59,240 --> 00:27:00,420
会被抄的捷径

1004
00:27:00,420 --> 00:27:01,480
先写死堵住

1005
00:27:01,480 --> 00:27:02,860
写完一定要测

1006
00:27:02,860 --> 00:27:04,280
那今天晚上你可以做什么

1007
00:27:04,280 --> 00:27:05,440
如果你是一般使用者

1008
00:27:05,440 --> 00:27:06,700
你把重复交代过

1009
00:27:06,700 --> 00:27:08,000
三次以上的那件事

1010
00:27:08,000 --> 00:27:09,300
写成一个资料夹

1011
00:27:09,300 --> 00:27:10,980
开头那句描述好好写

1012
00:27:10,980 --> 00:27:12,400
写清楚什么时候该用它

1013
00:27:12,400 --> 00:27:14,440
如果你在写程式或者做产品

1014
00:27:14,440 --> 00:27:16,140
你先问自己要改哪一层

1015
00:27:16,140 --> 00:27:18,000
能封进程式的就封进去

1016
00:27:18,000 --> 00:27:19,480
说明书留给判断题

1017
00:27:19,480 --> 00:27:20,900
然后把触发情境列满

1018
00:27:20,900 --> 00:27:22,520
如果你是内容创作者

1019
00:27:22,520 --> 00:27:24,340
你最该练的是品位那一层

1020
00:27:24,340 --> 00:27:26,000
把你不要的三种长相

1021
00:27:26,000 --> 00:27:27,920
写下来给模型角色跟反例

1022
00:27:27,920 --> 00:27:29,280
不要给检查表

1023
00:27:29,280 --> 00:27:30,560
最后问你一个问题

1024
00:27:30,560 --> 00:27:31,720
如果你只能写一只skill

1025
00:27:31,720 --> 00:27:33,360
长期放在你的AI旁边

1026
00:27:33,360 --> 00:27:34,980
你会拿它来改哪一件事

1027
00:27:34,980 --> 00:27:35,780
是它的知识

1028
00:27:35,780 --> 00:27:36,800
还是它的习惯

1029
00:27:36,800 --> 00:27:38,120
留言告诉我

1030
00:27:38,120 --> 00:27:39,860
喜欢的话订阅AI Notebook

1031
00:27:39,860 --> 00:27:40,500
开启通知

1032
00:27:40,500 --> 00:27:41,160
我们下次见