1
00:00:00,080 --> 00:00:04,040
Skills are, in my view, one of 
the most important and probably 

2
00:00:04,400 --> 00:00:07,200
one of the most underused 
features in the. 

3
00:00:07,240 --> 00:00:11,800
Cloud ecosystem right now. 
And even arguably beyond the 

4
00:00:11,800 --> 00:00:15,840
cloud ecosystem, if you're using
cloud for anything serious, you 

5
00:00:15,840 --> 00:00:17,440
probably hit. 
This problem. 

6
00:00:17,560 --> 00:00:20,400
You've got it working exactly 
how you want it for a specific 

7
00:00:20,400 --> 00:00:23,720
task that has a perfect output, 
and then the next time you have 

8
00:00:23,800 --> 00:00:25,720
a conversation it's just 
begotten everything. 

9
00:00:26,520 --> 00:00:28,560
You could copy and paste 
instructions into the chat every

10
00:00:28,560 --> 00:00:32,000
time, and many people do this, 
but that doesn't scale. 

11
00:00:32,080 --> 00:00:37,000
GPTS or clawed projects help, 
but they're not the same as it 

12
00:00:37,000 --> 00:00:40,960
understanding a process for a 
specific task, and they fill up 

13
00:00:40,960 --> 00:00:42,840
the context window far too 
quickly. 

14
00:00:43,120 --> 00:00:46,160
Because you preload them. 
Full of so much information. 

15
00:00:46,400 --> 00:00:49,160
To it has the information it 
needs to help you do a task. 

16
00:00:49,880 --> 00:00:52,200
And so you're constantly running
out of space in a conversation 

17
00:00:52,200 --> 00:00:53,480
and having to start a new 
thread. 

18
00:00:53,600 --> 00:00:56,160
The solution is skills. 
A skill is a set of instructions

19
00:00:56,160 --> 00:00:59,920
you write once give to claw. 
Encode your process and your 

20
00:00:59,920 --> 00:01:03,280
preferences, and today we're 
going to go deep on them. 

21
00:01:03,480 --> 00:01:07,040
We're going to go into the 
actual principles of context 

22
00:01:07,040 --> 00:01:09,080
engineering and why skills are 
so important. 

23
00:01:09,800 --> 00:01:13,000
The exact anatomy of a skill 
file with real examples that you

24
00:01:13,000 --> 00:01:16,480
can use today and how you can 
make them and how you can chain 

25
00:01:16,480 --> 00:01:18,160
different ones together and 
where to. 

26
00:01:18,160 --> 00:01:22,080
Find skills and install them. 
How they fit alongside MCPS. 

27
00:01:22,080 --> 00:01:24,480
Or plug insurance. 
And we shouldn't. 

28
00:01:24,480 --> 00:01:26,800
Even bother using them at all, 
this is in the loop. 

29
00:01:26,800 --> 00:01:29,400
With Jack Horton. 
Hope you enjoyed the show. 

30
00:01:40,800 --> 00:01:44,880
Before we get into what skills 
are, I think it's probably worth

31
00:01:44,880 --> 00:01:47,880
understanding the problem they 
were built to solve and kind of 

32
00:01:47,880 --> 00:01:49,880
what they consist of. 
What is a skill? 

33
00:01:50,000 --> 00:01:51,920
Because once you understand 
that, I think everything will 

34
00:01:51,920 --> 00:01:55,000
start to really fall into. 
Place so throughout. 2025 AI 

35
00:01:55,000 --> 00:01:59,160
coding agents were getting 
dramatically better and better. 

36
00:01:59,640 --> 00:02:03,040
Core code could suddenly, you 
know, build really complex 

37
00:02:03,040 --> 00:02:03,920
systems. 
They could. 

38
00:02:03,920 --> 00:02:06,200
Read an entire file system. 
They could. 

39
00:02:06,200 --> 00:02:08,039
Interact with Git. 
They can run tests, they can 

40
00:02:08,039 --> 00:02:11,120
deploy things, and the 
capability was obviously there 

41
00:02:11,200 --> 00:02:14,120
to do really cool stuff. 
But what most people found is 

42
00:02:14,120 --> 00:02:17,360
they hit a wall, and that wall 
really was the context window. 

43
00:02:17,560 --> 00:02:21,960
We've talked about context 
engineering before in July 2025,

44
00:02:21,960 --> 00:02:24,720
and it's the same for even those
that are not using core code. 

45
00:02:24,880 --> 00:02:27,200
The amount of times you just run
out of space and this happens 

46
00:02:27,200 --> 00:02:28,920
with GPTS. 
And core projects. 

47
00:02:29,000 --> 00:02:31,360
And what's actually happening 
here is that every time you want

48
00:02:31,360 --> 00:02:32,680
Core to obviously know 
something. 

49
00:02:32,680 --> 00:02:34,720
Whether? 
It was your, I don't know, 

50
00:02:34,720 --> 00:02:38,360
teams, coding standards, how an 
API works, the context of the 

51
00:02:38,360 --> 00:02:40,240
market or the business you're 
working in. 

52
00:02:40,600 --> 00:02:43,880
You basically be pasting lots of
information to the chat, which 

53
00:02:43,960 --> 00:02:46,280
essentially crams. 
Everything into the. 

54
00:02:46,280 --> 00:02:49,160
LLM System prompt which is this 
big long prompt. 

55
00:02:49,160 --> 00:02:52,200
It gets sent to an LLM that 
defines, you know, its. 

56
00:02:52,200 --> 00:02:53,720
Purpose. 
How it should behave? 

57
00:02:53,720 --> 00:02:57,280
What it should? 
Do and as agents and LLMS get 

58
00:02:57,280 --> 00:02:59,440
more powerful, the list of 
things that. 

59
00:03:00,080 --> 00:03:01,320
Is included. 
In that system. 

60
00:03:01,320 --> 00:03:03,720
Prompt is always growing, so 
there's more instructions, 

61
00:03:04,000 --> 00:03:06,160
there's more tools so the 
integrations it can use. 

62
00:03:06,160 --> 00:03:08,000
So that system prompt is just 
ballooning. 

63
00:03:08,600 --> 00:03:10,920
And the problem with that is 
that the context windows are 

64
00:03:10,920 --> 00:03:12,560
finite. 
The more you step in there, the 

65
00:03:12,560 --> 00:03:14,040
worse the agent starts to 
perform. 

66
00:03:14,600 --> 00:03:17,720
It gets slower and more 
expensive, less reliable, and 

67
00:03:18,000 --> 00:03:21,360
there's a real tension between 
giving an agent you know as much

68
00:03:21,360 --> 00:03:25,920
information as possible and 
making it able to execute 

69
00:03:25,920 --> 00:03:28,560
consistently well. 
And this insight is what led to 

70
00:03:28,560 --> 00:03:31,360
skills. 
Agents don't need all of that 

71
00:03:31,480 --> 00:03:33,520
knowledge and information all of
the time. 

72
00:03:34,240 --> 00:03:37,480
What they need is the right 
information, the right ability 

73
00:03:37,480 --> 00:03:40,200
to run a process, or awareness 
of a process at the perfect 

74
00:03:40,200 --> 00:03:42,720
moment. 
And so in in about late October 

75
00:03:42,720 --> 00:03:47,080
2025, Anthropic published a blog
called Equipping Agents for the 

76
00:03:47,080 --> 00:03:48,720
Real World. 
With agent skills. 

77
00:03:49,360 --> 00:03:50,720
And they framed. 
It kind of like this they. 

78
00:03:50,720 --> 00:03:53,800
Said Claude is really powerful, 
but real work requires 

79
00:03:53,800 --> 00:03:56,920
procedural information and just 
organizational contacts. 

80
00:03:57,400 --> 00:04:00,120
And as agents become more and 
more capable, we need more 

81
00:04:00,120 --> 00:04:03,840
composable, scalable and 
portable ways to equip them with

82
00:04:03,840 --> 00:04:08,040
that domain specific expertise. 
And that's what led to agent 

83
00:04:08,040 --> 00:04:11,040
skills, which essentially are an
organized set of folders. 

84
00:04:11,320 --> 00:04:14,400
Full of instructions and 
resources that help agents. 

85
00:04:15,240 --> 00:04:17,800
To basically discover and load 
the right thing at the right 

86
00:04:17,800 --> 00:04:20,360
time and I think the key phrase.
Here is dynamically. 

87
00:04:20,839 --> 00:04:23,840
So these skills. 
Aren't always there for. 

88
00:04:24,080 --> 00:04:27,400
The agent, they're not sitting 
in the context window taking up 

89
00:04:27,400 --> 00:04:29,480
space. 
They only get loaded when 

90
00:04:29,480 --> 00:04:31,760
they're relevant and stay out of
the way when they're not. 

91
00:04:31,920 --> 00:04:34,160
And that's A, and that's a 
fundamental. 

92
00:04:34,160 --> 00:04:35,640
Design. 
Principle of any modern day 

93
00:04:35,640 --> 00:04:37,720
agent. 
Really you know and topic is a 

94
00:04:37,720 --> 00:04:38,680
specific. 
Name for it. 

95
00:04:38,880 --> 00:04:40,160
That's been adopted. 
Everywhere. 

96
00:04:40,240 --> 00:04:43,360
Which is called. 
Progressive disclosure so you 

97
00:04:43,360 --> 00:04:46,080
progressively. 
Disclose more information to the

98
00:04:46,080 --> 00:04:47,720
agent needs. 
Only. 

99
00:04:47,720 --> 00:04:50,280
Actually when it needs it. 
So this progressive disclosure 

100
00:04:50,280 --> 00:04:51,320
happens. 
In three layers. 

101
00:04:51,640 --> 00:04:55,960
Layer one when Claude starts a. 
Session it basically preloads. 

102
00:04:55,960 --> 00:04:58,720
Just the name. 
And the description of the 

103
00:04:58,720 --> 00:05:00,480
skills. 
So not the full set of 

104
00:05:00,480 --> 00:05:02,160
instructions, just a one line 
summary. 

105
00:05:02,680 --> 00:05:04,320
You know, maybe 100 tokens per 
skill. 

106
00:05:04,320 --> 00:05:07,400
So again, a few words. 
This makes it possible for the 

107
00:05:07,400 --> 00:05:10,440
agent to know, I guess, what is 
available to it. 

108
00:05:10,640 --> 00:05:12,560
We just talked about this 
before, the tools that were 

109
00:05:12,560 --> 00:05:15,440
available to this agent and 
therefore it's cheap enough to 

110
00:05:15,440 --> 00:05:17,680
always have those small. 
Descriptions loaded. 

111
00:05:17,840 --> 00:05:20,240
Then you've got layer 2 when 
when essentially Claude 

112
00:05:20,240 --> 00:05:23,400
encounters a task and recognises
that one of those descriptions 

113
00:05:23,400 --> 00:05:25,560
is relevant. 
It then loads that full skill 

114
00:05:25,560 --> 00:05:28,640
file and as a result it will 
have the complete. 

115
00:05:28,640 --> 00:05:31,320
Rules and instructions that you.
Create for it, but only does it 

116
00:05:31,320 --> 00:05:33,520
for that specific skill, not all
of the skills. 

117
00:05:34,280 --> 00:05:35,920
So again, you could give it 
dirty. 

118
00:05:36,040 --> 00:05:39,240
Skills or tools? 
And it will only. 

119
00:05:39,240 --> 00:05:42,400
Use one of them so again. 
Protecting that context window, 

120
00:05:42,400 --> 00:05:47,400
and then you've got layer three.
If the skill MD file references 

121
00:05:47,400 --> 00:05:50,960
other files in the folder of 
templates or reference docs or 

122
00:05:51,120 --> 00:05:54,480
scripts, code will actually be 
able to navigate to those as 

123
00:05:54,480 --> 00:05:57,040
needed. 
It doesn't read everything 

124
00:05:57,040 --> 00:05:59,640
upfront, it follows it for 
references. 

125
00:05:59,840 --> 00:06:03,760
So if a process that it's 
following says use this extra 

126
00:06:03,760 --> 00:06:06,120
doc, it will then go and find 
that extra dock. 

127
00:06:06,120 --> 00:06:07,760
It won't then load every single 
skill. 

128
00:06:07,960 --> 00:06:09,320
Think about it like a set. 
Of instructions. 

129
00:06:09,320 --> 00:06:12,400
Basically layer 1 is like the 
table of contents, layer 2 is 

130
00:06:12,400 --> 00:06:16,160
that you know, relevant chapter 
I guess and layer. 3 is the 

131
00:06:16,160 --> 00:06:17,960
appendix. 
You don't need the whole thing 

132
00:06:17,960 --> 00:06:19,960
every time. 
So now you understand, you know,

133
00:06:19,960 --> 00:06:22,360
why we have skills, why they're 
important and what they're kind 

134
00:06:22,360 --> 00:06:25,800
of made-up of. 
I think then let's build a skill

135
00:06:25,880 --> 00:06:27,280
together. 
And I think the walkthrough. 

136
00:06:27,280 --> 00:06:28,320
Of this. 
Helps you. 

137
00:06:28,320 --> 00:06:31,720
Really understand what a skill 
is beyond just me using. 

138
00:06:31,720 --> 00:06:34,000
Kind of abstract words, 
instructions, because I think 

139
00:06:34,000 --> 00:06:35,720
it's really important to. 
Understand the mechanics. 

140
00:06:35,720 --> 00:06:37,920
And the specifics here. 
So if we build 1 together, 

141
00:06:37,920 --> 00:06:41,720
right, you have to picture this.
So let's say you write a team 

142
00:06:41,720 --> 00:06:45,080
update every single Friday, and 
every time you ask Claude for 

143
00:06:45,080 --> 00:06:47,360
help with this, you end up 
giving it basically the same set

144
00:06:47,360 --> 00:06:49,880
of prompts every single time. 
You know, keep it under 300 

145
00:06:49,880 --> 00:06:51,480
words. 
Lead with wins. 

146
00:06:51,520 --> 00:06:54,360
Plain language, no weird AI, 
slop language. 

147
00:06:55,000 --> 00:06:57,160
End with the blockers and every 
few. 

148
00:06:57,160 --> 00:06:58,440
Weeks Cloud does something 
annoying. 

149
00:06:58,440 --> 00:07:01,080
Maybe it allowed a random sign 
off or use the wrong language 

150
00:07:01,080 --> 00:07:03,120
again and you have to then. 
Correct it. 

151
00:07:03,280 --> 00:07:05,240
So here's how you would solve 
that with skills. 

152
00:07:21,150 --> 00:07:22,670
Now before I get into this, 
it's. 

153
00:07:22,670 --> 00:07:25,550
This is a few bits of a process 
here, and it's also important to

154
00:07:25,550 --> 00:07:28,950
note that on all the cloud 
interfaces you should have an 

155
00:07:28,950 --> 00:07:31,320
option of just either having a 
skills creator. 

156
00:07:31,520 --> 00:07:34,000
Or just telling. 
Claude, create all of this 

157
00:07:34,000 --> 00:07:36,960
process as a skill and review it
and collaborate with it. 

158
00:07:36,960 --> 00:07:38,280
Collaborate on this skill with 
me. 

159
00:07:39,120 --> 00:07:41,840
So you can actually tell it to 
turn things you've been working 

160
00:07:41,840 --> 00:07:45,720
on into a skill, then review it.
But I think for the purposes of 

161
00:07:46,040 --> 00:07:48,640
this discussion, it's still 
important to understand how to 

162
00:07:48,640 --> 00:07:51,080
create one because it'll give 
you an in depth understanding of

163
00:07:51,080 --> 00:07:53,280
what the heck this thing is. 
So step one would be create a 

164
00:07:53,280 --> 00:07:54,680
new. 
Folder on the computer, call it 

165
00:07:54,680 --> 00:07:59,720
weekly update or tell Claude to 
create a folder and call it 

166
00:07:59,720 --> 00:08:01,840
weekly update. 
Tell it to organise. 

167
00:08:01,840 --> 00:08:03,240
It as it. 
Sees fit or it. 

168
00:08:03,240 --> 00:08:05,720
Makes sense. 
Remember the core Co work 

169
00:08:05,720 --> 00:08:07,800
episode we've done I think about
5 weeks back? 

170
00:08:08,480 --> 00:08:11,240
If you haven't listened to that,
I'd recommend also listening to 

171
00:08:11,240 --> 00:08:16,040
that alongside this episode. 
Anyway, Step 2 would be to go 

172
00:08:16,040 --> 00:08:17,280
inside that folder, create a 
new. 

173
00:08:17,280 --> 00:08:20,600
File and it's a text file, you 
name it skill dot MD. 

174
00:08:20,680 --> 00:08:23,400
All caps for skill then dot MD. 
For the extension. 

175
00:08:23,560 --> 00:08:25,640
Step three. 
You open that file and at the 

176
00:08:25,640 --> 00:08:28,120
very top you're going to give. 
Cloud 2 piece of information, 

177
00:08:28,320 --> 00:08:31,960
you name the name of the skill 
and a one line simple 

178
00:08:31,960 --> 00:08:33,320
description of what it does and 
that's. 

179
00:08:33,320 --> 00:08:36,520
It below that you write actual. 
Instructions for example. 

180
00:08:37,000 --> 00:08:40,080
Write a concise weekly team 
update following our style of 

181
00:08:40,080 --> 00:08:43,280
format. 
Use when the user asks for help 

182
00:08:43,280 --> 00:08:46,120
with a weekly update or. 
Stand up summary or team status 

183
00:08:46,120 --> 00:08:47,520
report. 
Now notice that description. 

184
00:08:47,520 --> 00:08:49,440
I've included a few trigger 
phrases. 

185
00:08:49,720 --> 00:08:53,480
Weekly update stand up. 
Summary team status report. 

186
00:08:54,080 --> 00:08:56,880
So this gives Claude multiple 
ways to recognize when this 

187
00:08:56,880 --> 00:08:59,920
skill is needed. 
So if I then in a conversation 

188
00:08:59,920 --> 00:09:03,440
say I am writing this team 
status report, it would then 

189
00:09:03,440 --> 00:09:06,000
pull up and use this. 
This tool or this this? 

190
00:09:06,000 --> 00:09:08,920
Skill Step 4, you would then 
write the body, which is the 

191
00:09:08,920 --> 00:09:11,000
instructions. 
So you'd call it a heading 

192
00:09:11,000 --> 00:09:14,960
weekly team update format. 
And then when asked to write or 

193
00:09:14,960 --> 00:09:18,560
help with a weekly team update, 
all of these rules and then you 

194
00:09:18,560 --> 00:09:20,640
set the rules out. 
You can also have areas called 

195
00:09:20,640 --> 00:09:22,920
structure, you know open with a 
section called wins of the. 

196
00:09:22,920 --> 00:09:24,320
Week. 
Containing 2. 

197
00:09:24,320 --> 00:09:26,000
To four bullet. 
Points of complete work. 

198
00:09:26,160 --> 00:09:30,200
You could have tone so don't use
AI jargon and weird corporate 

199
00:09:30,200 --> 00:09:32,240
pillar. 
Keep it in first person. 

200
00:09:33,040 --> 00:09:37,440
You say LEMP, keep it under 300 
words and then add a section 

201
00:09:37,440 --> 00:09:40,720
called gotchas. 
So never use phrases like 

202
00:09:40,720 --> 00:09:42,760
moving. 
Forward or synergy or circle? 

203
00:09:42,760 --> 00:09:46,720
Back don't sign off saying kind 
regard because this isn't an 

204
00:09:46,720 --> 00:09:48,600
e-mail. 
This is going into a Slack 

205
00:09:48,600 --> 00:09:50,960
channel. 
And that's what a skill file is.

206
00:09:51,560 --> 00:09:53,760
And the gotcha section is 
something Anthropic specific 

207
00:09:53,760 --> 00:09:56,880
calls out as one of the highest 
signal content in any set of 

208
00:09:56,880 --> 00:09:58,360
skills, One of the. 
Claude. 

209
00:09:58,360 --> 00:10:02,760
Code team members actually wrote
that in a in a blog post that 

210
00:10:02,760 --> 00:10:05,200
the highest value thing you 
could possibly do is build up a 

211
00:10:05,200 --> 00:10:08,320
gotcha section over time so 
every time Claude makes a 

212
00:10:08,320 --> 00:10:11,000
mistake you can add to it. 
The skill becomes a living 

213
00:10:11,000 --> 00:10:13,960
document that essentially 
accumulates your hard won 

214
00:10:13,960 --> 00:10:16,200
knowledge about what goes wrong 
and how to prevent it. 

215
00:10:16,360 --> 00:10:19,200
So every time you're having a 
conversation with Claude, say I 

216
00:10:19,200 --> 00:10:23,640
want you to update the skill dot
MD file for XYZ skill with. 

217
00:10:23,640 --> 00:10:26,200
These new gotchas and. 
Again, you're just building up. 

218
00:10:26,200 --> 00:10:27,880
Constant information and 
knowledge on this. 

219
00:10:28,000 --> 00:10:29,840
And then finally, obviously you.
Save the file. 

220
00:10:30,600 --> 00:10:32,360
So that means you've now got a. 
Folder called weekly. 

221
00:10:32,360 --> 00:10:35,320
Update with a file called skill 
dot MD inside it. 

222
00:10:35,520 --> 00:10:37,640
You can then install it. 
And this depends on how you use 

223
00:10:37,640 --> 00:10:39,400
calls. 
If you're using cloud dot AI, so

224
00:10:39,400 --> 00:10:43,480
the web interface, you zip that 
folder, not the file, the actual

225
00:10:43,480 --> 00:10:46,320
folder itself. 
Right click the folder, compress

226
00:10:46,320 --> 00:10:48,640
it into a zip and then go to 
cloud AI. 

227
00:10:48,640 --> 00:10:51,320
And if you go into settings, you
can actually. 

228
00:10:51,320 --> 00:10:54,040
Go into customize and then 
skills and then you can upload 

229
00:10:54,040 --> 00:10:56,960
the skill using that zip file. 
If you are using the web 

230
00:10:56,960 --> 00:11:00,160
interface, which I recommend no 
one do, you should be using the 

231
00:11:00,160 --> 00:11:03,760
Co work interface or core code. 
But if you are using that you 

232
00:11:03,760 --> 00:11:06,680
need to go and turn on code 
execution and file creation 

233
00:11:06,800 --> 00:11:08,520
skills. 
Kind of need that to function if

234
00:11:08,520 --> 00:11:10,040
you're using. 
The core code interface. 

235
00:11:10,320 --> 00:11:13,040
It's really easy, you just tell 
it to do it and it will bash 

236
00:11:13,040 --> 00:11:15,040
command it. 
It'll just set it up for you now

237
00:11:15,040 --> 00:11:16,440
if you are. 
Writing Organization. 

238
00:11:16,440 --> 00:11:19,880
Wide skills you'll probably need
to speak to your admin, but you 

239
00:11:19,880 --> 00:11:23,520
can even tell it to make this a 
personal skill or a team wide 

240
00:11:23,520 --> 00:11:26,280
skill if you're using clouds COA
which is the. 

241
00:11:26,280 --> 00:11:29,640
Desktop interface skills can 
come bundled in plug insurance 

242
00:11:29,640 --> 00:11:32,000
or just upload them through the 
same settings panel. 

243
00:11:32,400 --> 00:11:35,120
As Claude dot. 
AI as I said Claude Co work is 

244
00:11:35,120 --> 00:11:37,200
the same agent as Claude code 
so. 

245
00:11:37,440 --> 00:11:39,400
You can basically ask. 
It to just do it for you. 

246
00:11:39,560 --> 00:11:42,080
And they've just released a new 
feature for most people called 

247
00:11:42,080 --> 00:11:44,520
the Skills Creator, which just 
helps you create a. 

248
00:11:44,520 --> 00:11:46,320
Skill. 
So yeah, once you've done that, 

249
00:11:46,680 --> 00:11:49,720
every time you start a new 
conversation, Claude sees all 

250
00:11:49,720 --> 00:11:52,360
the different skills that it's 
got available to it and it's 

251
00:11:52,360 --> 00:11:53,960
descriptions. 
That's that layer one. 

252
00:11:54,480 --> 00:11:58,760
It costs almost nothing in 
context window and capacity and 

253
00:11:58,760 --> 00:12:00,640
then when you say something like
here are my. 

254
00:12:00,640 --> 00:12:03,080
Notes for this week's. 
You know, write up our update, 

255
00:12:03,080 --> 00:12:05,480
it will then load up that skill 
and it will work. 

256
00:12:05,640 --> 00:12:07,680
Now really quick, really quick 
note here. 

257
00:12:08,560 --> 00:12:10,640
It could probably be quite 
attractive to start turning 

258
00:12:10,640 --> 00:12:13,320
everything into skills when 
that's just not necessary. 

259
00:12:14,120 --> 00:12:16,400
So if it's a one off task, just 
use prompts. 

260
00:12:16,400 --> 00:12:17,920
If it's a repeatable task that 
you do a. 

261
00:12:17,920 --> 00:12:20,520
Lot or you need other. 
Team members to fulfill. 

262
00:12:20,760 --> 00:12:22,600
Then you can make a skill. 
And when you write. 

263
00:12:22,600 --> 00:12:24,000
The skills. 
There's a trap. 

264
00:12:24,840 --> 00:12:26,960
You don't need to. 
Reinstate what Claude as a 

265
00:12:26,960 --> 00:12:30,760
baseline model already knows. 
Anthropics guidance on this is 

266
00:12:30,760 --> 00:12:33,480
really really direct, so don't 
tell it how to write Python 

267
00:12:33,480 --> 00:12:35,160
because it knows. 
How to write Python? 

268
00:12:35,320 --> 00:12:37,600
Don't tell it what good code 
looks like in general, tell it 

269
00:12:37,600 --> 00:12:40,320
what good code looks like in 
your specific code base. 

270
00:12:40,480 --> 00:12:44,120
So again, it's about making it 
contextual and writing it out 

271
00:12:44,120 --> 00:12:46,080
your process that you want. 
This thing to follow and then 

272
00:12:46,080 --> 00:12:48,040
once you've built a skill, test 
it. 

273
00:12:48,160 --> 00:12:49,920
Don't just use it once and go, 
oh, that was good. 

274
00:12:50,080 --> 00:12:51,720
Use the skill. 
Creator's Evaluation. 

275
00:12:51,720 --> 00:12:53,520
Tools run the skill against 
different prompts. 

276
00:12:53,640 --> 00:12:56,160
And then don't. 
Make it too prescriptive so 

277
00:12:56,160 --> 00:12:57,720
there's a difference between. 
A skill and a. 

278
00:12:57,720 --> 00:13:00,800
Template templates. 
Break the moment. 

279
00:13:01,200 --> 00:13:03,480
The situation changes. 
We talked about this last week 

280
00:13:03,800 --> 00:13:06,120
and you'll notice this is all 
about reliability. 

281
00:13:06,120 --> 00:13:08,600
And consistency. 
If you write something that is 

282
00:13:08,600 --> 00:13:10,520
if this, then that. 
It's. 

283
00:13:10,800 --> 00:13:11,920
Going to be. 
Quite rigid. 

284
00:13:12,200 --> 00:13:15,240
And therefore. 
If you need to be flexible, it's

285
00:13:15,240 --> 00:13:16,440
not going to work in the same 
way. 

286
00:13:16,960 --> 00:13:19,800
So again, Quad and Anthropica 
are very clear here. 

287
00:13:20,400 --> 00:13:22,480
It's about a skill. 
Not a template. 

288
00:13:22,680 --> 00:13:25,960
So yeah, that that's how to make
a skill and when not to make a 

289
00:13:25,960 --> 00:13:27,560
skill. 
But I also wanted to talk about 

290
00:13:27,560 --> 00:13:30,000
how to make skills extremely 
powerful. 

291
00:13:30,400 --> 00:13:32,640
So if you imagine, that's kind 
of like getting started. 

292
00:13:33,040 --> 00:13:36,280
This is now moving into. 
Advanced thinking, not because 

293
00:13:36,280 --> 00:13:40,080
it's super complex, but because 
it makes such an extra impact on

294
00:13:40,080 --> 00:13:42,400
how you work. 
And that's skill chains. 

295
00:13:56,400 --> 00:13:59,480
So yeah, the next level is when 
skills start working together, 

296
00:13:59,480 --> 00:14:01,840
and I'm going to give you an 
example of four skills that feed

297
00:14:01,840 --> 00:14:04,240
into each other to form a 
complete development process 

298
00:14:04,240 --> 00:14:07,080
from a vague idea all the way. 
Through to tested ship code. 

299
00:14:07,520 --> 00:14:09,720
And it came from a developer 
called Map Poco. 

300
00:14:10,080 --> 00:14:12,480
And you might be thinking, oh, 
but I'm not a developer now. 

301
00:14:13,240 --> 00:14:15,920
I'll also provide a couple of 
extra examples here about. 

302
00:14:16,320 --> 00:14:18,240
Non developer or product focused
tasks. 

303
00:14:18,240 --> 00:14:20,680
So the first skill is called. 
Grooming. 

304
00:14:21,480 --> 00:14:23,240
And I'll read it out in full 
because I think it's important 

305
00:14:23,240 --> 00:14:25,800
to hear just how simple a skill 
can be. 

306
00:14:26,280 --> 00:14:28,240
The entire skill body is 3 
sentences. 

307
00:14:28,560 --> 00:14:30,320
Interview me. 
Relentless about every. 

308
00:14:30,320 --> 00:14:31,600
Aspect of this. 
Plan until we. 

309
00:14:31,600 --> 00:14:32,920
Reach a shared. 
Understanding. 

310
00:14:33,320 --> 00:14:35,440
Walk down every single branch. 
Of the decision. 

311
00:14:35,440 --> 00:14:38,960
Tree resolve any dependencies 
between decisions 1 by 1. 

312
00:14:39,000 --> 00:14:42,520
And finally, if a question can 
be answered by exploring the 

313
00:14:42,520 --> 00:14:44,840
code base, explore the code base
instead, so. 

314
00:14:44,960 --> 00:14:46,600
Three sentences and that's the 
whole thing. 

315
00:14:46,920 --> 00:14:49,960
And he said Sessions where I 
called asked him 30-40 fifty 

316
00:14:49,960 --> 00:14:53,600
questions from that tiny skill 
before it was satisfied that 

317
00:14:53,600 --> 00:14:56,680
they understood the problem. 
And you might be thinking, gosh,

318
00:14:56,720 --> 00:14:59,120
that's a lot of typing. 
But if you've not got super. 

319
00:14:59,120 --> 00:15:02,240
Wispy yet? 
I recommend you to go and pay 

320
00:15:02,600 --> 00:15:04,280
just a few pounds a month. 
I think they. 

321
00:15:04,280 --> 00:15:08,040
Also have a pre tier but it's 
just an amazing basically voice 

322
00:15:08,040 --> 00:15:10,800
tool that you can press a 
shortcut, it appears and when 

323
00:15:10,800 --> 00:15:13,840
you're finished speaking the 
text appears in any text box. 

324
00:15:14,160 --> 00:15:15,920
It makes my life infinitely 
easier. 

325
00:15:16,760 --> 00:15:17,760
But anyway. 
And then? 

326
00:15:17,760 --> 00:15:18,880
Once the grilling. 
Is done. 

327
00:15:18,880 --> 00:15:21,360
It'll invoke a second skill, 
write a PRD. 

328
00:15:21,840 --> 00:15:24,880
So this is a structured multi 
step skill and apid is a 

329
00:15:24,880 --> 00:15:26,760
product. 
Requirement document says. 

330
00:15:27,560 --> 00:15:30,880
What people in product write. 
Before they build something 

331
00:15:31,080 --> 00:15:32,760
so-called. 
Will then start to ask the user 

332
00:15:32,760 --> 00:15:34,960
for a detailed. 
Description it'll explore. 

333
00:15:34,960 --> 00:15:38,600
The repo to verify any. 
Assertions it will interview. 

334
00:15:38,760 --> 00:15:40,960
And once all this is done, this.
Document's written. 

335
00:15:40,960 --> 00:15:44,480
It will also submit this as a 
GitHub issue, so actually raises

336
00:15:44,480 --> 00:15:46,520
it in the system of record to 
track work. 

337
00:15:46,640 --> 00:15:49,560
And So what comes out on the 
other end is a ticket or. 

338
00:15:49,960 --> 00:15:53,480
An item of a a task to be done. 
With a full document, with a 

339
00:15:53,480 --> 00:15:56,720
problem statement, a proposed 
solution, detailed stories. 

340
00:15:56,720 --> 00:15:59,360
In the right format, 
implementation decisions. 

341
00:15:59,480 --> 00:16:01,720
And then it will. 
Trigger 1/3 skill called the PRD

342
00:16:01,720 --> 00:16:03,520
to issues. 
That takes the document and 

343
00:16:03,520 --> 00:16:05,520
breaks it. 
Into a series of individual. 

344
00:16:05,520 --> 00:16:08,560
Tasks each task. 
Is essentially A deliverable 

345
00:16:08,560 --> 00:16:10,240
piece of work that the team can 
work on. 

346
00:16:10,480 --> 00:16:13,720
For example, it will start with.
What is most important to start 

347
00:16:13,720 --> 00:16:15,120
on? 
Usually the things with the 

348
00:16:15,120 --> 00:16:18,120
highest risk to a project. 
The skill will also establish. 

349
00:16:18,120 --> 00:16:20,000
Blocking relationships between. 
Tasks. 

350
00:16:20,200 --> 00:16:23,200
So task 3 might depend on task 
one, but task 2 might be 

351
00:16:23,200 --> 00:16:25,360
independent and should be 
therefore picked up in parallel.

352
00:16:25,360 --> 00:16:28,000
So this is really useful if 
you're running many different 

353
00:16:28,000 --> 00:16:30,080
agents at the same time. 
For example, this is what a 

354
00:16:30,080 --> 00:16:31,760
product manager and a developer 
has to be. 

355
00:16:31,880 --> 00:16:34,520
Finally, when it's time to 
implement, there's a Test Driven

356
00:16:34,520 --> 00:16:37,280
Development skill. 
So basically is a process that 

357
00:16:37,280 --> 00:16:40,560
enforces and confirms with the 
developer what interface. 

358
00:16:40,560 --> 00:16:42,320
Changes are going to be needed. 
Agreed. 

359
00:16:42,320 --> 00:16:44,840
Therefore, which behaviors we 
have to test and then enter a 

360
00:16:44,840 --> 00:16:47,600
loop and you can't? 
Have that thing finished until 

361
00:16:47,600 --> 00:16:50,520
it achieves those sets of 
behaviours that you were looking

362
00:16:50,520 --> 00:16:52,120
to. 
Actually achieve so. 

363
00:16:52,120 --> 00:16:54,640
That's a very. 
Product centric example that the

364
00:16:54,640 --> 00:16:56,560
take away here for non 
developers as well is that. 

365
00:16:56,800 --> 00:16:59,240
This principle. 
Of chaining works regardless of 

366
00:16:59,240 --> 00:17:01,200
domain, so and even that 
process. 

367
00:17:01,200 --> 00:17:03,320
We went through. 
You could build a research 

368
00:17:03,320 --> 00:17:04,800
brief. 
Skill that feeds into a draft. 

369
00:17:04,800 --> 00:17:08,280
Article skill that feeds into a 
format for newsletter skill. 

370
00:17:08,520 --> 00:17:12,280
The exact same architecture, 
discovery, planning, execution, 

371
00:17:13,160 --> 00:17:15,400
each skill picking up where the 
last one left on. 

372
00:17:15,520 --> 00:17:17,319
And this mechanism is domain 
agnostic. 

373
00:17:17,760 --> 00:17:20,240
The instructions inside is what 
makes. 

374
00:17:20,240 --> 00:17:22,920
It specific to your work. 
Now you might be thinking, gosh,

375
00:17:22,920 --> 00:17:25,720
it's really exciting, but maybe 
a lot to get. 

376
00:17:25,720 --> 00:17:28,160
Started on here. 
Well, there are places for you 

377
00:17:28,160 --> 00:17:31,800
to go find other skills, bring 
them in and get inspiration. 

378
00:17:32,280 --> 00:17:35,600
I will have one warning. 
Which is obviously when you're 

379
00:17:35,600 --> 00:17:38,600
downloading things from the 
Internet that are just very 

380
00:17:38,600 --> 00:17:41,440
public and inside a community. 
It does present risk. 

381
00:17:41,440 --> 00:17:43,480
So if you bring that straight 
into your. 

382
00:17:44,080 --> 00:17:46,920
Your environment that you're 
doing coding from or whatever it

383
00:17:46,920 --> 00:17:47,440
might. 
Be. 

384
00:17:47,760 --> 00:17:50,040
You need to make sure that 
you're going and getting trusted

385
00:17:50,040 --> 00:17:52,400
verified sources. 
So just just a. 

386
00:17:52,600 --> 00:17:54,960
A preface that. 
Super popular ones, probably 

387
00:17:54,960 --> 00:17:57,960
more say, but going on random 
websites and finding skills and 

388
00:17:57,960 --> 00:18:00,280
downloading them. 
Will always present some risk. 

389
00:18:00,360 --> 00:18:03,880
So anyway, Anthropic has got an 
entire skill set of a skill 

390
00:18:03,880 --> 00:18:07,160
library, should I say, on the 
GitHub reposogithub.com 

391
00:18:07,160 --> 00:18:08,480
Anthropic. 
Skills. 

392
00:18:09,280 --> 00:18:11,800
You know there's skills by 
creating word docs, excel 

393
00:18:11,800 --> 00:18:13,560
spreadsheets. 
PowerPoint presentations. 

394
00:18:13,560 --> 00:18:15,640
PDS. 
Beyond Anthropic, the community 

395
00:18:15,640 --> 00:18:18,160
is massive now. 
There are curated collections on

396
00:18:18,160 --> 00:18:20,280
GitHub with 10s of thousands of 
stars. 

397
00:18:20,560 --> 00:18:23,200
There are directories listing. 
Thousands of community skills 

398
00:18:23,200 --> 00:18:25,680
and there's even growing 
ecosystems of plug in. 

399
00:18:26,440 --> 00:18:29,680
Particularly in core Co work and
for those that don't understand 

400
00:18:30,240 --> 00:18:32,200
the difference. 
Between skills, MCPS and plug 

401
00:18:32,200 --> 00:18:34,440
insurance. 
There's often a lot of confusion

402
00:18:34,440 --> 00:18:36,600
here. 
Think of it in three layers. 

403
00:18:37,200 --> 00:18:39,960
Skills are the knowledge layer, 
so they tell Claude how to do a 

404
00:18:39,960 --> 00:18:40,440
thing. 
So a. 

405
00:18:40,440 --> 00:18:43,120
Process or a preference. 
A skill doesn't connect Claude 

406
00:18:43,120 --> 00:18:45,040
to anything in the external 
world. 

407
00:18:45,440 --> 00:18:48,640
Just makes it smarter about a. 
Particular task MCPS. 

408
00:18:48,960 --> 00:18:52,000
To model context protocol are 
the connection layer. 

409
00:18:52,440 --> 00:18:54,840
They let Claude reach out into 
your. 

410
00:18:54,840 --> 00:18:57,360
Slack your GitHub repo. 
Your CRM your. 

411
00:18:57,360 --> 00:18:59,960
Calendar. 
And plug insurance are basically

412
00:18:59,960 --> 00:19:00,720
the. 
Packaging layer. 

413
00:19:01,320 --> 00:19:05,720
A plug in bundles skills, MCPS, 
other resources together that 

414
00:19:05,720 --> 00:19:07,680
you can just then install in a 
single gap. 

415
00:19:08,160 --> 00:19:11,200
So instead of separately 
configuring a Slack MCP, writing

416
00:19:11,200 --> 00:19:14,200
a skill for how your team uses 
Slack, and then setting all that

417
00:19:14,200 --> 00:19:16,760
up manually, you can just. 
Create a plug in that does all 

418
00:19:16,760 --> 00:19:19,040
of that together. 
Skills of the brain, MCPS of the

419
00:19:19,040 --> 00:19:21,560
hands and. 
Plug insurance the toolbox, and 

420
00:19:21,560 --> 00:19:23,480
the best thing here is that 
skills aren't just an. 

421
00:19:23,480 --> 00:19:24,920
Anthropic thing. 
It was published. 

422
00:19:24,920 --> 00:19:26,720
As an open standard in December 
2020. 

423
00:19:26,720 --> 00:19:28,960
Five and. 
Since then it's been adopted by 

424
00:19:28,960 --> 00:19:33,240
loads of tools like Open AI 
codecs, GitHub Copilot VS Co 

425
00:19:33,240 --> 00:19:35,920
Cursor, Gemini. 
They're all supporting the same 

426
00:19:35,920 --> 00:19:37,600
format. 
It's one of the fundamental ways

427
00:19:37,600 --> 00:19:38,720
that would build a. 
Lot of new. 

428
00:19:38,720 --> 00:19:40,600
Functionality now. 
On our product. 

429
00:19:40,760 --> 00:19:42,640
Aligning ourselves with that 
ecosystem. 

430
00:19:59,390 --> 00:20:02,390
So yeah, let's conclude for 
today's episode because here's 

431
00:20:02,390 --> 00:20:04,630
the bigger picture. 
We're in this massive 

432
00:20:04,630 --> 00:20:06,030
transition. 
From using AI. 

433
00:20:06,030 --> 00:20:07,470
As a conversation. 
To using. 

434
00:20:07,470 --> 00:20:10,710
It as a library of reliable. 
Repeated. 

435
00:20:10,710 --> 00:20:13,870
Capabilities and every time you 
figure out how to get one of 

436
00:20:13,870 --> 00:20:16,110
these things, these systems, 
these large language models to 

437
00:20:16,110 --> 00:20:19,120
do something well, you can 
capture that knowledge. 

438
00:20:19,120 --> 00:20:21,800
So it compounds over time, and 
that's what skills are 

439
00:20:22,320 --> 00:20:25,720
compounding information. 
That you and your team can 

440
00:20:25,720 --> 00:20:28,280
leverage over time. 
Skills turn. 

441
00:20:28,280 --> 00:20:30,240
One person's expertise into 
everyone's. 

442
00:20:30,240 --> 00:20:33,840
Capability your best. 
Engineers review process 

443
00:20:33,840 --> 00:20:36,200
becomes. 
A skill the whole team can run. 

444
00:20:36,400 --> 00:20:40,080
Your best writer can be the 
foundation to how every person 

445
00:20:40,080 --> 00:20:42,600
in your company writes content. 
So if there's one thing to take 

446
00:20:42,600 --> 00:20:45,800
from this episode, it's this The
next time you find yourself 

447
00:20:45,800 --> 00:20:48,560
giving called the same 
instructions the third time in a

448
00:20:48,560 --> 00:20:52,440
row, just stop and write skill. 
We'll look to link everything 

449
00:20:52,440 --> 00:20:54,800
into the show notes. 
You know, documentation skills 

450
00:20:54,800 --> 00:20:57,640
repos basically as much as. 
Possible for you to. 

451
00:20:57,640 --> 00:21:00,440
Get started on. 
Skills anyway. 

452
00:21:00,720 --> 00:21:03,680
That's it for me. 
I hope you enjoyed the episodes.

453
00:21:04,320 --> 00:21:05,840
And I'll see you next week.