Create Your First Codex Skill: Turn a Workflow into SKILL.md — Transcript

Full transcript

1. 0:00Have you ever spent half a conversation
2. 0:04getting an agent to do something the way
3. 0:06you wanted?
4. 0:07You tried a few tools, fixed the output,
5. 0:10and finally you got a process that
6. 0:13worked. Then a new conversation comes
7. 0:15along and you're explaining the same
8. 0:17decisions again.
9. 0:19I've run into that making course
10. 0:21presentations.
11. 0:22So, the useful part of that conversation
12. 0:25was everything I learned about the
13. 0:27process and I want to keep that. Now,
14. 0:29when you interact with an agent, you're
15. 0:31often working through several steps of a
16. 0:33larger workflow.
17. 0:35For instance, I might be working on a
18. 0:36presentation for a training course I'm
19. 0:39developing.
20. 0:40There are many steps involved from
21. 0:42creating an outline to designing slides
22. 0:45and preparing speaker notes.
23. 0:48When I began asking the agent for help,
24. 0:50I found that some interactions went much
25. 0:53better than others.
26. 0:54A lot depended on how the agent went
27. 0:57about assisting me. Over many
28. 1:00interactions, I began to learn which
29. 1:02approaches worked best. Now, I want to
30. 1:05capture those guidelines, tool choices,
31. 1:08and actions so that the agent can reuse
32. 1:11that process.
33. 1:13That's what a skill lets me do.
34. 1:15Package a reusable workflow as
35. 1:17instructions with supporting files when
36. 1:20they're needed. Here's an example.
37. 1:23Often in my work, I have a video or an
38. 1:26audio recording of me explaining a
39. 1:28topic.
40. 1:29From that, I want to create a slideshow.
41. 1:31The slideshow is the ultimate goal, but
42. 1:34to get there, there are several steps
43. 1:37involved. One, transcribing the audio,
44. 1:40summarizing the key points, designing
45. 1:42the slides, and adding speaker notes.
46. 1:44One thing I learned in my own workflow
47. 1:46was that extracting audio from a video
48. 1:49before transcription gave me better
49. 1:51results.
50. 1:53And I found a tool named FFmpeg that did
51. 1:56that. I also learned that there are many
52. 1:58ways to transcribe audio. In some
53. 2:01conversations with the agent, it would
54. 2:03try several different approaches, fail,
55. 2:06and keep trying until it found something
56. 2:08that worked. I found that using Whisper
57. 2:11for transcription was the most reliable
58. 2:14approach for my needs. Now I can write a
59. 2:17script that extracts audio from video
60. 2:20and then transcribes it with Whisper.
61. 2:23That gives me a repeatable pipeline.
62. 2:26I'll still need to review the
63. 2:27transcription to correct errors or ask
64. 2:30the agent to do it. Summarizing key
65. 2:32points and designing slides require more
66. 2:35judgment. That's what I need the model
67. 2:37to do. And so when it comes to creating
68. 2:40the presentation, I have found I need to
69. 2:43provide a good bit of guidance on the
70. 2:46content of the slides as well as what
71. 2:48presentation application to use. Through
72. 2:51these iterations, I learned about tools
73. 2:54for creating different content and how
74. 2:56to direct the agent to produce content
75. 3:00and then verifying it. Capturing what
76. 3:02you know is the most critical piece.
77. 3:06Even if the model would ultimately come
78. 3:08up with an acceptable solution, your
79. 3:11knowledge can save time and money. So
80. 3:15when I tell the agent, "I have a video
81. 3:18where I talk about some topic in this
82. 3:20file,
83. 3:21file name.
84. 3:23Please create a slideshow for me.
85. 3:26I'm expecting it to follow the skill
86. 3:28I've developed. When that skill is
87. 3:31selected, the process is already there.
88. 3:33Extract the audio if needed, transcribe
89. 3:36it with Whisper, summarize the key
90. 3:38points, design the slides, add speaker
91. 3:41notes using the preferred presentation
92. 3:44tool.
93. 3:45The agent has my working tool choices
94. 3:48and options instead of having to
95. 3:51rediscover them. That can improve
96. 3:53efficiency and consistency and reduce
97. 3:56agent cost.
98. 3:58We still need to check the results, of
99. 3:59course,
100. 4:01but the goal is to reduce time and the
101. 4:03reduced cost. The AI client assembles
102. 4:06the model's context from your prompt,
103. 4:09the relevant conversation history,
104. 4:12the applicable instructions such as
105. 4:14agents.md,
106. 4:16and information about the available
107. 4:18tools and skills.
108. 4:21For skills, the model initially sees the
109. 4:24names and descriptions. In Codex, it
110. 4:27also sees their file paths.
111. 4:29When it decides to use a skill, it reads
112. 4:31that skill's skill.md file.
113. 4:35It can then read supporting references
114. 4:37or run scripts as needed.
115. 4:40That's what we call progressive
116. 4:41disclosure. The detailed instructions
117. 4:44and resources come into context only
118. 4:46when they're useful.
119. 4:48A task can use more than one skill.
120. 4:52The model continues working with these
121. 4:54instructions alongside the rest of the
122. 4:57context.
123. 4:58Selecting a skill isn't a guarantee that
124. 5:01every step will run correctly. So, in
125. 5:04Codex, start with an .agents folder at
126. 5:08the user or project level depending on
127. 5:10where you want the skill to be
128. 5:11available.
129. 5:13Other agent applications may use
130. 5:14different locations.
131. 5:16Next, create a subfolder named skills.
132. 5:19Within the skills folder, create a
133. 5:22separate folder for each skill and name
134. 5:25the folder the same name as the skill.
135. 5:28Use lowercase letters, numbers, and
136. 5:30hyphens with no spaces. Choose a name
137. 5:33that makes the task recognizable.
138. 5:36The description is where we make clear
139. 5:39when the agent should consider it. Now,
140. 5:42inside the skill folder, create a file
141. 5:44named skill.md.
142. 5:47This file will contain the documentation
143. 5:49and the instructions for the skill,
144. 5:51including steps, tools, and guidelines
145. 5:54the agent should follow when applying
146. 5:56the skill. The first part of the file is
147. 5:59YAML front matter. Between the lines of
148. 6:02three hyphens, it contains the skill's
149. 6:04name, matching the folder,
150. 6:07and a description. That description
151. 6:09tells the agent what the skill does and
152. 6:12when to consider using it. For automatic
153. 6:14selection, the skill's name and
154. 6:16description help the model decide
155. 6:18whether it's relevant. This is based on
156. 6:20meaning, so a request about making
157. 6:23slides from a video might match our
158. 6:25skill.
159. 6:26A match isn't guaranteed just because
160. 6:28the prompt contains certain keywords.
161. 6:31Once the skill is selected, the model
162. 6:34reads its full instructions. Then, it
163. 6:36uses any supporting resources it needs.
164. 6:39You can also explicitly name the skill
165. 6:42in your prompt.
166. 6:43In Codex CLI or the IDE extension, you
167. 6:46can use the skills selector or mention
168. 6:50the skill with a dollar sign. Other
169. 6:52clients use different selectors, so
170. 6:54check your local listings. So, here's a
171. 6:56a simple skill I created to change the
172. 7:00agent's conversational tone based on the
173. 7:03day of the week.
174. 7:04The workflow uses tool calls to find the
175. 7:07day and add a little randomness to the
176. 7:10mood. The description needs to match the
177. 7:12greetings I intend without matching
178. 7:16unrelated requests. The body of the file
179. 7:19provides the instructions I specifically
180. 7:22want the agent to use, for instance, two
181. 7:26PowerShell commands, get date and get
182. 7:28random.
183. 7:30The skill numbers Monday is one and
184. 7:33Sunday is seven, and then add some
185. 7:35random adjustment of minus one, zero, or
186. 7:38plus one. That gives us a mood index
187. 7:40from zero to eight. Zero is grumpy and
188. 7:44eight is cheerful and optimistic. The
189. 7:46numbers in between represent
190. 7:48intermediate moods.
191. 7:50So, when I tried this skill for the
192. 7:52first time, the agent would respond and
193. 7:56disclose the calculated index and say
194. 7:59things like, "Since it's Monday, I'm
195. 8:01supposed to be grumpy." So, I added an
196. 8:04instruction to keep the index out of the
197. 8:07greeting and instead use a persona that
198. 8:10fits the mood. When you submit a
199. 8:13personal greeting to the agent, it might
200. 8:15choose the skill to determine the
201. 8:17appropriate mood and respond accordingly
202. 8:20using a persona that matches that mood.
203. 8:23So, one Monday I I intentionally used a
204. 8:26happy upbeat greeting. The agent's
205. 8:29response acknowledged the good news, but
206. 8:31the tone was clearly dry and grumpy.
207. 8:35Then I tried a Friday. I did the
208. 8:38opposite with my own mood and gave it a
209. 8:40gloomy prompt. The agent responded in a
210. 8:43glass-half-full way. The reply looked
211. 8:46consistent with the skill, but the tone
212. 8:49alone doesn't prove that the skill ran.
213. 8:51I have to inspect the available tool
214. 8:54logs and
215. 8:55chat details to check whether it read
216. 8:57the skill and ran the required commands.
217. 9:00Now, my real goal wasn't the moody
218. 9:03greeting itself. I just wanted to show
219. 9:06whether the description led to skill
220. 9:08selection and whether the agent executed
221. 9:11the steps I'd defined.
222. 9:13The reply was a quick way to observe the
223. 9:16intended result. The execution record is
224. 9:20what lets us check the actual process.
225. 9:23We'll talk about testing skills in a
226. 9:25future lesson. So, for now though,
227. 9:28these examples give us a first check
228. 9:31rather than prove that the skill works
229. 9:33reliably. Now, you try. Create the
230. 9:35directory structure and a skill.md file.
231. 9:38Put some simple instructions and try
232. 9:41different prompts to see what triggers
233. 9:43the skill and what doesn't. Check which
234. 9:46instructions the agent actually
235. 9:48followed. Remember, this is
236. 9:50non-deterministic. Agents are still
237. 9:53using large language models. All you're
238. 9:55doing with the skill is augmenting the
239. 9:58context that goes to the large language
240. 10:01model. You're still following a
241. 10:03probabilistic
242. 10:05outcome. Say that fast three times. Now,
243. 10:07currently you might have a file full of
244. 10:09prompts that you copy and paste into a
245. 10:12conversation. You might even have
246. 10:14recipes where you drop several into the
247. 10:16dialogue sequentially. Well, that's a
248. 10:18that's a workflow. I was just talking to
249. 10:21a friend about this the other day, and
250. 10:23that's what she does. She just drops the
251. 10:25prompts in. I said, "Make a skill. Take
252. 10:28your knowledge and experience and turn
253. 10:31that into a skill. The next time you
254. 10:34have that task, you can just ask the
255. 10:36agent to use the process you've already
256. 10:39worked out. That saves you from feeding
257. 10:42it every step again, while you still
258. 10:45review the results and handle any
259. 10:47decisions that need your input. Now, in
260. 10:49our next lesson, we'll talk about the
261. 10:51best practices [music] in skill
262. 10:53creation. Hey, I'm Chuck McCollum, and
263. 10:55if you like this video, click like and
264. 10:57share your skill creation experiences in
265. 11:00the comments below. [music] I'd love to
266. 11:02hear from you, and I'll see you next
267. 11:05time.