YouTube2Text

Matt Pocock's EXACT System To 10x Your Claude Skills — Transcript

by Austin Marchese · 3,394 words · 496 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Mattacott has the most installed Claude
  2. 0:02skills on the planet with over 22
  3. 0:04million people using that. So I decided
  4. 0:06to do a deep dive in how exactly he
  5. 0:08creates his skills and the results
  6. 0:10shocked me. I went in expecting
  7. 0:11extremely complicated processes, but
  8. 0:13after analyzing every video he's ever
  9. 0:15posted and every skill he's created, I
  10. 0:17realized that he has a five-step process
  11. 0:19he uses to create every skill. But
  12. 0:21unfortunately, Matt's videos are
  13. 0:23extremely technical and the majority of
  14. 0:24his content is for engineers. So in this
  15. 0:26video, I'll break down each of the five
  16. 0:28steps he uses to create skills, simplify
  17. 0:31it and then tell you exactly how you can
  18. 0:32apply this process in your day-to-day to
  19. 0:35make your skills better and build 10
  20. 0:36times faster. And at the end of the
  21. 0:38video, I'll cover the reason why Matt
  22. 0:40thinks you should never blindly copy
  23. 0:42anyone else's skills. The step one is
  24. 0:44about notice and this is the thing that
  25. 0:46you keep sending back. Now, I've heard a
  26. 0:48lot of breakdowns on what a Claude skill
  27. 0:50actually is, right? It's a reusable
  28. 0:51prompt. It's a process of teaching
  29. 0:53Claude something new and these aren't
  30. 0:55wrong, but Matt has one of my favorite
  31. 0:57definitions that I've come across. A
  32. 0:58skill is a correction you got tired of
  33. 1:01making. Here's someone on podcast
  34. 1:02explaining exactly this.
  35. 1:03>> Uh I went to a restaurant and I ordered
  36. 1:07a chorizo salad and I got a chorizo
  37. 1:09sandwich. Right? So I had to send it
  38. 1:12back. And what I was finding is I was
  39. 1:14just having to send things back all the
  40. 1:16time to AI.
  41. 1:16>> Said another way, the key reason to
  42. 1:18create a skill isn't what you do most
  43. 1:19often with Claude. It's what you keep
  44. 1:21having to correct. If the model can
  45. 1:23already complete a task correctly the
  46. 1:24first time, a skill for that is just
  47. 1:26unnecessary bloat. And that's actually
  48. 1:28part of the reason why the creator of
  49. 1:29Claude code told people to go delete all
  50. 1:31of their Claude skills. If Claude
  51. 1:32already gets it right, the skill isn't
  52. 1:34doing anything except costing tokens.
  53. 1:36With Matt's framework, he's essentially
  54. 1:38saying if Claude doesn't get it right
  55. 1:39the first time, send it back. And this
  56. 1:41kind of remind me of that scene in The
  57. 1:43Office where he says, "I'll have the
  58. 1:44spaghetti with a side salad. If the
  59. 1:46salad is on top, I'll send it back."
  60. 1:48It's that same Michael Scott energy. If
  61. 1:50the model gets it wrong, create a skill
  62. 1:52so you stop having to send it back. But
  63. 1:54the question is, why do you have to keep
  64. 1:55sending things back when the model gets
  65. 1:57better and better. Matt has a specific
  66. 1:59answer for that and he calls it an asset
  67. 2:01rush.
  68. 2:02>> It has a little bit of what I'm going to
  69. 2:04call an asset rush. It's rushing towards
  70. 2:06creating an asset. It's like a premature
  71. 2:08completion.
  72. 2:09>> In plain English, this means the models
  73. 2:11are trained to finish, to produce
  74. 2:13something right now. The framework I
  75. 2:14like is think about models as an over
  76. 2:16eager student. They raise their hand,
  77. 2:18the answer's ready before it's even sure
  78. 2:20if it's correct. So, when doing this for
  79. 2:22yourself, be mindful to avoid rushing
  80. 2:25steps end to end. When your skills
  81. 2:26aren't focused on a single step of the
  82. 2:28process, that lands you in the danger
  83. 2:30zone for this asset rush problem. More
  84. 2:33bluntly speaking, if you don't do this,
  85. 2:35you'll generate AI slop. So, the mental
  86. 2:37shift here is this. Throughout your day,
  87. 2:38stop asking yourself "What did I use AI
  88. 2:41for?" and start asking "What did I have
  89. 2:43to send back?" And then once you do
  90. 2:45that, you then say "What specific step
  91. 2:47of the process was that?" So, here's a
  92. 2:49prompt you can run right now to scan
  93. 2:51your conversation history for the past
  94. 2:5230 days and flag places you send
  95. 2:55something back to see if those
  96. 2:56candidates can become a skill that's
  97. 2:58reusable. So, at this point, you know
  98. 3:00what requires a skill, but how do you
  99. 3:01actually create one? Step two is about
  100. 3:03using leading words where you borrow a
  101. 3:05word the model already knows. When you
  102. 3:07create a skill, you want to avoid
  103. 3:09starting from scratch or re-explaining
  104. 3:11something Claude already understands.
  105. 3:12LLMs have so much training data that
  106. 3:14there are specific words called leading
  107. 3:16words that immediately direct Claude to
  108. 3:18pull from its massive training data. So,
  109. 3:20if you choose the right word, the one
  110. 3:22the model already knows deeply, you
  111. 3:24radically simplify the whole entire
  112. 3:26skill. Here's Matt again explaining it
  113. 3:28and if it sounds technical, don't worry,
  114. 3:30I'll simplify it right after.
  115. 3:31>> Refactoring is such an old book, such a
  116. 3:33well cited book that these uh kind of
  117. 3:36smells are deep in the agent's priors.
  118. 3:38>> So, what he's actually saying is simple.
  119. 3:40Just by saying the word refactor, to
  120. 3:42pull from its billions of lines of
  121. 3:44training data specifically about
  122. 3:46refactoring code. And the alternative in
  123. 3:48this case, let's say you're trying to
  124. 3:49explain to Claude what refactoring is,
  125. 3:52you're starting from scratch. And the
  126. 3:53reality is this is a lot of wasted
  127. 3:55effort, the output won't as good, and
  128. 3:56this is what a lot of people actually
  129. 3:58do. The model already read all the
  130. 4:00information. You don't need to re-teach
  131. 4:02it. You just need to use a specific
  132. 4:04leading word to point it in the right
  133. 4:06direction. Here's how Matt put it in his
  134. 4:08own words. A made-up word recruits no
  135. 4:09priors. You pay in definition tokens
  136. 4:12what a pre-trained word gives free.
  137. 4:14Logically this should start making
  138. 4:15sense, but what are some other leading
  139. 4:16words? Here are some of the words that
  140. 4:18Matt uses in his Grill Me skill. He uses
  141. 4:20relentlessly. This shows up in both the
  142. 4:22description and the very first line.
  143. 4:24This single word is doing most of the
  144. 4:25work in the entire file. He used
  145. 4:27decision tree, which is decision
  146. 4:28branching into other decisions hanging
  147. 4:31off them. Work the tree in rounds, not
  148. 4:33one question at a time. And then
  149. 4:34frontier. Every decision whose
  150. 4:36prerequisites are already settled. Okay,
  151. 4:38it should start making sense, but how do
  152. 4:40you apply this to your day-to-day?
  153. 4:41Identify what you're actually trying to
  154. 4:43do and consider if there's already a
  155. 4:45leading word for it. There are a ton of
  156. 4:48established frameworks the model is
  157. 4:49already trained on. You just have to
  158. 4:51point it in the right direction. So
  159. 4:53before you write a skill from scratch,
  160. 4:54ask Claude directly, is there an
  161. 4:56established framework or term for what
  162. 4:58I'm describing? Then you'll just use
  163. 5:00whatever word it gives you instead of
  164. 5:02your own clunky description. For
  165. 5:04example, on screen are a bunch of
  166. 5:06leading words that apply to totally
  167. 5:08non-technical workflows, but here are
  168. 5:09some of my favorites. You have a smart
  169. 5:11goal framework which will help you set
  170. 5:12achievable goals. Steelman, which argues
  171. 5:15the opposing case. Premortem, imagines
  172. 5:17it already failed. Bluff, puts
  173. 5:19conclusions first and then spin
  174. 5:20discovery, ask pain point questions. And
  175. 5:23here's a prompt you can run to audit
  176. 5:25your own setup for missing leading
  177. 5:26words. So you've got the trigger and the
  178. 5:28word. Now before we get to the part that
  179. 5:30everyone messes up including me, which
  180. 5:32is step three, we've covered the power
  181. 5:34of skills in this video, but one
  182. 5:35limitation is video generation, which
  183. 5:38brings us today's video sponsor AI Cove,
  184. 5:40an AI video company out of Germany. If
  185. 5:42you've ever played with AI video, you
  186. 5:43know one of the biggest problems. You
  187. 5:44get a amazing clip and then you generate
  188. 5:47the next clip, and either the characters
  189. 5:48look entirely different, or the room
  190. 5:50changes, and suddenly the whole story
  191. 5:52kind of falls apart. And so, AI Core has
  192. 5:54fixed that with something they just
  193. 5:55launched called reference mode inside
  194. 5:57their Cinema AI Engine. And the
  195. 5:58character consistency that this
  196. 5:59establishes is a major unlock, because
  197. 6:02once your characters and locations stay
  198. 6:04the same from the first frame to the
  199. 6:06last, AI video stops being this cool
  200. 6:08clip generator, and it instead becomes
  201. 6:10something more than that. You can create
  202. 6:11full-length documentaries or music
  203. 6:13videos from a single idea in the future.
  204. 6:15I think that everyone's actually going
  205. 6:16to have their own personal documentary,
  206. 6:18and this is something that would have
  207. 6:19just never existed before. Now, that's
  208. 6:21just one impact this feature has, but
  209. 6:23two other features have surprised me.
  210. 6:24The first is that you can stay in
  211. 6:26control of the creative direction. AI
  212. 6:27Core will suggest the references in the
  213. 6:29video, but you can add your own. So, you
  214. 6:31could add a specific prop you want to
  215. 6:32keep coming up, or a visual style for
  216. 6:34the entire video. The camera, the color,
  217. 6:36the whole look, you can set this
  218. 6:38yourself. And now, the second it takes
  219. 6:39you from idea to finished video. So,
  220. 6:41with one click, you can generate a
  221. 6:42script, a voice, and the visuals. And
  222. 6:44now, for me, this isn't about replacing
  223. 6:46creativity, it's about expanding it. It
  224. 6:48lets you tell stories, make fun ads, and
  225. 6:50try concepts that you never could have
  226. 6:52made on your own before. And this is
  227. 6:54flexing a creative muscle that way too
  228. 6:56many people neglect. So, to go use
  229. 6:57reference mode, check out AI Core in the
  230. 6:59first link in the description, and use
  231. 7:01code Austin to get started for free, and
  232. 7:03get an extra 10 credits. Now, getting
  233. 7:05back to step three, less is more. Make
  234. 7:07skills smaller and harder to use. This
  235. 7:10is the part of the research that
  236. 7:11surprised me the most, and there are two
  237. 7:12things going on here. The first part is
  238. 7:14a skill length. On screen, you can see
  239. 7:16Matt Grillme's skill. It's seven lines,
  240. 7:1822 words. That's it. This single skill
  241. 7:21has 1.1 million installs. So, clearly,
  242. 7:24skills don't need to be long to be
  243. 7:25effective. Here's Matt talking about
  244. 7:27that directly.
  245. 7:28>> From this, skills don't have to be long
  246. 7:29to be impactful. You've just got to
  247. 7:31choose the right words for the LLM at
  248. 7:32the right time.
  249. 7:34>> This quote builds directly on what we
  250. 7:35just covered with leading words, but now
  251. 7:37we're looking at the skill as a whole,
  252. 7:38not just one line or one word inside of
  253. 7:41the skill. The second part is making
  254. 7:42them harder to use. For every skill, you
  255. 7:45can actually enable or disable Claude's
  256. 7:47ability to call it automatically. This
  257. 7:48is called model invocation, and all it
  258. 7:50means is can the model use the skill
  259. 7:52without you telling it to. And on the
  260. 7:54surface, it feels like you want the
  261. 7:56model to be able to use the skills,
  262. 7:57right? That's why you're creating it.
  263. 7:58But when you actually look at Matt's
  264. 8:00skills, he has disabled the ability for
  265. 8:02models to call them, essentially saying
  266. 8:04a human has to call this skill. And when
  267. 8:07I dug into why, I found two tweets that
  268. 8:08explain it perfectly. First is the
  269. 8:10reasoning itself, cognitive load versus
  270. 8:12context load. Cognitive load lessens
  271. 8:14over time as you get better, context
  272. 8:16load costs you tokens forever. If you
  273. 8:18let the model auto-invoke skills, each
  274. 8:20of those skill descriptions gets loaded
  275. 8:22into context on every message you sent.
  276. 8:25That costs you tokens on every single
  277. 8:26run. If you remove that ability, you
  278. 8:28remove the constant contextual load, and
  279. 8:30it forces you to think about when to use
  280. 8:32a skill. And over time, this makes you
  281. 8:34better at using AI, and in turn fights
  282. 8:37intellectual obesity, a concept you guys
  283. 8:39know is so important to me. Then in that
  284. 8:41tweet, he lays out the actual benefit of
  285. 8:43doing it this way. One benefit of being
  286. 8:45a primarily user-invoked skill set is
  287. 8:48that we get rid of a whole category of
  288. 8:49errors you never need to worry about.
  289. 8:51Did {slash} two spec fire? Why did
  290. 8:54{slash} two spec fire there? Why did it
  291. 8:56choose two tickets instead of two spec?
  292. 8:58You stay in control, plus it means most
  293. 9:01of the skills don't cost any tokens
  294. 9:03until you invoke them. So, Matt is
  295. 9:04intentionally making these skills harder
  296. 9:06to use on purpose, so the model
  297. 9:08literally can't call them without him.
  298. 9:10And from my own experience, I've
  299. 9:12definitely seen the downside of not
  300. 9:14doing this. Claude will just
  301. 9:16automatically fire a skill that I didn't
  302. 9:17want it to, and it'll in turn slow down
  303. 9:19the whole process or trigger something
  304. 9:21that's totally unnecessary for what I'm
  305. 9:23actually trying to do. So, here's how
  306. 9:25you can apply this. Here is a prompt
  307. 9:27that simplifies your existing skills and
  308. 9:29flags which ones are good candidates to
  309. 9:31make user-invoked only. And now, before
  310. 9:33we get to step four, this is our
  311. 9:34anti-slop agreement. Everything in this
  312. 9:36video is built by human for humans. So,
  313. 9:39all that I ask is that you subscribe to
  314. 9:41help this content reach more people. In
  315. 9:42every video, I give away a Claude Max
  316. 9:44subscription as a thank you for everyone
  317. 9:45who supports. So, shout out this video's
  318. 9:47winner, journeymengeorgie7000,
  319. 9:50for building a project management
  320. 9:52software. And to enter this video's
  321. 9:53giveaway, comment with what you're
  322. 9:54building, and let's chat about how
  323. 9:56skills can help that. Step four, iterate
  324. 9:58over time and don't overthink if it's
  325. 10:01working. So, you've written your skill,
  326. 10:02it's short, it's got the right leading
  327. 10:04word, but is it working? Did your
  328. 10:06changes help the final result? How do
  329. 10:08you even measure that? This is where
  330. 10:09most people mid-curve the whole skills
  331. 10:11thing. They get worked up trying to
  332. 10:13quantify whether a skill has improved or
  333. 10:15not, and that's exactly why this step in
  334. 10:16Matt's process is so refreshing. Matt
  335. 10:18doesn't quantifiably measure whether a
  336. 10:20change made a skill better. He mostly
  337. 10:23goes by vibes. Here's him talking about
  338. 10:25exactly this. Actually, at least, you
  339. 10:27know, in terms of vibes. I'm not doing
  340. 10:28evals here. Maybe I should be doing
  341. 10:30evals on this particular case, but what
  342. 10:32I'm seeing is that people aren't
  343. 10:33reporting this anymore. Guy with 22
  344. 10:35million installs has no testing setup.
  345. 10:37He uses the skill himself for about a
  346. 10:39week and just watches whether people
  347. 10:41stop complaining. Now, sure, you won't
  348. 10:43have as many testers as Matt does, but
  349. 10:45the stakes are just as high for you in
  350. 10:47your own world. And he still just goes
  351. 10:49by vibes. So, to do this testing
  352. 10:51yourself, you don't need to go crazy
  353. 10:53trying to evaluate whether a skill is
  354. 10:55improving. You just need to use it and
  355. 10:57run what I'll call the send it back
  356. 10:59test. If you don't have to send back the
  357. 11:00result, the skill is working. It's
  358. 11:02really that simple. And when you're
  359. 11:04making skills, don't try and be perfect
  360. 11:06on version one. Perfection is the enemy
  361. 11:08of good. Good is good enough. Just the
  362. 11:10skill you write, just iterate from the
  363. 11:12first version. If you break down how
  364. 11:13some of Matt's most famous skills got
  365. 11:15built, every single line tells a story.
  366. 11:17On screen, here is a history of the
  367. 11:19change log for his skill. "Give your
  368. 11:21recommended answer" was added because
  369. 11:23conversations were dragging. "Ask one
  370. 11:25question at a time" was added because
  371. 11:26firing off multiple questions was
  372. 11:28bewildering. What a great word,
  373. 11:30bewildering. "Do not enact the plan
  374. 11:32until I confirm was added because the
  375. 11:34agent kept skipping ahead and just
  376. 11:36building and split facts from decisions
  377. 11:38was added because the agent started
  378. 11:40grilling itself. None of that was
  379. 11:41planned in advance. Each line is
  380. 11:43something that broke got noticed and was
  381. 11:45fixed with a one sentence addition. So
  382. 11:48wrapping up the step there are three
  383. 11:49things to actually do. The first is run
  384. 11:51your skill on real work not test cases.
  385. 11:54The second is at the end of the week
  386. 11:55have Claude analyze your conversation
  387. 11:57history for anything you sent something
  388. 11:58back and turn recurring corrections into
  389. 12:01a one sentence fix. And then three once
  390. 12:03you have that fixed stop and try again.
  391. 12:05Don't add anything you haven't
  392. 12:06personally watched fail. So at this
  393. 12:08point you'll keep iterating and
  394. 12:09improving your skills which creates a
  395. 12:11new problem that nobody warned you
  396. 12:13about. Step five you need to cut it.
  397. 12:15Everyone treats a skill like a thing
  398. 12:17that only ever grows. Nobody schedules
  399. 12:19the day to make one smaller and honestly
  400. 12:21this is a step that I forget all of the
  401. 12:23time. Matt's most engaged post about
  402. 12:25skills isn't even about writing. It's
  403. 12:27about deleting. Check for lines like be
  404. 12:29thorough. They're no-ops. Try removing
  405. 12:31that. Does the output change? No, then
  406. 12:33the line was a no-op. Agent authored
  407. 12:36skills are littered with no-ops. In
  408. 12:38plain English go through your skill one
  409. 12:40sentence at a time and ask yourself
  410. 12:42would Claude have done that anyway? If
  411. 12:44the answer is yes, the line is
  412. 12:45decoration and Matt calls these no-ops
  413. 12:47and they're costing you tokens to say
  414. 12:49absolutely nothing. And so the decision
  415. 12:50you have here is you can be precise or
  416. 12:52aggressive. The precise version go line
  417. 12:54by line, remove the line you suspect is
  418. 12:56a no-op, run it again and if the output
  419. 12:59didn't change it was never doing
  420. 13:00anything in the first place. The
  421. 13:02aggressive version which I personally
  422. 13:04prefer is just have Claude scan the
  423. 13:06whole thing for no-op language and
  424. 13:08delete those words. Then just see what
  425. 13:10happens. And so either way there are a
  426. 13:12couple of rules when you're cleaning up
  427. 13:13these skills. The first is delete the
  428. 13:15whole sentence. Don't just trim words
  429. 13:17out of it. Then say what you want not
  430. 13:20what you don't want. Words like don't do
  431. 13:22this drags the banned thing right back
  432. 13:24into context. Matt has a line on this
  433. 13:26where he says, "If you say don't think
  434. 13:28of an elephant, and then suddenly the
  435. 13:30elephant is all that's there." Just by
  436. 13:32saying the word elephant, you've already
  437. 13:34put it in the model's head. So, how does
  438. 13:36Matt apply this himself? He's constantly
  439. 13:38shortening and deleting skills, and he's
  440. 13:40just as careful about creating new ones.
  441. 13:43His rule is, "If you can compose the
  442. 13:45behavior from existing skills, then
  443. 13:47don't create a new skill." You can apply
  444. 13:49this thought process by first doing a
  445. 13:51full audit of your skills for no-op
  446. 13:53lines, then deleting whole sentences.
  447. 13:55And then two, by checking if your skills
  448. 13:56overlap, then condensing them or cutting
  449. 13:59them entirely. On screen, you can see a
  450. 14:01prompt that will help you do both of
  451. 14:02these. So, let's recap all five steps.
  452. 14:04The first is notice what you keep
  453. 14:06sending back to identify where you need
  454. 14:08skills. Two is use leading words the
  455. 14:11model already knows to simplify your
  456. 14:13skills. Three, keep skills small and
  457. 14:15make them harder to auto-invoke to avoid
  458. 14:18bloat. Four is iterate by using the
  459. 14:19simple send back test, and five, cut the
  460. 14:22no-ops before you ever add something
  461. 14:24new. There's one thing that Matt said
  462. 14:25that reframed the entire video for me.
  463. 14:27>> AI has eaten tactical programming. It's
  464. 14:30gone. There's a sort of crisis of
  465. 14:32responsibility when it comes to working
  466. 14:34with AI, which is that people fire off
  467. 14:36these things, they merge into main, and
  468. 14:38then no one is responsible for it. You
  469. 14:40know, it's the AI who did it, wasn't me,
  470. 14:42it was not my code.
  471. 14:43>> So, he's saying that AI ate the doing,
  472. 14:45but it can't remove ownership. The
  473. 14:47reality is it comes down to people, and
  474. 14:49you have to own your process, and a
  475. 14:50skill is the way to do that. And as a
  476. 14:52result, you do not get anything out of
  477. 14:54blindly copying other people's setups,
  478. 14:56because you don't have a deep
  479. 14:57understanding of how it works. And Matt
  480. 14:59agrees with this. He said this while
  481. 15:00replying to someone who called his
  482. 15:01skills awful. Most folks should write
  483. 15:03their own skills and own their own
  484. 15:05processes. So, don't blindly copy
  485. 15:07things. Instead, create your own using
  486. 15:09the five-step framework we've covered in
  487. 15:11this video. Now, once you do that, you
  488. 15:12may have a problem where you say, "I
  489. 15:14send back a ton of things. Where do I
  490. 15:15start?" So, to solve that, check out
  491. 15:17this video, where I break down exactly
  492. 15:19how Entropic's own team uses skills.
  493. 15:21That'll inspire you on exactly where to
  494. 15:22start. And if you like this video, I
  495. 15:24guarantee you'll love that. I'll see you
  496. 15:26over there. Peace.

About this transcript

This page contains the full transcript of Matt Pocock's EXACT System To 10x Your Claude Skills by Austin Marchese, generated from the public captions YouTube serves with the video. The transcript has 3,394 words across 496 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.