New Skills! v1.2 brings /wait-what, /writing-for-agents, and fixes /grill-me — Transcript

Full transcript

1. 0:00Skills, skills, skills, folks. We have
2. 0:02yet more skills for you. This is version
3. 0:041.2.0,
4. 0:06which adds a bunch of improvements to my
5. 0:08already extremely popular engineering
6. 0:10skills. The skills repo is now, I think,
7. 0:13the 24th most starred repo of all time
8. 0:15on GitHub up to 204K stars. And as a
9. 0:18mature project, we need some proper
10. 0:20documentation, and so we have it. This
11. 0:23is at aihero.dev/skills,
12. 0:25and I've been working with my team to
13. 0:27produce this absolutely wonderful site.
14. 0:30The best way to navigate this is by
15. 0:31going to the main page, /skills, and
16. 0:34then looking at the kind of groupings
17. 0:36that we've got here. For instance, we
18. 0:37can see here the main flow, where you
19. 0:39start with grill with docs, then go to
20. 0:40to spec, then to ticket, implement, and
21. 0:42then code review. In the left-hand panel
22. 0:44here, you've got a full reference of
23. 0:45every single skill and what they do. It
24. 0:48used to be my actual uh full-time job to
25. 0:50write documentation, so I was pretty
26. 0:51happy to go back to it. There's even a
27. 0:53list of common questions, which are
28. 0:55actually sourced from a personal wiki
29. 0:57that I maintain of all the questions
30. 0:59people ask me. These skills are also
31. 1:01linked to an AI coding dictionary that
32. 1:03I've been putting together. For
33. 1:04instance, this little ticket link here
34. 1:06is the dictionary definition for what I
35. 1:08think of as a ticket. So, these docs,
36. 1:10not only can you use them to learn my
37. 1:13skills better, you can also use them
38. 1:14just to learn about AI coding and
39. 1:16understand how things work at a deep
40. 1:18level. The second big change is we are
41. 1:20now an official part of the Claude Code
42. 1:22official plugin marketplace. What that
43. 1:24means is you can open Claude Code, you
44. 1:26can say plugin like this, you can search
45. 1:29for Matt Pocock here, or Matt Pocock
46. 1:31skills, and then you can install it
47. 1:33right from here. This means you end up
48. 1:35with a read-only bundle. There's no
49. 1:37extra steps here, and any updates I make
50. 1:41to the skills will be automatically
51. 1:42pulled down to you. That sounds pretty
52. 1:44good if you're a Claude Code user, but
53. 1:45what if you're a Codex user? What's in
54. 1:47this release for you? Well, I now
55. 1:49include open ai.yaml files for every
56. 1:53single skill, which means that it should
57. 1:56properly work out of the box with UIs
58. 1:58like Codex UI. And also, it means that
59. 2:00this allow implicit invocation false
60. 2:03travels over to Codex. Before, we had
61. 2:05model invoked skills and user invoked
62. 2:07skills. And the benefit of user invoked
63. 2:09skills, which we can see inside the main
64. 2:11skill.md file, is these are hidden from
65. 2:14the agent's context window until you
66. 2:16invoke them. That works for Claude Code,
67. 2:18it works for Pi, it works for a couple
68. 2:20of other harnesses, but it doesn't work
69. 2:21for Codex. I didn't quite realize that.
70. 2:24And so, now we ship these sidecar with
71. 2:26every single skill to mean that we have
72. 2:28this allow implicit invocation false, so
73. 2:30that works out of the box with Codex,
74. 2:32too. The next thing I wanted to ship a
75. 2:33skill for is that Opus, especially Opus
76. 2:365, is talking garbage at the moment. I
77. 2:39don't understand what it is. I think it
78. 2:41might just be a new model thing. But,
79. 2:43for some reason, every time I interact
80. 2:45with Opus, I it just goes right over my
81. 2:47head. It's incredibly verbose, it uses
82. 2:49really weird LLM phrases, and I just
83. 2:53find it really, really hard to read, and
84. 2:54a lot of people will feel the same. The
85. 2:55honest options are to accept it and
86. 2:57reword the AC or to persist the digest
87. 2:59along the export, so a resumed unchanged
88. 3:01publish can skip the read to flagging
89. 3:03rather than picking. What does that
90. 3:04mean? It is genuinely load-bearing,
91. 3:06though, so that's how you know it
92. 3:07matters. I've tried all sorts of ways to
93. 3:09fix it, such as output styles or adding
94. 3:11things to my agent.md, but what I
95. 3:13realized I needed was just a skill to
96. 3:15say, "Wait, what did you just say?" And
97. 3:17so, the skill is called Wait What. It's
98. 3:19a very, very simple skill that just does
99. 3:22a couple of things. First, it says, "Use
100. 3:24this specific standard ASD STE 100
101. 3:28Simplified Technical English," which is
102. 3:29essentially just a leading word to the
103. 3:31agent to say, "Use very simple language.
104. 3:34Speak in clear declarative sentences."
105. 3:36And secondly, it also tells it to ground
106. 3:38itself in the ubiquitous language in
107. 3:41context.md. The real cure for verbosity
108. 3:44is not to tell it to use simple
109. 3:46language, although we are doing a bit of
110. 3:47that, is to tell it to use your
111. 3:49language, the stuff that you have come
112. 3:51up with in Grill with Docs. You use this
113. 3:53skill whenever the agent just creates
114. 3:55some random garbage and you've no idea
115. 3:57what they just said. You just say,
116. 3:58"Wait, what?" and it should reply with a
117. 4:00much better alternative. The next thing
118. 4:02that's changed is a pretty big one. It's
119. 4:04an update to my most popular skill,
120. 4:06which is Grill me. Previously, Grill me
121. 4:08used to grill you one question at a
122. 4:11time. In other words, you'd get a
123. 4:12question and then you would respond and
124. 4:14then another question, then you would
125. 4:15respond. So, one question per turn.
126. 4:18However, this has a horrible failure
127. 4:20mode, which is when you get to the end
128. 4:22of a grilling session and most of the
129. 4:24hard stuff is already done, you just
130. 4:25have a bunch of easy questions, one per
131. 4:28turn, that you're just basically saying,
132. 4:30"Yeah, that sounds good. Yeah, that
133. 4:32sounds good, too." This is incredibly
134. 4:33frustrating and it feels dead slow. So,
135. 4:36for a while, I've wanted to find a way
136. 4:38to speed this up. The thing that makes
137. 4:39most sense to me is to have multiple
138. 4:41questions per turn. So, if you have lots
139. 4:44of easy ones, you can just say, "Yes,
140. 4:46yes, yes, yes, yes." But, what happens
141. 4:48when the answer to question three
142. 4:50depends on the answer in question one?
143. 4:53Before, you never had that problem
144. 4:55because you would just answer them one
145. 4:57at a time. So, if a question depended on
146. 4:59an earlier question, then that one had
147. 5:01already been answered. But, in this new
148. 5:02setup, surely that's more confusing
149. 5:04because you're going to be answering
150. 5:05erroneous questions or questions at the
151. 5:07wrong time or questions that need to be
152. 5:09answered before other questions. How is
153. 5:10that going to work? Well, if you think
154. 5:11about it, these questions are really
155. 5:13like a graph. There might be one
156. 5:15critical question at the very start that
157. 5:17needs answering first that opens up a
158. 5:19whole raft of other questions, and maybe
159. 5:21they lead on to other questions
160. 5:23themselves. And so, now Grill me is
161. 5:25designed to follow this graph, to answer
162. 5:28only the questions that are available
163. 5:29now, but to ask them in rounds. And so,
164. 5:32you get the first round of questions,
165. 5:34which might be as little as one critical
166. 5:36question. Then, you go to the next round
167. 5:38of the questions that opens up, and it's
168. 5:40always pushing you as fast as possible
169. 5:43down the frontier of questions. And this
170. 5:45is what it looks like. You can see I'm
171. 5:46at round one here, and we have Q1, and
172. 5:50then we get this little recommendation
173. 5:52here. I've decided to use I know I know
174. 5:55you probably don't like it, but I love
175. 5:57it, emojis here for a little pop of
176. 5:59color. This makes it super easy to
177. 6:01navigate with my eyes. I can see, okay,
178. 6:03this is Q1, and then I can see the
179. 6:05recommended answer. Q2, see the
180. 6:07recommended answer. And because I'm
181. 6:08using dictation, I just get to blast
182. 6:10out, okay, Q1, I agree. Q2, I agree. Q3,
183. 6:14we need something to change there. Q4,
184. 6:16we need something to change. And then
185. 6:17further down, we can see round two came
186. 6:19in and the same setup. So, we answered
187. 6:22the first round of questions, and then
188. 6:24we got another round of questions, and
189. 6:26we kept pushing the frontier back. The
190. 6:27next big change is that I've modified my
191. 6:30writing great skills skill, because I
192. 6:32noticed that I was using it for stuff
193. 6:34other than skills. I was using it for
194. 6:37anything that an agent read to make the
195. 6:40agent better, to make the doc more
196. 6:41concise, and to make the output a bit
197. 6:43more predictable. And so, now we have
198. 6:45writing for agents, where you can modify
199. 6:47it for agents.md or Claude.md, or when
200. 6:50you're creating editing skills. And it
201. 6:52has some extra skill mechanics stapled
202. 6:55on inside this reference file. So, you
203. 6:57can use it for writing skills, you can
204. 6:59use it crucially for taking your
205. 7:01agents.md, pulling it out into skills to
206. 7:05stop that horrible front loading. So,
207. 7:07overall, I find myself using this skill
208. 7:08so so often whenever I need to write any
209. 7:11agent configuration, and I want it to
210. 7:13read well, I want the agent to perform
211. 7:15well with it, writing for agents is the
212. 7:17right call. And it's also
213. 7:19model-invocable as well, and it's a
214. 7:21really clean description here. So, it
215. 7:23should pull it in whenever it's
216. 7:25modifying agents.md, which is just
217. 7:27great. The next one is a new skill, one
218. 7:29that I've been keeping under wraps for a
219. 7:30little while. This is the wizard skill.
220. 7:33I found myself recently having to
221. 7:34provision a bunch of uh infrastructure
222. 7:36and walk through a bunch of AWS stuff,
223. 7:39and I I
224. 7:40hated it. And I thought, surely there's
225. 7:42a way that the agent can help me here.
226. 7:44And so, I created this wizard skill that
227. 7:46generates an interactive bash wizard
228. 7:48that walks the human through steps only
229. 7:50they can perform. In other words, I
230. 7:51probably could have got the agent to,
231. 7:53okay, go into AWS and provision the
232. 7:55stuff with computer use or something,
233. 7:57but it just felt pretty icky. I wanted
234. 8:00to have control over all of it, but I
235. 8:02wanted it to be as easy as possible. And
236. 8:04here is what one looks like. This was
237. 8:06for, I think, migrating to a remote box
238. 8:08setup. So, you just kick it off, and
239. 8:10then you say, "Ready to start?" Yes,
240. 8:12please. And then it sends me off to
241. 8:13exactly the page I need to go. That
242. 8:15opened the script, by the way. I didn't
243. 8:16need to do anything there. Sends me off,
244. 8:18gets me to log in, gets me to change the
245. 8:20exact thing we need to. I can then paste
246. 8:22in my API keys, and remember this is a
247. 8:24deterministic script, so nothing's
248. 8:26touching an agent here. It's not sending
249. 8:28it off to Anthropic or anything. So, I
250. 8:30paste it in here. It saves it into the
251. 8:33files it needs to. It's even saving them
252. 8:34into GitHub secrets if it needs to. And
253. 8:37it walks through these separate stages
254. 8:38with me, keeping going with this nice
255. 8:40little UI until we end up with what we
256. 8:43need. I found that this takes
257. 8:44provisioning services from incredibly
258. 8:46painful to weirdly joyful because I just
259. 8:48know how long it took me before. So,
260. 8:50wizard has become an essential part of
261. 8:53my skill set. I had it in in progress
262. 8:55for a couple of months, and I just
263. 8:56think, "Okay, it's time for you guys to
264. 8:57see it." The next one is certainly the
265. 8:59most boring skill name that I've ever
266. 9:01produced, but I found myself really,
267. 9:03really needing it, and I realized it
268. 9:05would be useful for more than what I had
269. 9:07just used it for. It is called to
270. 9:09questionnaire, and its purpose is that
271. 9:11you take the decisions inside a grilling
272. 9:13session, and you pull them out into a
273. 9:15document that you can walk through with
274. 9:17someone else. I was doing a wayfinder
275. 9:19session to try to build a garden office
276. 9:23in my back garden, and I realized that
277. 9:25the agent was grilling me, but the
278. 9:27person I really needed to speak to about
279. 9:28it was my wife because my wife,
280. 9:30obviously, was going to use this space,
281. 9:32too, and so we're going to design it
282. 9:33together. And so, I got the agent to
283. 9:35take the stuff it was going to ask in
284. 9:36the grilling session and turn it into a
285. 9:38markdown document. I then put that into
286. 9:40a Google Doc. I sent that over to my
287. 9:42wife and we walked through it together
288. 9:44and then we pulled the answers back into
289. 9:47the agents. Now, this is a skill that I
290. 9:49hope someday to delete because it's sort
291. 9:51of like a patch for the fact that agents
292. 9:54are kind of hard to collaborate with at
293. 9:56the moment. Some teams who have like
294. 9:58what I consider a really good setup is
295. 10:00they have the agent in Slack and they
296. 10:02have these conversations they can then
297. 10:03tag the agent in. So, they can
298. 10:05collaborate together, answer questions
299. 10:07together, especially with stakeholders
300. 10:09and then the agent can actually pile in
301. 10:11and go, "Okay, I'll implement it based
302. 10:12on that." But, a lot of people are not
303. 10:14in that world right now. A lot of people
304. 10:17are not AI native. Maybe they don't even
305. 10:18have Slack or Microsoft Teams or
306. 10:20something. And so, actually having a
307. 10:22document that you can just pull out,
308. 10:24give to someone, they can comment on it
309. 10:25on Google Docs or whatever is actually
310. 10:28really useful. And so, that folks is
311. 10:29basically it. We've got a tranche of new
312. 10:31skills and especially looking forward to
313. 10:34your reactions to wait, what? Because I
314. 10:36think it will actually save your ass a
315. 10:38couple of times when Opus is doing
316. 10:40strange things. There were a few things
317. 10:42that were just a bit too small to
318. 10:43mention here and so, you can go to
319. 10:45V1.2.0
320. 10:46in the GitHub releases to take a look at
321. 10:48the full change log. And if you're
322. 10:50digging this stuff, then I'm putting out
323. 10:51an AI coding crash course, which is
324. 10:54pretty different to the paid courses
325. 10:55that I've put out before. Instead of
326. 10:57being a cohort where you do it over a
327. 10:59couple of weeks and everyone's doing it
328. 11:01together, this is going to be a cheaper
329. 11:03course that's available throughout the
330. 11:04year. My personal opinion about AI
331. 11:06coding is that it's kind of settling
332. 11:08down a little bit. There's not actually
333. 11:11that much has changed since about
334. 11:13December last year and so, I now feel
335. 11:15confident enough that I can put out a
336. 11:17self-paced course that hopefully won't
337. 11:19change too much at even as things
338. 11:21develop. I'm hoping to give you a stable
339. 11:23foundation from which you can work and
340. 11:25actually ship amazing stuff. And the
341. 11:27link to this is down below and you can
342. 11:29join the waitlist. It should be coming
343. 11:30out only in a couple of weeks, I think.
344. 11:32So, thank you so much for watching.
345. 11:33Enjoy the new skills, and I will see you
346. 11:36very soon.