This is a live recording from Localist Lab, the in-person version of The Localist, where specialists come in to talk to small business owners about the work everyone’s trying to figure out. This month Carrie was the speaker, talking about the thing that finally made AI click for her: using AI well isn’t a tech skill, it’s a management skill. Anyone who has trained an intern, written an SOP, or told somebody how to cover their desk while they were out already has most of what they need.

The session walks through how to build an AI skill start to finish, first using Infomedia’s new hire Emery as the comparison and then using the podcast as the working example. It covers the 70-30 rule for deciding what to hand off, why a skill is really just a standard operating procedure written in plain English, why the context window works like a meeting whiteboard, and the eight things a good set of instructions has to include. Carrie also shows the skills she actually uses to make The Localist: a researcher, an interview writer, a title writer, a social media writer, and a strategic advisor skill that the others pull from.

There’s also a story about an espresso machine named Richard that will convince you to read your AI’s research before you sit down at a microphone. Then the audience asks questions on whether skills are just fancy form letters, whether you need to pay for an AI account, and what actually happens to your data once you put it into a chatbot.

If AI still feels overwhelming, or if you’ve built a skill and it keeps handing you mediocre work, this session gives you a real process for fixing it.

Watch the Full Episode on YouTube

Build an AI Skill: Topics Covered in This Episode

Why using AI well is a management skill and not a tech skill
Onboarding an AI the same way you onboard a new employee
The strengths and the weaknesses your AI shows up with
Why giving feedback is the step most people skip
The 70-30 rule for deciding what to delegate and when
What a skill actually is: a saved set of instructions in plain English
How to build an AI skill by just asking your AI to build it
Breaking a big project into several small skills instead of one giant one
The context window explained as a meeting whiteboard
The five skills Carrie uses to make The Localist
Chaining skills together so one pulls from what another created
Why Carrie keeps herself in the middle of the research step
Richard the espresso machine, and why you check the research
The eight parts of a good SOP: task, parameters, decision points, process, mission, audience, meta prompt, examples
Why three examples beat a perfect explanation
Giving your AI long-form writing so it gets your short-form voice right
Stress testing a skill on real work instead of a made-up test
Auditing results and fixing the skill instead of the output
Where to start: something personal, something annoying, or something constant
Audience questions on skills versus templates, paid accounts, and AI training on your data

Mentioned in This Episode

Follow Carrie Rollwagen

About the Speaker

Carrie Rollwagen is senior vice president of strategic planning at Infomedia and the host of The Localist. She is a former bookstore owner, a career writer, and the author of “The Localist: Think Independent, Buy Local, and Reclaim the American Dream.” She teaches AI classes for teams alongside Russell Marbut of Infomedia.

Listen to The Localist on Spotify & Apple Podcasts

Thanks to Our Sponsor, Infomedia

The Localist is sponsored by Infomedia. They are a Birmingham-based web and digital marketing company. They help small businesses grow online.

Contact Infomedia

Join Us at Localist Lab

Localist Lab is a free event series for small business owners. It meets on the third Thursday of most months at Saturn in Avondale. Free tacos and coffee too.

See upcoming events and register

Subscribe to Carrie’s Newsletter

Want more small business tips from Carrie? Sign up for her newsletter.

FULL EPISODE TRANSCRIPT

Below is the full transcript of this Localist Lab session with Carrie Rollwagen. This transcript is provided for accessibility and SEO.

_______________________________________

Carrie (00:12): Hi, I’m Carrie Rollwagen, host of The Localist. What you’re about to hear is a live recording from one of our Localist Lab events. This is when we bring in experts to talk about practical insights on marketing for small businesses. This session was recorded in front of a live audience, so you may hear some room noise and some shuffling around during the Q and A. But hopefully that just makes you feel more like you are actually there. For more background on the speaker, or to view slides from the event, check out our show notes, or you can also grab a ticket there to our next event. I hope you enjoy this conversation.

Carrie (00:48): Welcome to the Localist Lab. My name is Carrie Rollwagen. I am the host of The Localist podcast. The Localist goes out every week wherever you listen to podcasts, and I have local business owners come on the podcast every week and talk about their experience: what’s working for them, what is not working for them, what tools they’re using, all of that kind of thing. So I hope you will like and follow and all of that stuff.

Carrie (01:20): But the Localist Lab is the live version of The Localist podcast. Here, instead of having business owners come on, we have different specialists from around the community come and talk to local business owners like you, and people who work in local businesses, about how to do the things that we are all struggling to do: work on marketing, work on our website, use AI, the things that we’re going to be talking about today.

Carrie (01:50): The Localist and the Localist Lab are sponsored by the Lovoy family of companies. That includes Infomedia, Uptick, Tempo, and Infomedia Studios. Between all of our sister companies, we help with pretty much everything digital. Whether it’s your website or your marketing or your social media, talk to one of us. Most of us are wearing name tags, if you’re interested and if you’re not a client already.

Carrie (02:11): So today we’re actually going to just dive in, because I am the host and the speaker. Again, my name is Carrie. I’m senior vice president of strategic planning, which is kind of a mouthful, at Infomedia, and we’re going to be talking about the topic: if you can write instructions, you can build an AI skill. I’m happy to have everybody here today.

Carrie (02:45): Oh yeah, I’m supposed to promo this class. Sorry, we don’t spend a lot of time selling stuff here, but we do have a paid class that we offer if you want to deep dive more into AI. If you are wanting your whole team to come and learn about AI, we have a couple of options. We partner with Caravan, who are AI trainers, if you want to do a really full day deep dive course. And then we also have this jump start course, where Russell, who’s also from Infomedia, and I teach AI basics. This is a couple of hours in the morning, and your team will leave understanding Claude, hopefully, and actually building a skill and doing some practical work. So that’s my promo for that, and I’ll put that up at the end. We normally just sell that to teams, but we’re going to do a special session on Sept. 2 where you can buy a single seat if you’re interested.

Carrie (03:35): Okay, so AI. We’re all at different points in the spectrum on AI, right? Most of us, if you’re here, you’re using it in some capacity. But some of us are using it and we’re just like, okay, this is overwhelming, I don’t feel like I really know how to use this well. Some of us are already experts and we’re creating new agents and creating all of this stuff all the time. But for a lot of us, we do find this to some extent kind of mysterious and scary, hard to learn. It feels like this overwhelming technology that’s changing all the time. And how can we possibly keep up?

Carrie (04:10): It is overwhelming. It is changing all the time. But the good news is we can keep up, because using AI well is actually not a tech skill. It’s a management skill.

Carrie (04:25): For me, this was really the huge unlock that changed it from AI being something I have to learn because my boss is making me and the world is going in this direction, to something where I actually feel empowered and I understand how to use this.

Carrie (04:55): So what does that mean? If you manage a team at work, or you’ve ever trained an intern, or just told somebody, here’s how you cover the desk for me while I’m gone, if you’ve ever trained a person, or if you’ve ever written instructions, and that could be an SOP or it could just be instructions for the babysitter or a pet sitter, if you know how to do that, then you can use AI. You can get it to work for you. Onboarding your AI and teaching it skills is actually a lot like onboarding a new employee.

Carrie (05:35): All right, so let’s meet Emery. He’s over there. He’s super, super pleased to be on the screen. This is also mean of me, because this was a Halloween costume, but I thought it was funny. Emery started working for us a few months ago, and this was his first job out of school. He’s a blank slate. Just kidding, he did know stuff. But when Emery started working for us, he had some strengths and he had some challenges, just like anybody who works for you does. He has a lot of potential. He’s smart. He’s very helpful. He helps us understand what Gen Z is saying. He knows jujitsu. That hasn’t come in handy at the office yet, but it could. But he also has some challenges. It’s his first professional job, so the way we email is different, the way we communicate is different. He’d never coded a site before, so we have to teach him how to do that. We all know that when we have a human who comes into our organizations, we have to teach them how to work to some extent, and the same is true for AI.

Carrie (06:40): So when your AI agent comes on board, so to speak, and this can be different, there are different names for this. I’m going to use Claude Skills as an example a lot in this talk, partially because that’s what I use heavily and that’s what I teach. But whatever your AI is, what we’re going to talk about today is pretty platform agnostic. So if you’re thinking, how does Grok do this, or how does ChatGPT do this, or Copilot, just say, hey, Carrie said I could do this, how do I do that with you, and it will tell you. It’s usually called skills or an agent or something like that, but you want to onboard this agent into your workflow as well.

Carrie (07:10): Just like a human employee, the AI also has strengths and weaknesses. It has a lot of potential. It seems to know a lot. It works very quickly, and it doesn’t ask for PTO. But it also has challenges. It hasn’t worked for your company before, so it probably only knows what’s out there on the internet, and maybe it doesn’t even have up-to-date information about that. It doesn’t know your workflow. It’s very confident even when it’s wrong, and it’s kind of a suck-up. It’s also really good at some tasks, almost miraculously good, and then other things that seem pretty simple, it’s really bad at. So again, there are strengths and weaknesses that come with AI.

Carrie (07:45): One of the important next steps, after you start onboarding somebody, is to give feedback consistently. Again, here’s Emery learning from his elders. When we had Emery first adding content to a site, we showed him how to do it. But we didn’t just say, oh, that’s probably fine, let’s send it to the client. We looked at it. We assessed it. We said, this is good, this is okay, and actually when you do this, do this better, that kind of thing. And then we go through that process several times until he’s going to work autonomously. The same is true for AI.

Carrie (08:25): We have all seen things that come from AI where you think, this is so basic and this is obviously AI and not good. That’s because they’re not spending time in this stage. You have to create the skill, but then you have to assess the skill, update it, give it feedback, and do that for a few cycles before you get something that you can pretty much trust.

Carrie (09:00): Now, when we are delegating, whether or not this is to a new person who’s working for us or to AI, I see that the same people who struggle with delegating to people are also, in my experience, struggling to delegate to AI. I work with managers a lot, and I had a manager come to me and say, I can’t give anything away. They were talking about people, but this is also true of AI. They’re like, I can’t give anything away because I do it better. And I know we’re all a little judgy about that, but we all probably think that a little bit too, right?

Carrie (09:45): What I had to tell her is, you’re right. You’re working at 100%, and nobody is going to be able to work to 100. But you don’t delegate when somebody is at 100. You delegate when they’re at more like 70 or 80. So I call this the 70-30 rule. I just made that up, well, years ago I made that up. But if somebody is at 70 or 80%, you delegate to them then, and then they get to practice. You give them feedback. They get better. They’re not going to get to that 100% mark until you give them the actual task and work with them to get better. So that’s when you delegate, not when they’re at 100%.

Carrie (10:25): And that’s true of AI too, because you’ll hand something off to AI and be like, well, I still had to fix it. Okay, that’s not a reason not to give it away. That’s a reason to go and iterate on it more. On the other hand, when a person or your AI is at 30%, don’t delegate then. You’re going to overwhelm the person. You’re going to be frustrated with them. It’s not going to go well. Same with AI. But you may check in in six months, and your AI has learned more about you, and AI, of course, has grown leaps and bounds by that point, and you may be able to delegate then.

Carrie (11:05): Okay, so how do we actually teach our AI skills? Again, it does have a lot of correlation to how we teach people skills. The actual way to build an AI skill is just to tell your AI you want it to build a skill. So this is not going to be the big part of this talk, because the actual logistics are just that you tell it to build you a skill and then it builds you a skill. It may look a little different from this. It may give you a little thing to click to say yes, depending on whatever platform you’re using. But essentially you just say build me a skill, and it builds you a skill. There’s a lot of information that we need to give it to build the best skills, though. And this is a demo. I would never write a prompt that was this short. My prompts are ridiculously long.

Carrie (11:45): Okay, so what actually is a skill? A skill is just a set of instructions that is saved, so you don’t have to explain it over and over. Again, I’ll use Emery as an example. Emery sends the email that you all probably got yesterday that says the Localist event is tomorrow. That’s part of what he does, but he only does it once a month, so he doesn’t have to remember every step in that process. He goes back to his set of instructions. He follows the instructions, and then you get the email. AI is the same way. So what a skill is is actually extremely basic. It’s just a list of instructions, and it’s also written in English, as long as you’re using your AI in English.

Carrie (12:19): It’s not a weird coding language. And coding languages aren’t weird, sorry to the devs here. But it’s not in another language where you have to go and sort through everything. You can just literally look at that file, which is written in English, and see, why is this messing up? Oh, it’s probably messing up because these instructions are weird. And then you can fix it. So essentially, that is what a skill is. It’s just a standard operating procedure, an SOP. All of this stuff is AI technical, yes. Is it advanced and confusing sometimes? Yes. But it doesn’t have to be confusing to work with, because creating skills is actually a very basic thing.

Carrie (13:05): I want to use an example of making a podcast, because I make a podcast and I think it’s a good example. There are a lot of skills that go into making a podcast, and I think this example will also help you come up with ideas that correlate to what you’re doing. There are a lot of different tasks involved in making The Localist. These are just some of them, but there are more things. There’s just a lot of different steps.

Carrie (13:50): But I don’t want to write a skill that’s going to do all of those steps, for a variety of reasons. One of those reasons is that I don’t do all these things. I have a team that does a lot of the stuff for me, so that automatically knocks out a lot of this. Now, you can use skills where you’re interacting with your team members. You can share skills, absolutely. But I would recommend, when you’re just getting started and you want to get better at this, start creating skills that you use, that only relate to you, so that you can learn the process better and iterate on it and you don’t have a lot of different factors involved. You can kind of expand after that.

Carrie (14:35): So first I knock out everything that’s not going to belong to me anyway, and then I knock out things that I’m not ever going to outsource, either to AI or to a person. For example, record the interview is on this list. Well, that’s me sitting across from the person and asking them questions. I’m not going to outsource that, so I’m not going to build a skill for that, because I’m always going to do that personally. So then I still have research guests, interview, write titles, write social media, and those are good candidates for skills. If the task is big, like making the podcast, it’s good to separate it into separate skills.

Carrie (15:20): There are a lot of reasons for this, but one of the reasons is because you want to think about what’s called the context window. The metaphor that I like to use is that it’s kind of like a whiteboard in a meeting. If you’re sitting in a meeting and one person gets up and draws an illustration on the whiteboard, that’s helpful, right? It helps you understand what they’re talking about. But then if everybody gets up and they’re scribbling on that, and they’re like, well, let’s change that, and then they’re writing their own things, and then the whiteboard is totally filled with scribbles, you’re not only losing the new stuff, you’re losing the original stuff too. That is just too much information, and to some extent AI does the same thing. If you write a skill that is super long, for example if I said I’m going to write one skill for the whole Localist podcast, that would be so long. Or if I was going back and forth with Claude, like, should I include this, should I do this, should I do that, oh, never mind, I don’t want to do that, and then we’re taking that whole thing and throwing it into a skill. That skill is huge, and I’m also forcing AI to read through the entire skill every time I ask it to do something. If all I want it to do is write a title for an episode, that’s a pretty simple skill, but it would have to read through that entire thing.

Carrie (16:06): So I want to keep that context window small, keep it really clear and distinct and focused. I mentioned chatting back and forth about ideas to build a skill. That’s a really good thing to do, but when you’re done with that chatting and going back and forth, just tell AI, hey, I want to go build a skill, can you write me a summary, or write instructions to build a skill? And then it will summarize everything that was important in that chat. It will leave out all the extraneous things, and then you can copy that into a new window, so you have a new context window, and it will build the skill from there. So just try to keep it as clean as possible when you’re doing that. You can also add context by creating projects or doing a markdown or a file on your computer. That’s a little bit more into the weeds, but there are some other ways to give it context.

Carrie (16:50): So when you have this big task and you want to map it out, how do you do that? I just recommend that you do it in whatever way makes the most sense for your brain. I like to whiteboard things, or I like to write things on three by five cards or Post-its and move them around. If you like to do that, do it. If you don’t, that’s okay. You can feed in resource docs that you already have that describe the task, the things that you use internally. You can also just dictate your process. I’ll pace around in my office and say, well, then I do this, and then I do this, and all of that stuff. So all of these are ways to just get the task on paper, and then you can map it out.

Carrie (17:45): This is just an example of those skills that I mentioned for The Localist. So I have the researcher, the interview writer, the title writer, and the social media writer, and all of those tasks are separate. But some of them can also talk to each other, and then some of them use the output that another skill made to get their information.

Carrie (18:25): Here’s an example of that. I do a research skill first, so that goes and researches the guest. Well, I’m really picky about my research, because my background is in journalism, so I want to make sure this is actually correct. I give it direction about how it should research, the sites that it should favor, the sites that it should avoid. But then I go and have it create a document of that research. I go into the research document, and I also change and edit in that document. I find things all the time that are either wrong or misconstrued.

Carrie (19:05): For example, I was doing a podcast with Trattoria ZaZa, and the research document told me, hey, somebody on their team just passed away, they post about it a lot, his name is Richard, you might want to bring it up. Well, Richard is an espresso machine. I knew that because I’d looked at the actual source material. So yes, Richard broke, and they did post about it like they lost a team member. But I would have looked really stupid if I got on the podcast and said, I’m so sorry for your loss. So it is important to check the things. That’s also why I don’t have the interview skill going directly from the researcher, because I want to put myself in that process in the middle. I want to review the research, take out the stuff that is wrong or that I don’t want to cover, and then I have the interview skill go back to that document.

Carrie (19:54): I don’t know how much of this you can see, but when you are calling up a skill, or when you want it to do something, you can use the name of the skill if you want to. But also, when AI creates a skill, it will create inside the skill some common questions or some common things that you would say that would trigger it to start the skill. So you may see here, I just said, can you write the interview questions for this guest, and it already knows, because I wrote it into the skill, that that trigger pulls the interview question skill. I don’t have to say go to backslash Localist podcast interview skill or whatever. You can do that, though. But the skill tells it to go to the research document that you created with the research skill. So this is when skills can get really powerful, when they can be chained together, or they can reference things that are created by other skills.

Carrie (20:50): So when you’re breaking that large task into the smaller skills, how do you know where the breaks are? Well, first of all, think about the natural break times. Most of these things I do in different work sessions. I write the title as soon as we record. I write the social media captions after I get everything back from our media department. So those I’m doing at different times anyway. That’s a natural time to break the skill apart. Also, what skills would do better pulling from a completed task? So like the interview skill, I want it to be separate because I want it to pull from the research document that the researcher created, if that makes sense. And then if you don’t know where to break the skills apart, ask AI, where should I break these skills apart, and it will tell you, and it does a pretty good job of that. If you put things on a whiteboard or three by five cards, or you recorded your process, you can just upload all of that and say, I’m trying to decide which skills to break this into, and it will recommend for you.

Carrie (22:05): So you might have noticed that I do have a central skill here called a strategic advisor. I only do this for big projects. The Localist is a big project. I also have some other skills that aren’t super relevant. I knew I was going to be building more skills for the podcast, so when I started, I first built a strategic advisor skill, and I taught that. I uploaded a lot of transcripts for The Localist, and I’ve been doing this since 2019, so it’s like, here’s a lot of historical information about it, here’s what I’m trying to do with this, here’s what I’m not trying to do with this. And that way, when I create those other supplementary skills, I can say, hey, also get your information from the strategic advisor, and I don’t have to explain that over and over. There are also a lot of other ways to do that. You can do that with projects. You can do it with things on your desktop. I did it with a strategic advisor.

Carrie (23:05): I can also go to that strategic advisor skill and ask it for advice. So if I say, hey, I’m thinking of changing something on the podcast, maybe I want to have more experts on the podcast, what do you think? Because I’ve taught that strategic advisor skill how I think through things, it can give advice, if that makes sense.

Carrie (23:39): Okay, so if a skill is just a standard operating procedure, how do you write a good standard operating procedure? This is something that we have found is actually a pain point. Some people write great instructions and some people don’t. So it does help to think through how you write a good SOP. I’m going to walk through each of these, and this will be applicable to people also, but this is really optimized for things that your AI really needs, or that will really benefit you when you’re building a skill with AI. First, define the task. That’s kind of what we talked about, right? Decide what your skill is actually going to do. When you’re choosing a task, it’s good to start with something that happens repeatedly. If you do something twice a year and it’s complex, this is probably not a great skill candidate. It may take you longer to write the skill than just to do it. But if you’re doing it even twice a month, it probably is worth the time. Start at least with something that could be handed off to somebody else anyway. So if you’re thinking, I wish I had an assistant because they would do these things, those are great candidates to start building skills for. And it does help if your task typically gets broken up in predictable ways anyway. So if you’re like, well, I do this at this point, but then if this department gets me this, I go this way, but if they don’t, I do this and then this. You can do that eventually, but I wouldn’t start there. I would start with something that’s predictable. That’s going to be the easiest thing to build into a skill.

Carrie (24:35): Then give it its parameters. Tell it what done looks like. So for my title writing skill, for example, the first thing I have it do is write kind of a long form title for the podcast, and then I tinker with it. I approve it or not or whatever. And then after that, it also writes a title for the YouTube slide, which is different, and a title for the Instagram reel, which is different. So it’s not done until it’s created all three of those. So figure out what does done actually look like here, and then what just belongs in a different skill, and you can build another skill for that.

Carrie (25:25): It also helps to address decision points if you know they’re already going to be decision points. So there may be a point where you’re like, hey, if you get to this point, I want you to stop and ask me what to do. Or it may be, no, I want you to keep going, but just tell me later that this is what you did. So if you know those things happen in a process anyway, or this project is really likely to go off the rails if you do X, Y, Z, tell it those things when you’re defining the task.

Carrie (26:05): And then document your process. You do want to give it the step by step instructions, but in addition to that, also give it the mission of what it is supposed to be doing. We can go back to that example of Emery sending that email to you all and letting you know the date and time of the event. I can tell him, the point of this email is to remind them to come to the event. So maybe the instructions in the email say get five bullet points of what you’re talking about. Well, if he didn’t know the point of the email, he may spend time chasing down those five bullet points and be like, well, I don’t have these, so should I send the email or should I not? And then finally realize, oh, those were not important, because really the point of the email is just to say, hey, remember this is happening, here’s the time and date. So tell your AI that same thing. This is the point of this. That will help with some of that decision point information. Also, if it knows what you’re trying to get out of this, it will know more what to do when it hits a step that is unknown.

Carrie (26:50): Also tell it who your audience is, because this really affects tone and voice and things too. If I’m writing an email to my staff, I’m going to have a really different tone than if I’m emailing a client. And if I emailed them with that client tone, they’re going to be like, why are you mad at me? So there are also different ways you can drill into voice and tone, but telling it who the audience is can help a lot.

Carrie (27:11): When you’re recording the steps of the podcast or the process, you can just write up the steps. If you have an existing SOP, just use that. But you can also do the task and dictate while you’re doing it. When you do that, those things that come up, like, oh, this is so annoying, or this always goes wrong, or I hate when this happens, make sure you’re voicing those things, because that will actually help the AI know how to overcome some issues. You can also record your workflow with AI and it will create steps for you. I believe GPT and Grok do this now. I don’t think the others have it yet. That could be wrong. Probably they launched yesterday. But typically with these services, if somebody has it, everybody will have it soon enough. So if you can’t do it now, you probably can soon.

Carrie (27:55): You also want some meta prompting. So this is asking AI how to make your prompts better. After you give it the task and give it the instructions, ask AI, what am I missing here? What are the gaps? How should I write this better? You can say, I’m going to have you build a skill, write me instructions for building a skill. It’s weird to us to think that way, because usually we’re like, well, just do it, why wouldn’t you just make it better? But it really helps to say, hey, what am I missing, make this a better prompt for you, and then I will feed it back to you. It’s kind of weird, but it really helps a lot.

Carrie (28:45): Also, add examples. This is so important, and we know this is actually important for us as humans too, right? If I went to our creative director Caleb and I said, I want my site to be happier, he would make a site that he thinks is happier, and maybe it would be what I think, and maybe it wouldn’t. But if I said, make me a site that’s happier, and I showed him three examples of sites I thought were happy, he can look at that and think, okay, Carrie thinks that yellow and orange are happy, and Carrie thinks the sans serif font is happy, and this kind of photo is happy. So he’s way more likely to show me a site that looks like what I would consider happy, because he saw examples plus my explanation. The same is true for AI. We think we’re explaining things so well. I do this all the time. I’m like, I nailed it on this explanation, and then for people too, they’re like, what are you talking about? But if I can show them examples of what I’m talking about, they’re way more likely to pair that with my explanation and realize, oh, this is what she means. So I like to give it at least three examples if I can. Something similar is better than nothing. So if I’m creating a tool for sales and we haven’t done this before, I can’t show them this thing that doesn’t exist, but I could show them a one-page PDF that we’ve done in the past and a proposal we’ve done in the past, and it’ll still get some more ideas from that.

Carrie (30:05): One very specific tip is that if you’re creating a skill to write something short form, like a title, I would at least throw one example in there that’s long form writing. When I did my title writing skill, I originally gave it all the titles I’ve ever written for The Localist since 2019, and it still did kind of a meh job. But then when I showed it blogs I had written and a chapter from my book and things like that, and my book’s already been stolen by AI, so I don’t care, take it more, whatever, when I did that, it could see from my long form writing more about my thought pattern and my writing style, and then it translated that into better short form work too. So if you’re looking for examples, you can use screenshots, existing SOPs, screen recordings, templates, finished product. There are all kinds of examples. But if you can give AI some examples of what you’re doing, you’re just going to get a way better product.

Carrie (31:05): And then stress test your skill. Again, just like we’re going to review Emery’s work and iterate on it, we need to do this with AI. So one tip I have is that typically, if you tell Claude to create a skill and you don’t give it something to test on, it will then create something to test on and test on that. I personally just don’t like that. I think it’s a waste of time and tokens, and I don’t feel like it gets a good result. So I would say, when we’re creating the skill, before it goes to create the skill, tell it, don’t run a test, don’t make up a test, here’s something to test on. I think that we get better results that way.

Carrie (31:45): Then you want to audit those results. Are the facts correct? Is the formatting correct? Did it act the way you wanted it to? Did it ask too many questions? Did it stop in the middle? Was it too long, too short? Be really picky here, because you want to iterate on the skill so you aren’t fixing the same things over and over.

Carrie (32:25): One example is when I first started creating that Localist researcher and Localist interview skill, it kept putting a bunch of questions in the interview about how did you navigate COVID. And I’m like, I was podcasting during COVID. I have podcasts from 2019 to 2021 that are all about businesses navigating COVID. I don’t need to cover that. I promise you, they don’t want to talk about it, and the audience probably doesn’t want to talk about it either. So I said, stop giving me a whole interview about COVID, and limit yourself to one question about COVID, if anything. I didn’t just change that in that research document or that interview. I actually went back to AI and said, hey, change the skill so you stop doing this. You don’t have to do that for every little thing, but if you find yourself changing this over and over, just go and tell the skill, hey, stop it.

Carrie (33:10): Okay, so where do you start? I like to start with either something personal, or something annoying, or something that you do constantly. I feel like those are good tests, just to get used to building a skill and iterating on that skill. And if you want more of an audit, so if you have a big task and you’re thinking, how do I break this task down, I have a little quiz. If you go to this, you’re signing up for my newsletter, so thank you. But I have a little file that’s like good things from Carrie, and this audit is in there. It’s from “The AI Daily Brief” podcast. So if you would rather listen to it, you can go find it there. The podcast is also linked in the thing. And if you’re already signed up for my newsletter, you got this yesterday. There should be a link at the bottom where you can find that stuff.

Carrie (34:05): All right, so if you want to deep dive more into this class, this training, I’m going to leave this one up here. This is our half day training. It’s $350 a person right now. That’s a special price through this group. So if you’re interested in signing up, go ahead and do that. If you’re an Infomedia client and you would rather put this on your bill instead of paying for it here, talk to your account rep or me after, and we can work that out. So now I’m going to take questions, if you have any.

Carrie (34:42): The question is, are skills kind of like a form letter or a template? And I would say yes and no. Yes, that is theoretically how they work. But because we’re using them with agentic AI, they can actually go and do things for you. It’s kind of like supercharged. So for example, I write an email after we record The Localist to the team to say, here’s the titles and all that stuff. Well, I have a skill that writes that email, so that part would be similar.

Carrie (35:11): It’s working off of a form, but it’s also pulling in all kinds of information from other documents that skills have created, that I’ve created, and then it goes into my email and creates a draft for me. So it’s all linked up, and it’s formatted, and it has a subject line and things like that. And I always have things drafted. I don’t have it send emails. Don’t do that. So yes, it’s simple instructions. But whereas with a template you have to copy and paste and do that, instructions are something somebody can actually go and do. That is the difference, that the AI can go do something. So yes and no, I think, would be the answer.

Carrie (36:10): David. Yeah, I mean, most of this agentic AI, I may actually throw some of this to Russell, who I do this training with. You can come up if you want to. He knows more about the technical things. Most of this stuff, when you’re using agentic AI, you do want the paid version. You’re going to get way more. Sometimes you have to have it. You can answer this.

Russell Marbut (36:39): Yeah, this stuff changes all the time. Honestly, I haven’t been on the free version for so long, I don’t know what they don’t give you and what they do. If you have an issue writing a skill or anything like that, you’re probably going to need the paid version. The pro version is like 20 bucks a month, I think, so it’s not that big of a stretch. And you’re going to get more tokens, obviously, as well.

Carrie (37:03): And this has come up in our classes. I think typically, if I were trying something new, I would just start with month to month and see, does this even work for you? Because we’ve had some people pay for a year of Claude and then be like, well, I’m on Outlook, so I need to be using Copilot. So the platform you use is going to be dependent on the kind of work that you do. I think for what you do, Claude is great, is perfectly good.

Carrie (37:50): Yeah, I was just being fatalistic. I mentioned that AI already has my book. So the question is really, when you’re feeding this information into AI, how much of it is going into the public domain? And this will depend a lot on the version you’re using. So Claude specifically, if you’re using a Teams account, contractually at least they can’t, they don’t train on that data. So it is true that I was a little wrong in that I do have a paid account, so it wouldn’t be training on that data anyway. Things can still be hacked, so just use your discretion. I wouldn’t put my Social Security number into Claude regardless. And there will be different limitations on the kind of work that you do.

Carrie (39:08): But yeah, the LLMs were initially trained on a lot of training data, and my book is in that training data. So that’s what I meant by that. It is what it is. Maybe AI will start encouraging you to buy local. That’s what the book is about. Did you have anything?

Russell Marbut (39:36): No, just, again, with Claude specifically, they will train on your data if you’re just using a personal account. The teams and enterprise accounts, they claim to not train on your data, but that’s going to be different depending on any of these services you use. So I would just check with them on that and use some discretion on what kind of information you want to give these tools.

Carrie (39:56): I will say personally too, my thoughts on this have kind of shifted as the risk reward changes. Probably six months ago, I wouldn’t have dreamed of giving Claude access to my emails, but now it can do so many things with email. And I don’t send out sensitive data through email anyway. For example, at Infomedia, all of our sensitive data is in different systems. It’s not being sent through email. It is password protected, and we’re using 1Password or LastPass or things like that, so that the client data is really secure. We’re not sending that through an email anyway. So eventually AI became able to do so many things that I was like, okay, if somehow they get hacked and my emails to Pam about stop telling me not to spend over budget are out there, then okay. So some of this depends on what the reward is, and some of it depends on what your company is doing. I would be a lot more careful with client data, all of that. So I hope this means I’ve thoroughly explained AI to you. Please tell everyone to go buy Carrie and Russell’s class, because you will not have any questions after.

Carrie (41:13): All right, thank you. They’re clapping for you. Our next Localist Lab is going to be back here. We’re going to be on the third Thursday of next month, which is, I think, Sept. 17. Laine White is going to be talking about hacking the algorithms. So we’re talking about social media, but we’re also talking about LLMs, so that’s AI. We’re also talking about Google. How do you actually get past these algorithms that feel like they’re changing all the time? Laine spoke a couple of months ago and was a very popular speaker, and people asked me to have her back, so I’m going to. So come to that. It’s going to be great. Also, thank you to our sponsors. Like and subscribe to The Localist, and we’ll bring you this kind of content every single week. Thank you so much.

Carrie (42:07): Thank you so much for listening to this Localist Lab session. Again, if you’d like tickets to the event, check out the show notes. The tickets are free, and we would love to see you there. And as always, whether you’re buying from a local business or running one, remember that what you do makes our community stronger every day.