Matt Pocock's EXACT System To 10x Your Claude Skills — Transcript
Full transcript
- 0:00Mattacott has the most installed Claude
- 0:02skills on the planet with over 22
- 0:04million people using that. So I decided
- 0:06to do a deep dive in how exactly he
- 0:08creates his skills and the results
- 0:10shocked me. I went in expecting
- 0:11extremely complicated processes, but
- 0:13after analyzing every video he's ever
- 0:15posted and every skill he's created, I
- 0:17realized that he has a five-step process
- 0:19he uses to create every skill. But
- 0:21unfortunately, Matt's videos are
- 0:23extremely technical and the majority of
- 0:24his content is for engineers. So in this
- 0:26video, I'll break down each of the five
- 0:28steps he uses to create skills, simplify
- 0:31it and then tell you exactly how you can
- 0:32apply this process in your day-to-day to
- 0:35make your skills better and build 10
- 0:36times faster. And at the end of the
- 0:38video, I'll cover the reason why Matt
- 0:40thinks you should never blindly copy
- 0:42anyone else's skills. The step one is
- 0:44about notice and this is the thing that
- 0:46you keep sending back. Now, I've heard a
- 0:48lot of breakdowns on what a Claude skill
- 0:50actually is, right? It's a reusable
- 0:51prompt. It's a process of teaching
- 0:53Claude something new and these aren't
- 0:55wrong, but Matt has one of my favorite
- 0:57definitions that I've come across. A
- 0:58skill is a correction you got tired of
- 1:01making. Here's someone on podcast
- 1:02explaining exactly this.
- 1:03>> Uh I went to a restaurant and I ordered
- 1:07a chorizo salad and I got a chorizo
- 1:09sandwich. Right? So I had to send it
- 1:12back. And what I was finding is I was
- 1:14just having to send things back all the
- 1:16time to AI.
- 1:16>> Said another way, the key reason to
- 1:18create a skill isn't what you do most
- 1:19often with Claude. It's what you keep
- 1:21having to correct. If the model can
- 1:23already complete a task correctly the
- 1:24first time, a skill for that is just
- 1:26unnecessary bloat. And that's actually
- 1:28part of the reason why the creator of
- 1:29Claude code told people to go delete all
- 1:31of their Claude skills. If Claude
- 1:32already gets it right, the skill isn't
- 1:34doing anything except costing tokens.
- 1:36With Matt's framework, he's essentially
- 1:38saying if Claude doesn't get it right
- 1:39the first time, send it back. And this
- 1:41kind of remind me of that scene in The
- 1:43Office where he says, "I'll have the
- 1:44spaghetti with a side salad. If the
- 1:46salad is on top, I'll send it back."
- 1:48It's that same Michael Scott energy. If
- 1:50the model gets it wrong, create a skill
- 1:52so you stop having to send it back. But
- 1:54the question is, why do you have to keep
- 1:55sending things back when the model gets
- 1:57better and better. Matt has a specific
- 1:59answer for that and he calls it an asset
- 2:01rush.
- 2:02>> It has a little bit of what I'm going to
- 2:04call an asset rush. It's rushing towards
- 2:06creating an asset. It's like a premature
- 2:08completion.
- 2:09>> In plain English, this means the models
- 2:11are trained to finish, to produce
- 2:13something right now. The framework I
- 2:14like is think about models as an over
- 2:16eager student. They raise their hand,
- 2:18the answer's ready before it's even sure
- 2:20if it's correct. So, when doing this for
- 2:22yourself, be mindful to avoid rushing
- 2:25steps end to end. When your skills
- 2:26aren't focused on a single step of the
- 2:28process, that lands you in the danger
- 2:30zone for this asset rush problem. More
- 2:33bluntly speaking, if you don't do this,
- 2:35you'll generate AI slop. So, the mental
- 2:37shift here is this. Throughout your day,
- 2:38stop asking yourself "What did I use AI
- 2:41for?" and start asking "What did I have
- 2:43to send back?" And then once you do
- 2:45that, you then say "What specific step
- 2:47of the process was that?" So, here's a
- 2:49prompt you can run right now to scan
- 2:51your conversation history for the past
- 2:5230 days and flag places you send
- 2:55something back to see if those
- 2:56candidates can become a skill that's
- 2:58reusable. So, at this point, you know
- 3:00what requires a skill, but how do you
- 3:01actually create one? Step two is about
- 3:03using leading words where you borrow a
- 3:05word the model already knows. When you
- 3:07create a skill, you want to avoid
- 3:09starting from scratch or re-explaining
- 3:11something Claude already understands.
- 3:12LLMs have so much training data that
- 3:14there are specific words called leading
- 3:16words that immediately direct Claude to
- 3:18pull from its massive training data. So,
- 3:20if you choose the right word, the one
- 3:22the model already knows deeply, you
- 3:24radically simplify the whole entire
- 3:26skill. Here's Matt again explaining it
- 3:28and if it sounds technical, don't worry,
- 3:30I'll simplify it right after.
- 3:31>> Refactoring is such an old book, such a
- 3:33well cited book that these uh kind of
- 3:36smells are deep in the agent's priors.
- 3:38>> So, what he's actually saying is simple.
- 3:40Just by saying the word refactor, to
- 3:42pull from its billions of lines of
- 3:44training data specifically about
- 3:46refactoring code. And the alternative in
- 3:48this case, let's say you're trying to
- 3:49explain to Claude what refactoring is,
- 3:52you're starting from scratch. And the
- 3:53reality is this is a lot of wasted
- 3:55effort, the output won't as good, and
- 3:56this is what a lot of people actually
- 3:58do. The model already read all the
- 4:00information. You don't need to re-teach
- 4:02it. You just need to use a specific
- 4:04leading word to point it in the right
- 4:06direction. Here's how Matt put it in his
- 4:08own words. A made-up word recruits no
- 4:09priors. You pay in definition tokens
- 4:12what a pre-trained word gives free.
- 4:14Logically this should start making
- 4:15sense, but what are some other leading
- 4:16words? Here are some of the words that
- 4:18Matt uses in his Grill Me skill. He uses
- 4:20relentlessly. This shows up in both the
- 4:22description and the very first line.
- 4:24This single word is doing most of the
- 4:25work in the entire file. He used
- 4:27decision tree, which is decision
- 4:28branching into other decisions hanging
- 4:31off them. Work the tree in rounds, not
- 4:33one question at a time. And then
- 4:34frontier. Every decision whose
- 4:36prerequisites are already settled. Okay,
- 4:38it should start making sense, but how do
- 4:40you apply this to your day-to-day?
- 4:41Identify what you're actually trying to
- 4:43do and consider if there's already a
- 4:45leading word for it. There are a ton of
- 4:48established frameworks the model is
- 4:49already trained on. You just have to
- 4:51point it in the right direction. So
- 4:53before you write a skill from scratch,
- 4:54ask Claude directly, is there an
- 4:56established framework or term for what
- 4:58I'm describing? Then you'll just use
- 5:00whatever word it gives you instead of
- 5:02your own clunky description. For
- 5:04example, on screen are a bunch of
- 5:06leading words that apply to totally
- 5:08non-technical workflows, but here are
- 5:09some of my favorites. You have a smart
- 5:11goal framework which will help you set
- 5:12achievable goals. Steelman, which argues
- 5:15the opposing case. Premortem, imagines
- 5:17it already failed. Bluff, puts
- 5:19conclusions first and then spin
- 5:20discovery, ask pain point questions. And
- 5:23here's a prompt you can run to audit
- 5:25your own setup for missing leading
- 5:26words. So you've got the trigger and the
- 5:28word. Now before we get to the part that
- 5:30everyone messes up including me, which
- 5:32is step three, we've covered the power
- 5:34of skills in this video, but one
- 5:35limitation is video generation, which
- 5:38brings us today's video sponsor AI Cove,
- 5:40an AI video company out of Germany. If
- 5:42you've ever played with AI video, you
- 5:43know one of the biggest problems. You
- 5:44get a amazing clip and then you generate
- 5:47the next clip, and either the characters
- 5:48look entirely different, or the room
- 5:50changes, and suddenly the whole story
- 5:52kind of falls apart. And so, AI Core has
- 5:54fixed that with something they just
- 5:55launched called reference mode inside
- 5:57their Cinema AI Engine. And the
- 5:58character consistency that this
- 5:59establishes is a major unlock, because
- 6:02once your characters and locations stay
- 6:04the same from the first frame to the
- 6:06last, AI video stops being this cool
- 6:08clip generator, and it instead becomes
- 6:10something more than that. You can create
- 6:11full-length documentaries or music
- 6:13videos from a single idea in the future.
- 6:15I think that everyone's actually going
- 6:16to have their own personal documentary,
- 6:18and this is something that would have
- 6:19just never existed before. Now, that's
- 6:21just one impact this feature has, but
- 6:23two other features have surprised me.
- 6:24The first is that you can stay in
- 6:26control of the creative direction. AI
- 6:27Core will suggest the references in the
- 6:29video, but you can add your own. So, you
- 6:31could add a specific prop you want to
- 6:32keep coming up, or a visual style for
- 6:34the entire video. The camera, the color,
- 6:36the whole look, you can set this
- 6:38yourself. And now, the second it takes
- 6:39you from idea to finished video. So,
- 6:41with one click, you can generate a
- 6:42script, a voice, and the visuals. And
- 6:44now, for me, this isn't about replacing
- 6:46creativity, it's about expanding it. It
- 6:48lets you tell stories, make fun ads, and
- 6:50try concepts that you never could have
- 6:52made on your own before. And this is
- 6:54flexing a creative muscle that way too
- 6:56many people neglect. So, to go use
- 6:57reference mode, check out AI Core in the
- 6:59first link in the description, and use
- 7:01code Austin to get started for free, and
- 7:03get an extra 10 credits. Now, getting
- 7:05back to step three, less is more. Make
- 7:07skills smaller and harder to use. This
- 7:10is the part of the research that
- 7:11surprised me the most, and there are two
- 7:12things going on here. The first part is
- 7:14a skill length. On screen, you can see
- 7:16Matt Grillme's skill. It's seven lines,
- 7:1822 words. That's it. This single skill
- 7:21has 1.1 million installs. So, clearly,
- 7:24skills don't need to be long to be
- 7:25effective. Here's Matt talking about
- 7:27that directly.
- 7:28>> From this, skills don't have to be long
- 7:29to be impactful. You've just got to
- 7:31choose the right words for the LLM at
- 7:32the right time.
- 7:34>> This quote builds directly on what we
- 7:35just covered with leading words, but now
- 7:37we're looking at the skill as a whole,
- 7:38not just one line or one word inside of
- 7:41the skill. The second part is making
- 7:42them harder to use. For every skill, you
- 7:45can actually enable or disable Claude's
- 7:47ability to call it automatically. This
- 7:48is called model invocation, and all it
- 7:50means is can the model use the skill
- 7:52without you telling it to. And on the
- 7:54surface, it feels like you want the
- 7:56model to be able to use the skills,
- 7:57right? That's why you're creating it.
- 7:58But when you actually look at Matt's
- 8:00skills, he has disabled the ability for
- 8:02models to call them, essentially saying
- 8:04a human has to call this skill. And when
- 8:07I dug into why, I found two tweets that
- 8:08explain it perfectly. First is the
- 8:10reasoning itself, cognitive load versus
- 8:12context load. Cognitive load lessens
- 8:14over time as you get better, context
- 8:16load costs you tokens forever. If you
- 8:18let the model auto-invoke skills, each
- 8:20of those skill descriptions gets loaded
- 8:22into context on every message you sent.
- 8:25That costs you tokens on every single
- 8:26run. If you remove that ability, you
- 8:28remove the constant contextual load, and
- 8:30it forces you to think about when to use
- 8:32a skill. And over time, this makes you
- 8:34better at using AI, and in turn fights
- 8:37intellectual obesity, a concept you guys
- 8:39know is so important to me. Then in that
- 8:41tweet, he lays out the actual benefit of
- 8:43doing it this way. One benefit of being
- 8:45a primarily user-invoked skill set is
- 8:48that we get rid of a whole category of
- 8:49errors you never need to worry about.
- 8:51Did {slash} two spec fire? Why did
- 8:54{slash} two spec fire there? Why did it
- 8:56choose two tickets instead of two spec?
- 8:58You stay in control, plus it means most
- 9:01of the skills don't cost any tokens
- 9:03until you invoke them. So, Matt is
- 9:04intentionally making these skills harder
- 9:06to use on purpose, so the model
- 9:08literally can't call them without him.
- 9:10And from my own experience, I've
- 9:12definitely seen the downside of not
- 9:14doing this. Claude will just
- 9:16automatically fire a skill that I didn't
- 9:17want it to, and it'll in turn slow down
- 9:19the whole process or trigger something
- 9:21that's totally unnecessary for what I'm
- 9:23actually trying to do. So, here's how
- 9:25you can apply this. Here is a prompt
- 9:27that simplifies your existing skills and
- 9:29flags which ones are good candidates to
- 9:31make user-invoked only. And now, before
- 9:33we get to step four, this is our
- 9:34anti-slop agreement. Everything in this
- 9:36video is built by human for humans. So,
- 9:39all that I ask is that you subscribe to
- 9:41help this content reach more people. In
- 9:42every video, I give away a Claude Max
- 9:44subscription as a thank you for everyone
- 9:45who supports. So, shout out this video's
- 9:47winner, journeymengeorgie7000,
- 9:50for building a project management
- 9:52software. And to enter this video's
- 9:53giveaway, comment with what you're
- 9:54building, and let's chat about how
- 9:56skills can help that. Step four, iterate
- 9:58over time and don't overthink if it's
- 10:01working. So, you've written your skill,
- 10:02it's short, it's got the right leading
- 10:04word, but is it working? Did your
- 10:06changes help the final result? How do
- 10:08you even measure that? This is where
- 10:09most people mid-curve the whole skills
- 10:11thing. They get worked up trying to
- 10:13quantify whether a skill has improved or
- 10:15not, and that's exactly why this step in
- 10:16Matt's process is so refreshing. Matt
- 10:18doesn't quantifiably measure whether a
- 10:20change made a skill better. He mostly
- 10:23goes by vibes. Here's him talking about
- 10:25exactly this. Actually, at least, you
- 10:27know, in terms of vibes. I'm not doing
- 10:28evals here. Maybe I should be doing
- 10:30evals on this particular case, but what
- 10:32I'm seeing is that people aren't
- 10:33reporting this anymore. Guy with 22
- 10:35million installs has no testing setup.
- 10:37He uses the skill himself for about a
- 10:39week and just watches whether people
- 10:41stop complaining. Now, sure, you won't
- 10:43have as many testers as Matt does, but
- 10:45the stakes are just as high for you in
- 10:47your own world. And he still just goes
- 10:49by vibes. So, to do this testing
- 10:51yourself, you don't need to go crazy
- 10:53trying to evaluate whether a skill is
- 10:55improving. You just need to use it and
- 10:57run what I'll call the send it back
- 10:59test. If you don't have to send back the
- 11:00result, the skill is working. It's
- 11:02really that simple. And when you're
- 11:04making skills, don't try and be perfect
- 11:06on version one. Perfection is the enemy
- 11:08of good. Good is good enough. Just the
- 11:10skill you write, just iterate from the
- 11:12first version. If you break down how
- 11:13some of Matt's most famous skills got
- 11:15built, every single line tells a story.
- 11:17On screen, here is a history of the
- 11:19change log for his skill. "Give your
- 11:21recommended answer" was added because
- 11:23conversations were dragging. "Ask one
- 11:25question at a time" was added because
- 11:26firing off multiple questions was
- 11:28bewildering. What a great word,
- 11:30bewildering. "Do not enact the plan
- 11:32until I confirm was added because the
- 11:34agent kept skipping ahead and just
- 11:36building and split facts from decisions
- 11:38was added because the agent started
- 11:40grilling itself. None of that was
- 11:41planned in advance. Each line is
- 11:43something that broke got noticed and was
- 11:45fixed with a one sentence addition. So
- 11:48wrapping up the step there are three
- 11:49things to actually do. The first is run
- 11:51your skill on real work not test cases.
- 11:54The second is at the end of the week
- 11:55have Claude analyze your conversation
- 11:57history for anything you sent something
- 11:58back and turn recurring corrections into
- 12:01a one sentence fix. And then three once
- 12:03you have that fixed stop and try again.
- 12:05Don't add anything you haven't
- 12:06personally watched fail. So at this
- 12:08point you'll keep iterating and
- 12:09improving your skills which creates a
- 12:11new problem that nobody warned you
- 12:13about. Step five you need to cut it.
- 12:15Everyone treats a skill like a thing
- 12:17that only ever grows. Nobody schedules
- 12:19the day to make one smaller and honestly
- 12:21this is a step that I forget all of the
- 12:23time. Matt's most engaged post about
- 12:25skills isn't even about writing. It's
- 12:27about deleting. Check for lines like be
- 12:29thorough. They're no-ops. Try removing
- 12:31that. Does the output change? No, then
- 12:33the line was a no-op. Agent authored
- 12:36skills are littered with no-ops. In
- 12:38plain English go through your skill one
- 12:40sentence at a time and ask yourself
- 12:42would Claude have done that anyway? If
- 12:44the answer is yes, the line is
- 12:45decoration and Matt calls these no-ops
- 12:47and they're costing you tokens to say
- 12:49absolutely nothing. And so the decision
- 12:50you have here is you can be precise or
- 12:52aggressive. The precise version go line
- 12:54by line, remove the line you suspect is
- 12:56a no-op, run it again and if the output
- 12:59didn't change it was never doing
- 13:00anything in the first place. The
- 13:02aggressive version which I personally
- 13:04prefer is just have Claude scan the
- 13:06whole thing for no-op language and
- 13:08delete those words. Then just see what
- 13:10happens. And so either way there are a
- 13:12couple of rules when you're cleaning up
- 13:13these skills. The first is delete the
- 13:15whole sentence. Don't just trim words
- 13:17out of it. Then say what you want not
- 13:20what you don't want. Words like don't do
- 13:22this drags the banned thing right back
- 13:24into context. Matt has a line on this
- 13:26where he says, "If you say don't think
- 13:28of an elephant, and then suddenly the
- 13:30elephant is all that's there." Just by
- 13:32saying the word elephant, you've already
- 13:34put it in the model's head. So, how does
- 13:36Matt apply this himself? He's constantly
- 13:38shortening and deleting skills, and he's
- 13:40just as careful about creating new ones.
- 13:43His rule is, "If you can compose the
- 13:45behavior from existing skills, then
- 13:47don't create a new skill." You can apply
- 13:49this thought process by first doing a
- 13:51full audit of your skills for no-op
- 13:53lines, then deleting whole sentences.
- 13:55And then two, by checking if your skills
- 13:56overlap, then condensing them or cutting
- 13:59them entirely. On screen, you can see a
- 14:01prompt that will help you do both of
- 14:02these. So, let's recap all five steps.
- 14:04The first is notice what you keep
- 14:06sending back to identify where you need
- 14:08skills. Two is use leading words the
- 14:11model already knows to simplify your
- 14:13skills. Three, keep skills small and
- 14:15make them harder to auto-invoke to avoid
- 14:18bloat. Four is iterate by using the
- 14:19simple send back test, and five, cut the
- 14:22no-ops before you ever add something
- 14:24new. There's one thing that Matt said
- 14:25that reframed the entire video for me.
- 14:27>> AI has eaten tactical programming. It's
- 14:30gone. There's a sort of crisis of
- 14:32responsibility when it comes to working
- 14:34with AI, which is that people fire off
- 14:36these things, they merge into main, and
- 14:38then no one is responsible for it. You
- 14:40know, it's the AI who did it, wasn't me,
- 14:42it was not my code.
- 14:43>> So, he's saying that AI ate the doing,
- 14:45but it can't remove ownership. The
- 14:47reality is it comes down to people, and
- 14:49you have to own your process, and a
- 14:50skill is the way to do that. And as a
- 14:52result, you do not get anything out of
- 14:54blindly copying other people's setups,
- 14:56because you don't have a deep
- 14:57understanding of how it works. And Matt
- 14:59agrees with this. He said this while
- 15:00replying to someone who called his
- 15:01skills awful. Most folks should write
- 15:03their own skills and own their own
- 15:05processes. So, don't blindly copy
- 15:07things. Instead, create your own using
- 15:09the five-step framework we've covered in
- 15:11this video. Now, once you do that, you
- 15:12may have a problem where you say, "I
- 15:14send back a ton of things. Where do I
- 15:15start?" So, to solve that, check out
- 15:17this video, where I break down exactly
- 15:19how Entropic's own team uses skills.
- 15:21That'll inspire you on exactly where to
- 15:22start. And if you like this video, I
- 15:24guarantee you'll love that. I'll see you
- 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.