So What? Getting started with Claude Skills

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

You can watch on YouTube Live. Be sure to subscribe and follow so you never miss an episode!

In this episode of So What?, you’ll learn about Claude skills and what they can do for you. Create and implement your own custom skills within Claude to streamline your workflows. You will understand how Claude skills differ from other AI models and why portability matters for your business, and you’ll gain practical insights into maximizing efficiency and avoiding common pitfalls when using Claude skills.

Watch the video here:

So What? Getting started with Claude Skills

Watch this video on YouTube

Can’t see anything? Watch it on YouTube here.

In this episode you’ll learn:

The new Claude Skills function
How Claude Skills is different from a GPT or Gem
How to get started using Claude Skills

Transcript:

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Katie Robbert: 00:34 – Happy Thursday. Welcome to “So What? The Marketing Analytics and Insights Live Show.” I am Katie, joined by Chris and John. We got the band back together this week, guys.

Christopher Penn: 00:43 – We did. We did kind of a.

John Wall: 00:45 – Wayo, half high five, but we’ll take it.

Katie Robbert: 00:48 – So, we’re together this week. The next couple of weeks we’re going to be up, down, left, right, center. So, we are knee-deep in event season. We hope to catch you at some of the live events, but this week we are here, and we are bringing you the live stream: “So What: Getting Started with Claude Skills?” I’m actually really excited to learn about it because I know Chris, obviously the second something is released, you’re on top of it, you’re playing with it, you’re testing it. I was using Claude the other day, and Skills started kicking in that I didn’t ask for, and I was like, “Whoa, wait a second,” but I also couldn’t find where the Skills were to actually look at them, to see what they were comprised of. So, I have lots of questions, as expected.

Katie Robbert: 01:35 – Where should we start?

Christopher Penn: 01:37 – We should probably start with what the heck is this and why would anyone care? For folks who’ve been using ChatGPT for a while, you probably know that you can invoke a GPT within a chat, within a new chat by saying, “At writing style analysis,” or “At, virtual Katie, what do you think about this thing?” It can reference that, but those GPTs are somewhat limited, and they’re not available in other systems and things like that. So, if you are in, say, OpenAI’s Codex, their coding environment, your GPTs are not there. If you’re using the API, it’s very difficult to do that. But that was sort of the beginning of this concept: if you have a pre-packaged thing, like a mini app, you can call it within a conversation.

Christopher Penn: 02:24 – About a week ago, Anthropic, which is the maker of Claude, introduced their version of this, called Skills. Skills are essentially the ability to take a similar concept, like a writing style analyzer, a business requirements document analyzer—whatever the things that you would normally use as Claude projects—but turn them into Skills that you could invoke in a conversation at any time: new chat, old chat, etc. Two things make it very interesting. One, you can take the Skill, and you can hand it as a plug-in file to another Claude user. Any other Claude users can drop it into theirs and immediately have access to it. Two, it’s usable in environments like Claude Code, the coding tool. Katie, you look like you have some questions.

Katie Robbert: 03:13 – I do. It might just be me. John, I would love to get your perspective. I didn’t know, I wasn’t aware, that in something like an OpenAI, you could add a custom GPT in a different conversation. So that’s my learning number one today. I wasn’t aware that was a thing, which is actually pretty cool. If I’m understanding correctly, you’re in the “Just What Can I Help With?” You’re not in a custom GPT, right? So you could start a conversation and almost like you’re just tagging a person in Slack or any other social media platform, for example, you can tag a custom model that you’ve built inside that same system.

Christopher Penn: 04:04 – Sometimes it’s really unreliable.

Katie Robbert: 04:07 – Can you do the same thing in Gemini?

Christopher Penn: 04:09 – You cannot.

Katie Robbert: 04:10 – Okay.

Christopher Penn: 04:12 – I’m trying to actually get this work.

John Wall: 04:14 – No, in fact, I’m like ten steps behind you. I’m just like, Chris thought this was way exciting, and I need to know what this is, and I have no friggin’ idea. So.

Katie Robbert: 04:25 – Well, I mean, I can kind of see the usefulness of it. If you’re in the middle of trying to do something and you’re like, “Oh, I probably should have had this conversation in the custom model instead,” you can bring it into the custom model if you forget to actually have the conversation in the custom model. Sort of. It sounds like, because when I look at this, I’m like, “Well, why wouldn’t I just have the conversation in the FAQ generator versus tagging it into a separate conversation?”

Christopher Penn: 04:56 – Right. The idea being that you could reference that GPT as you were having a conversation midway through. You’re like, “Hey, let’s work with our ideal customer profile.” Maybe you’re working on it, and then you’re like, “Oh, I need to talk to my co-CEO about this.” So you could say, “At co-CEO, what do you think about whatever’s going on?” It could invoke and bring that GPT into that conversation. So this has been a capability. It’s been in Open Chat GPT for about a year now, give or take. It’s useful. To be clear, this is not a novelty. But as I said, there are scoping issues with it that Claude, in their implementation, has done a really good job of addressing. Our buddy.

Katie Robbert: 05:44 – Brian brings up a really good use case, which is I love bringing multiple custom GPT personas into one prompt and doing a virtual focus group. That is a fantastic use case because you likely have custom GPT personas for maybe different segments of your audience or different members of your team internally. Bringing them all into that one conversation is actually a very smart way to use them. So thank you for that, Brian.

Christopher Penn: 06:14 – Exactly. So, Claude, in their implementation, it’s called Skills. You will need to go into your settings, into your capabilities menu in Claude—and this can be on the web or in desktop—and scroll down until you get to the Skills section. If you have not enabled it, it will ask you if you want to enable it. It will then give you some feedback like, “Hey, you need to turn this, this, and this on in order to enable it.” After that, your Skills become available, and you can reference them by using a hashtag. In ChatGPT, you use an “@” sign. In Claude, you use a hashtag, and the reason you use the hashtag is because Claude reserves the “@” sign for agents, which is another show.

Katie Robbert: 06:57 – I was going to say, God forbid everything be standard across the platforms, but okay, exactly.

Christopher Penn: 07:05 – And so what I’ve got, I’ve already built a few just for testing. When I was testing it, one is to apply the Trust Insights brand standards to any project. So, “Hey, I’m in the middle of this thing, I made a cool infographic.” I’m like, “Oh, it doesn’t look like our brand standards.” So I can invoke the skill and say, “Use the Trust Insights brand skill. Rebuild this thing, make it look the way it’s supposed to with the correct fonts and stuff like that.” I have a co-CEO, one that I built based off of yours, Katie, but I added a whole bunch of stuff to it. So now it probably has more frameworks and things available to it.

Katie Robbert: 07:42 – I was going to say, would real Katie CEO approve of the things that you added that I haven’t seen?

Christopher Penn: 07:50 – Conditionally.

Katie Robbert: 07:53 – Front row.

Christopher Penn: 07:55 – And, of course, the most important one, at least that we recommend people have, is “Ask the ICP,” which is actually loaded with seven different ICPs. When you invoke it, it will say, “Which one do you want to use?” because this is based on what was in our sales playbook to make Claude Skills. You just have a conversation with Claude. So there’s no special interface, and they make it really simple. In fact, there is a Skill Creator skill, called Skill Creator, to allow you to build these things. When it’s done, you get a zip file. If you ever use and installed a WordPress plugin and you get a zip file, you just drag it into WordPress and turn it on—exactly the same thing.

Christopher Penn: 08:37 – That’s what I find makes this really cool that you can’t do with the GPT or Gem. I could zip up co-CEO and just hand it to you, Katie. Now you have co-CEO in your instance of Claude, if you had a separate one, and we all use the same one.

Katie Robbert: 08:51 – I see what you’re saying.

Christopher Penn: 08:52 – But I could put, say, my Trust Insights brand thing, I could put that on Google Drive, and if we had multiple instances of Claude, people could just drop it into theirs, and they’d be up and running. If I was really feeling enterprising, maybe I had a data analyst skill. I could package that up, put it on Gumroad, and put it up for $49. Anyone who bought it would get the zip file, drop it into their Claude, and suddenly their Claude has that skill. You can’t do that with GPGs and Gems.

John Wall: 09:22 – So when they, when you first of all, do you export it just from the window there at the normal interface?

Christopher Penn: 09:29 – No, you export from the builder. When you build it, when you.

John Wall: 09:33 – And then when it’s passed off, is it visible? Can somebody open it up and see everything that’s in there, or does it remain opaque so they can’t see what you actually built?

Christopher Penn: 09:43 – No, it’s just a straight-up zip file, and like a WordPress plugin, you can open it right up and just see what’s inside the box.

Katie Robbert: 09:52 – I assume we’re going to be doing a demo of this.

Christopher Penn: 09:55 – We will, but I’ll show you what’s inside the box. So inside the “Ask the ICP,” you get our IC, there’s the ICPs and personas, which is straight out of our sales playbook. Right. So there’s, it’s literally a copy-paste from our sales playbook. So there should be no surprises there. Claude creates itself a skill system, instructions that it uses so that when it is invoked, it knows what to do. That’s the content. The thing that’s really interesting is you can stuff a lot of stuff in here, so it doesn’t have to be just this. This is like the slim version of one. But you could put in multiple support documents, examples, you could put in templates, you could put all kinds of things in the same way you would for a WordPress plugin.

Christopher Penn: 10:40 – When Claude is told what to do with those things, it knows how to manage them and what to make of them.

Katie Robbert: 10:49 – It strikes me as just another way to build a custom thing. I was building custom gems in Gemini earlier. You have the system instructions, but then you can also add additional documents for contextual knowledge. This seems very much like that, where you can add as much context as you need to, provided you work within the limits. So I built one that gathers business requirements. You have the system instructions, and that is the contextual knowledge. It was the deep research I did on the SDLC, and I could also add in the way Trust Insights onboards or something like that. It seems very similar in that respect.

Katie Robbert: 11:35 – If I’m breaking it down to its most simplistic thing, you have the skill, which is the system instructions, and then all of the background information that supports that skill for it to reference.

Christopher Penn: 11:47 – Exactly right. Exactly right. So, that’s why this is valuable. The thing I really like about it is the portability of it. The fact that I could put that thing on a SharePoint and suddenly add everyone at a company who’s, we have one client that is a Claude shop. They can take their skills, put them on their SharePoint, drop them in, and they’re up and running. Let’s look at how to build one of these things. The nice thing is it’s done right in the chat. There’s no special interface, there’s no crazy stuff. It’s just literally just making a thing. But we still have to obey general best practices for software development because we’re making a piece of software.

Christopher Penn: 12:38 – So you can just wing it and say, “Hey, I want a skill that does this.” But it’s probably not going to be the best use of your time, and there will be a lot of trial and error if you do it that way. So, Katie, is there something that—

Katie Robbert: 12:50 – Before we get to that, Chris, we do have a clarifying question from Jessica. So, is this just Claude’s version of a custom GPT or a Gem?

Christopher Penn: 13:00 – No, that’s where this gets confusing. Claude’s equivalent of a GPT are Claude Projects.

Katie Robbert: 13:06 – I thought they were artifacts.

Christopher Penn: 13:08 – No, artifact is what they call Canvas.

John Wall: 13:13 – Who’s on here?

Katie Robbert: 13:15 – Exactly. Thank you, John. Thank you, John. All right.

Christopher Penn: 13:22 – It actually gets even worse than that because OpenAI’s project is different than a Claude project.

Katie Robbert: 13:28 – Let’s not bring them into this. No, let’s not. They’re, we’re booting them from the chat. We’re staying focused on Claude. Okay.

Christopher Penn: 13:37 – Okay. So we can sit, and we can. We’re in this interface here, which is the standard stock interface. We can build any kind of skill. So what would you like to build?

Katie Robbert: 13:52 – Since we’ve been talking a lot about the SDLC for the past couple of days, for those who don’t know, the SDLC is the software development life cycle. Actually, I have a better idea, because that could get unwieldy, and we only have so much time on the live stream. Let’s actually build a 5P framework skill. So that way, if you’re in a conversation, and you’re kind of going off the rails, or you’re like, “I forgot why I was doing this,” you can invoke the 5P framework to bring you back to where you started. Like, “Here’s what I know about your conversation so far, here’s your purpose, here’s your people,” and so on, so forth. So I think that would be a useful skill.

Christopher Penn: 14:35 – Okay, so let’s do that. Here’s how we’re going to do this. We’re going to go first to the Trust Insights website. You can get, and you can get this at Trust Insights AI/5P framework. We’re going to grab the 5P framework, the PDF version of it, and we’re going to just drag and drop that right into our window here. You can read. It says “upload fail due to a network connection.”

Katie Robbert: 14:59 – Great.

Christopher Penn: 14:59 – Love that. Okay, I’m going to switch to the web version because it is unhappy with me. There we are. There should be no network problems now because you’re literally on the Internet.

Katie Robbert: 15:15 – You jinxed it.

Christopher Penn: 15:16 – Theory. I clearly have. In theory, I have. Unless.

Katie Robbert: 15:22 – While you’re doing that, to answer Brian’s question of, “Can’t we get these foundational model companies some branding help?” Ideally, yes, but they’re named by developers and no shaded developers. They have no business naming things.

Christopher Penn: 15:37 – I can’t challenge that.

Katie Robbert: 15:41 – It’s almost like I’ve been there, done that.

John Wall: 15:43 – John, if your product name has a decimal point and more than four digits, back to the drawing board.

Christopher Penn: 15:50 – Mm. So we’re going the long way around. I said, “Oh, good, read the PDF,” which is the least efficient way to do this possible. But that’s what the PDF says. So now we could say, “Let’s build this into a Claude skill using,” and we’re going to invoke it Skill Creator. Now, we want this to basically, in conversation, allow someone to say, “Have I thought,” like, “How do you want it to work?”

Katie Robbert: 16:26 – I think there are a couple of different things. So one is you’re in a conversation, you are starting a conversation. You’re like, “Let me use the 5P framework to structure what I’m thinking.” Or the other scenario could be, “I’m all over the place. How do I bring this back? Let’s invoke the 5P framework.”

Christopher Penn: 16:47 – Got it. The purpose of the skill is to help a user either frame out a thought using the 5P framework or to organize their thoughts when they feel overwhelmed or distracted, or all over the place, and use the 5P framework as a framework to reframe their thoughts. So this is a consulting skill, an advisory skill, where the skill will be a trusted advisor to the user. As always, we ask the meta question, “Ask me one question at a time until you have enough information to successfully complete this task.” Let’s see if we can get the Skill Creator to invoke itself and to come up with some designs. All right, so it’s going to help me build this skill? Reading the Skill Creator documentation, its own Skill Creator is, by the way, a very large thing of its own. Okay.

Christopher Penn: 17:43 – It says, “Can you give me two or three examples of what a user might say?” Those are really good examples that you’ve listed there: “I’m feeling overwhelmed,” or “Help me think this through?” Yeah, that’s exactly what you would say.

Katie Robbert: 17:58 – That’s the idea. Yay.

Christopher Penn: 18:02 – Yay. It’s pretty straightforward. How should Claude do this? Should it ask all of them upfront, walk them through one at a time, or assess which one they need the most help with?

Katie Robbert: 18:15 – I’d say two and three. So first, it should assess which P’s the user needs help with and then walk the user through each P one at a time.

Christopher Penn: 18:24 – Okay, so first, assess which P the user needs help with most based on what they’ve shared, and then secondarily, walk them through HP1 at a time sequentially so that the skill has enough information to be helpful. When Claude has gathered enough information, what should the final output look like? A structured summary, strategic recommendations, an action plan, all of the above, something else?

Katie Robbert: 18:59 – I would say all of the above. It should provide a summary, then the recommendations of next steps. Depending on what the thing is, there could be a lot of action, so helping it prioritize, like, “Okay, you have chosen a platform, but you don’t have a purpose. So start there.”

Christopher Penn: 19:20 – What tone of voice should I have? Directive and prescriptive? Socratic and question-based? Balanced? Empathetic and supportive? Business and efficient?

Katie Robbert: 19:29 – Well, Chris, with the gem that I built earlier today, you said it sounded incredibly optimistic, and I asked if I should make it sound more mean. So I mean, my gut says it should be somewhere between direct to prescriptive and empathetic. If you want it to be all business and efficient, I’m fine with whatever. I don’t get easily offended by machines.

Christopher Penn: 20:00 – Make it a good balance of directive and prescriptive, and empathetic and supportive. That’s not contradictory at all, actually.

Katie Robbert: 20:10 – Well, I mean, you best. You basically just describe me.

Christopher Penn: 20:15 – That’s true. All right, so now what it’s going to start doing is it’s going to start building the skill. Skills are composed of a master skill document called Skill MD, which is, you can see in the download that has some front matter, which is basically like a little miniature table of contents in the summary. Then it has the system instructions that it builds. Then if there are supporting documents, it will typically include them. Once it’s done, it makes a little zip file, and that zip file you download to your computer. Then in whatever version of Claude you’re using, you can drop that into your account, and it becomes available across your account. That part’s also really important: no matter what version of Claude you’re using, the skill becomes available.

Christopher Penn: 21:01 – It says, “Why did you? Why did you fail? Remove the template. We’re burning tokens here.” Oh, dear.

Katie Robbert: 21:14 – It’s a day ending in why it is.

Christopher Penn: 21:17 – One of the things that’s really interesting is that Claude, in particular, burns a lot of tokens. It overthinks everything. It has packaged up the skill, and there it is: 5P framework. I don’t know why it tries to show you the binary file, because that’s completely unhelpful and makes you think that something broke. Nothing broke. Because if I look at my downloads folder, I have the 5P framework skill. This is just the skill document itself. There are no supporting files because the PDF was enough. It put that in here. You can modify this before you upload it. You could modify it if you forgot something. So let’s take that raw, the final version. Let’s go back into our settings. Here, let’s go into settings, capabilities, skills, and upload our new skill.

Christopher Penn: 22:18 – We just drag and drop our zip file in. It says, “Skill upload successfully, 5P framework.” So we have that skill available. Now, if I start a new chat, and let’s say today I want to think through how to create a fun, mildly entertaining video game that’s a weird mix of Donkey Kong and Contra. That’s a very strange request, and we’ll see what it comes up with. What I’m going to prompt it next is, “Oh, I forgot this is going to be an important project, so I probably should use a skill. Let’s use the 5P framework skill,” and we’re going to say “available at 5P framework.”

Katie Robbert: 23:14 – I feel like that’s a good plug for some governance around naming conventions. Make sure that you’re naming things in a way that makes sense, and perhaps, depending on how large your organization is or how many different people have access to create skills, maybe create a master dictionary or some sort of inventory of the skills that have been created and what they’re meant to do. Just a quick plug for that. While that’s running, a quick question from a LinkedIn user: “I missed the start of this. Will it be available to watch again?” The answer is yes. Go to Trust Insights AI, YouTube, go to the “So What” playlist, and all past episodes should be there.

Christopher Penn: 24:07 – So it did exactly what was requested, which is it pulled up the framework, and it looked at it, and it started with. It identified the most important. “Why are you doing this?”

Katie Robbert: 24:17 – “Why are you doing this?” That is a great question to start with.

Christopher Penn: 24:21 – “What problem does this game solve? What gap does it fill?” Even entertainment has a purpose. This is about nostalgic arcade vibes, specific gameplay itch, or proving you can execute on a weird concept. “What’s driving this project for you?” Proof that we can execute on a weird concept.

John Wall: 24:40 – It’s giving Mario guns. That’s really what we want.

Katie Robbert: 24:45 – But see, there’s the empathetic part of it. It’s like, “No, I can see that you’re trying to do something. So before we just shoot it down and say no, let’s make sure that we’ve sort of really thought this through.”

John Wall: 24:55 – Let’s give you a chance to explain that, stupid.

Christopher Penn: 25:01 – I was thinking about making this for a customer. What if we use our ICP skill? Use “Ask the ICP” to help with our people. Now, let’s see if we can invoke that second skill in the conversation.

Katie Robbert: 25:20 – This goes back to what our friend Brian was saying earlier: he likes to bring in custom GPTs into one conversation for a focus group. This is a great way to think about why you would have different skills versus one custom project. Trying to have one thing that does it all is never a great idea because it can’t certainly. You can’t do everything great every single time. But having individual purposes for each project and then bringing them into the conversation is actually going to make whatever it is you’re doing that much stronger.

Christopher Penn: 25:58 – It said, “Hey, I loaded your ICPs. This makes no sense at all. Are we, are you actually planning to make this Donkey Kong Contravention for one of your customers?” Because that would be really interesting and weird in a good way.

Katie Robbert: 26:17 – I mean, this is about the time I would tell you to invoke the CEO to start asking, “What the heck are you doing?”

Christopher Penn: 26:23 – Exactly. Maybe we need extra help from our co-CEO skill. To clarify, let’s ask them. Use that skill.

Katie Robbert: 26:38 – This is not going to go well.

John Wall: 26:39 – This is going to be funny if it just, like, straight up vetoes. It’s just like, “What the hell are you doing?”

Katie Robbert: 26:46 – Well, I mean, according to you, the CEO skill that you built was a lot meaner than me.

Christopher Penn: 26:51 – Yes, it is. But you could see how, to your point, Katie, if you have a mental palette of all the different skills that you’ve built, you don’t have to use them all at once. You can bring them in as needed. Claude also is smart enough to look at the conversation context and say, “Hey, I have this skill, like the Ask the ICP. So, by saying you’re talking about ICPs a lot, maybe I should use the Ask the ICP skill,” which is.

Katie Robbert: 27:18 – What happened the other day when I was using Claude? It brought in the co-CEO skill without my asking it to. Which makes sense because I’m me, and I was the one using it.

Christopher Penn: 27:30 – Exactly. So the co-CEO says, “I have questions.”

Katie Robbert: 27:36 – It really is me.

Christopher Penn: 27:39 – “Is this a business project or is this a personal side project?” If it’s a business project, what’s the use case? “How does building a video game align with this at all? If it’s a personal project, that’s fine. When you say ‘for a customer,’ what do you mean? Is this for a client? Is it for sale? Or is it people who play video games?” So essentially, this is really is you. It’s basically saying, “Yeah, these are.”

Katie Robbert: 28:02 – The exact questions I would be asking you.

Christopher Penn: 28:05 – Exactly. These are the Claude skills. This is essentially what this is. You build these skills, then you get the download, and the download again is nothing more than a zip file. If you didn’t want to use the Claude Skill builder, all you need to do is have a root file at the top of a folder called Skill MD that has the expected structure, which, by the way, this structure is identical to Claude Code sub-agents. If you’ve used those, you know exactly how to build these. It’s the name, the description up top as a little bit of front matter, and then your system instructions after that. Within there, you would have a Docs folder with reference documents if you wanted, with templates, with whatever the thing is.

Christopher Penn: 28:54 – For example, let’s say you are an agency that builds websites for customers, and you have a stock list of, like, 10 different templates that you use. You could have a skill that interviews the user that says, “Of these 10 different templates that I know that we produce for our clients on a regular basis, which template should we start with?” Have that conversation, and it can immediately bring stuff out of its own little archive here and say, “Okay, we’re ready to start working on template 3 or template 4,” or what have you.

Katie Robbert: 29:26 – If there are a couple of assumptions here. Number one, you have to know that the skills exist. Number two, you have to know what the names of the skills are. Is there a scenario where you could say, “Here’s what I’m thinking, what skills are available to use,” and would it give you a list?

Christopher Penn: 29:50 – That’s a great question. Let’s see if it knows because I don’t know if it’s aware of that. Ted was asking in the chat, “Can we invoke skills within a project?” Yes, you can, and that’s very, very useful. Oh, look, I can just ask it. Claude says I can make Word Docs, I can make PDFs, I can make PowerPoint presentations.

Katie Robbert: 30:17 – Shut the front door. We’re going to test that one out. Okay, sorry, keep going. I just got really excited.

Christopher Penn: 30:26 – I had the Trust Insights business tools, the 5P framework. Ask the ICP, the co-CEO, and the brand, and they have some skills development stuff. So these are the things that I, it says that’s what the, what’s available to you.

Katie Robbert: 30:38 – But that’s really helpful because I feel like if you are, it’s hard if you’re like midstream thinking, like in the middle of a conversation with your LLM, and you’re like, “Oh, I could really use some help on this.” I think we had something that could do that. It’s nice to know you can ask, “Is there a skill that can help me with this?” It might just give you this list again, which is fine. Then you would have to invoke it or, theoretically, be like, “Yeah, you have an ICP. I’m going to bring it right into the chat.” My other question—and you may not know the answer to this—is, I’m assuming the answer is no.

Katie Robbert: 31:17 – Like, let’s say I brought my ICP “Ask the ICP” skill into the conversation and then decided, “I don’t actually want the ICP. It’s too constraining.” Can you boot the skill from the conversation?

Christopher Penn: 31:33 – No, because at that point, it’s been loaded into the context window.

Katie Robbert: 31:36 – Okay, that’s what I thought you were going to say, but I figured it’s always worth asking.

Christopher Penn: 31:42 – Yep. I’m now curious about the PowerPoint thing. Let’s say.

Katie Robbert: 31:46 – Yeah, I am too.

Christopher Penn: 31:47 – If I provide you with a slide deck template, can you use the PPTX skill to modify it? Let’s ask, and I’m going to invoke it and see what it says.

Katie Robbert: 32:07 – I mean, I can’t see our audience, but raise your hand if you’ve been waiting for this one.

Christopher Penn: 32:11 – It says, “Upload your template. I’ll read the skill, I’ll modify it, and deliver the results.” All right, well, I have our Trust Insights master slide template here. Let’s see.

John Wall: 32:26 – Where?

Christopher Penn: 32:27 – All right, so here is our slide template. We’re going to turn this into a 10-slide sales presentation. Our target customer that we’re going to be giving this presentation to, we’ll call Skill Point Consulting, and we are pitching them on an AI project. Let’s go ahead and get started framing out what this presentation might have in it using the PowerPoint skill. Let’s see if it works. Let’s see. I read that by all the best practices, “Examine your template.”

John Wall: 33:12 – It’s funny, they should have the token wheel spinning berserkly.

Christopher Penn: 33:18 – It does have a usage meter. It’s in the preferences. You can see how much of your—

John Wall: 33:22 – Oh yeah, you can turn it on.

Christopher Penn: 33:23 – To watch weekly allotment. You’ve burned up fuel burning. Let’s see. It’s creating a thumbnail grid of template slides for visual analysis. Now let me see the visual thumbnails.

Katie Robbert: 33:44 – I think we’re all holding our breath, hoping that this actually works for once.

Christopher Penn: 33:50 – Yeah. For folks who don’t know, this is the document that is in question. It’s just literally our blank template we got from our designer. Let’s see, section two design elements. So architecting presentation structure. “Show me your thinking.” I want to see what you’re thinking here. Plain text tent slides. Today’s data is correct. The challenge section divider. Okay. So it knows that our section dividers exist, which is interesting.

Katie Robbert: 34:28 – Well, if you look at the PowerPoint, that’s the slide template that’s selected.

Christopher Penn: 34:34 – Yeah.

Katie Robbert: 34:35 – Because each slide, you choose a template from PowerPoint. Like, I want a header slide or text and body, or two side-by-side, or text and image. So it’s probably looking at whatever that background information is, slide by slide, and saying what’s available to me to use. So I’m assuming—and this is a big assumption—that unless you give it all the different slide templates that you may want to use, it’s only going to use what you give it. So if you want to use section headers or dividers or whatever they’re called, but you don’t give that to this, it’s not necessarily going to say, “Oh, and by the way, that’s a big assumption.”

Christopher Penn: 35:24 – That is a big assumption. Interesting. So it’s got a combination of Python code on the backend and object manipulation to try to essentially frame out. I see what it’s doing. That’s very clever. What it did is it took apart our slides. Because PowerPoint, a PPTX slide, is nothing more than a big XML document with some images attached to it, so it basically deconstructed that, figured out what fields are where on our slides, and was able to understand the general structure of our slides and what goes where. Because every object in a PowerPoint presentation has an ID number, like a text field is ID12, an outline is ID14, etc.

Christopher Penn: 36:15 – It’s essentially disassembled that, looked at our content, put together an outline in a JSON object, and now it looks like it’s trying to reassemble the slides with the changes to the XML tree. Yep. In fact, that’s literally what it’s doing to try and rebuild the slides with the changes to them. I don’t know how long this is going to take, but that’s a very clever way of doing it. The only other service I’ve seen that does something like this is Moonshot AI’s Kimmy K2 model, which you can’t upload a template. It has its own templates, but it knows what regions those templates it’s allowed to modify.

Katie Robbert: 37:00 – This is another plug. This is great, but first, you still need to go through probably the 5P framework to figure out what it is you’re trying to create. You need to have the template, so you need to have all that background information first. Like, “Here’s my template, here’s what I’m trying to accomplish, here’s what the purpose of this slide deck, here’s who it’s for, how we’re going to present it.” We want slides that include slots for video, or we’re going to cut out of the slide deck to do a live demo, and then the overall. So, you still want to do your upfront requirements gathering before you start using a skill like a slide builder, just like you would any other PowerPoint deck.

Katie Robbert: 37:47 – When I was building our AI Ready Strategist course, I had that outline, like slide-by-slide outline, and then I just had to fill it in. Chris was able to create an application that did that for me automatically and got me, 60% of the way there, which was super helpful. But I would still, for this, I would still be like, “Here’s our slide deck template, here’s my 30-page outline. Now can you put it together?” That would definitely be helpful because the widget, Chris, the application that you had created, didn’t take the full template. It was just a bunch of blank slides, which was still really helpful.

Christopher Penn: 38:28 – It says the presentation is complete, and now it’s got to move files around, and it’s providing an outline of the presentation itself.

Katie Robbert: 38:50 – It’ll be interesting to see what the output is because if it’s just an outline, we didn’t need to use the skill for that.

Christopher Penn: 38:56 – Right. It will be very interesting to see that. I will be honestly surprised if it finishes at all because, knowing the backend of what it’s doing, this is a very, very computationally intensive process. This is probably churning up a kilowatt of power just to do this.

Katie Robbert: 39:20 – So let’s add a little bit of context to what that is that you’re saying. Anyone who’s using any of these large language models, we’ve talked about how each of these large language models use tokens. Every word, every ask, everything you’re trying to do with it, say you have a bag of tokens that renews every week. As you’re asking these large language models to do something, you’re withdrawing those tokens from whatever your allotment is. I think that changes with whatever it is you’re paying. So the paid. For the free, you probably have certain limits; for paid, so on and so forth. It sounds like, Chris, if I’m following, asking it to do a skill like create a PowerPoint is very—

Katie Robbert: 40:04 – When you say “computational heavy,” it means if you start with a thousand tokens, you might end up using 999 of those tokens just by doing this one thing.

Christopher Penn: 40:15 – Yep. In our Claude account, we are on Claude Max, which is the most expensive plan. Oh, good. Before we started, we were using about—we were around 12% of usage for the week because this is the same model we use for a lot of coding. Just the 10-slide deck alone used 2% of the max plan’s computation, which is not an insignificant amount of compute that it used. However, it successfully figured out our slide templates, put it in the correct fonts, used the section headers appropriately, and came up with essentially a decent starter that’s in our template.

Katie Robbert: 41:04 – Yeah, that’s not terrible.

Christopher Penn: 41:06 – Yeah, it definitely could use some more images and graphics and things like that. It violates the general standard of no more than nine words on a slide. But it’s a start.

John Wall: 41:19 – I’m thrilled it opens. That’s really. I was like, that’s not going to open. You can’t mess. You can’t plug the metadata like that.

Christopher Penn: 41:26 – And make it work so.

Katie Robbert: 41:28 – Well. And I feel like everyone has different standards for how they create slides, so that’s probably something additional you’d want to give. Like, “Here’s our brand guidelines,” or “Here’s my best practices for creating a slide deck.” Because, Chris, you and I do it very differently. So we’d want to include that kind of information in the prompt upfront. Like, “You must not use more than nine words on any given slide,” or “Add clip art wherever applicable,” which is your favorite thing to do.

Christopher Penn: 42:01 – Exactly. But again, it managed 10 slides. It did. At the very end of the process, it did crash. So it skated through just barely to the end before it blew up.

Katie Robbert: 42:17 – What didn’t it do?

Christopher Penn: 42:19 – No, it did everything. But it then gave me a lost connection with Claude error message. So I don’t know that I would try a 200-slide epic in here. I think that could go very badly.

Katie Robbert: 42:33 – Well, that’s what I was getting to. When you’re saying that tasks or skills like this are computationally heavy, and for us at Trust Insights, it used 2%. Imagine exponentially more at an enterprise-sized company where everybody’s trying to get their weekly reports done or their monthly reports done, and all of those slide decks are in the hundreds of slides, and everybody’s using the one same account all at the same time. That’s when you’re going to start to see those tokens get eaten up really quickly.

Christopher Penn: 43:11 – Exactly. The difference between the plan we’re on and your average 20-month plan is 20x. So if we use 2% of our compute budget for the week, because Claude meters weekly, on the lowest paid plan you will have, you would have used 40% of your week’s compute on a single slide presentation.

Katie Robbert: 43:37 – And that is significant.

Christopher Penn: 43:38 – That is significant. That is problematic if you were also using it for other things because everything counts against your quotas. This is true of all the systems and all the vendors. But in general, these skills are super handy, and you can see how you can invoke them. If you think about it like those basic document skills, if you’re having a long chat and conversation, you could at the very end and say, “Hey, turn this into a slide deck for me,” right?

Christopher Penn: 44:07 – So not just create a summary but actually turn this to a slide deck. Or you’re working on some accounting and say, “Put this into an Excel spreadsheet for me,” and it’s capable of doing that in the same way that in Gemini, you can say, “Put this into a Google Sheet for me,” and it will also do that in there. The skills are portable, which is super handy. You can build once and hand it to as many people as you want. I have already seen people on LinkedIn offering their Claude’s Ultimate Skills pack for $99, which, as we saw, is just nothing more than their previously prompted, their previous packs of prompts they used to sell. But it is super useful, and we didn’t cover it because I would have to go and set it up. But the skills are usable within Claude Code.

Christopher Penn: 44:59 – So, if you’re using the Claude coding tool to do batch tasks, I’m going to be testing this myself. I have Katie’s writing style, and I have several hundred transcripts of YouTube minis that I’ve done over the years. I’m going to put 10 of them through and see if I can have it rewrite my transcripts as though Katie was writing about those particular topics, and see what comes out using Claude Code and a “write like Katie” Skill.

Katie Robbert: 45:29 – Hey, friends, here’s an agentic AI, because for some reason, you think that’s how I sound.

Christopher Penn: 45:38 – I don’t think that’s how the machines think you sound.

Katie Robbert: 45:42 – Well, we’ll be interested to see the results of that, and perhaps we’ll post some of those results in our free Slack community, which you can get to at Trust Insights AI Analytics for Marketers. Free to join. Always a good time.

Christopher Penn: 45:57 – Questions? Every day.

Katie Robbert: 45:59 – Yeah, different questions. Fridays are unhinged. Tomorrow should be interesting.

Christopher Penn: 46:04 – Every day is unhinged.

Katie Robbert: 46:06 – No, this is true. But on Friday, we purposely have unhinged questions.

Christopher Penn: 46:11 – Any final thoughts on Claude Skills before you go off and start creating your own?

Katie Robbert: 46:16 – I don’t know, John, what kind of skills do you think would be useful from your perspective on the sales side, on the biz dev side?

John Wall: 46:25 – The thing is, that’s the first time I’ve actually seen a slide builder that I was like, “Okay, yeah, that could be used.” I don’t even care if it burns up all the tokens on a side account. If I can throw an outline in there and get 50 slides with the outline plugged into it, that’s huge. So I’m taking the win there.

Katie Robbert: 46:44 – All right.

Christopher Penn: 46:47 – All right. Well, that’s going to do it for this week’s show, folks. Thanks for tuning in, and we will catch you all on the next one. Thanks for watching today. Be sure to subscribe to our show wherever you’re watching it. For more resources and to learn more, check out the Trust Insights podcast at Trust Insights AI, a weekly email newsletter at Trust Insights AI. Got questions about what you said in today’s episode? Join our free Analytics for Marketers Slack Group at Trust Insights AI Analytics for Marketers. See you next time.

Need help with your marketing AI and analytics?

So What? Marketing Analytics and Insights Live

airs every Thursday at 1 pm EST.

In this episode you’ll learn:

Transcript:

One thought on “So What? Getting started with Claude Skills”

Leave a Reply Cancel reply

Subscribe to our Weekly Newsletter

Pin It on Pinterest