Webinar: How to Build an AI Skill for Your SEO Workflow

You've probably typed out the same instructions to a chatbot more than once this week: the same context, the same steps, the same explaining-it-from-scratch because starting a fresh conversation felt easier than figuring out how to save what you did last time.

That's the exact itch an AI Skill scratches, and it's a lot less technical than it sounds. This session is built for people who haven't built an AI Skill yet.

Tory Gray is joins Jojo to explain why and how to build one. Expect real SEO examples, a step-by-step process using the LLM to help draft itself, and time to ask your own "but what about..." questions as they come up.

What we cover

• Why this feels like a bigger task than it actually is, and what's really standing between you and building one today

• The repeated-prompt problem a Skill solves

• What counts as a "Skill" across different tools and when something's actually worth turning into one

• Characteristics of a good Skill

• How to build one, using the tool itself to help draft the structure

• How to use your Skill in a real SEO scenario

• Where a Skill's judgement runs out, and what still needs a human to check before you trust what it hands back

Webinar recording

Tory's slide deck.

Webinar transcript

Jojo Furnival: Hi, everyone. We have got a special tutorial webinar today from the fabulous Tory Gray, who is going to break down the barrier to entry for those of you who aren't yet building AI skills and would like to. I'm Jojo Furnival. I'm the marketing manager at Sitebulb.

If you are joining us for the first time, Sitebulb is a website crawling tool that makes technical SEO easier. Guess what? You don't have to trawl through spreadsheets of data to use it. Unless you want to, that is. It's got helpful hints and [unclear: possibly "SEO education"?] literally baked into the tool, and you can find out more about that and register for the free trial at sitebulb.com.

As a reminder, this session is being recorded, and I will email the recording out to you tonight or tomorrow. There should be time for Q&A at the end of today's session. So if you have a question for Tory,

Tory Gray: Yep.

Jojo Furnival: please put it in the Q&A tab next to the chat. It looks like two round speech bubbles. You can also upvote other people's questions in there.

So, now to our guest. Tory, as founder and CEO of Gray Dot Co, a senior-level SEO consultancy specializing in digital strategy and tech solutions. She's a familiar face and mentor in the Women in Tech SEO community. And is also one of the trainers on Sitebulb's free JavaScript SEO training course. If you haven't taken that one already, definitely check it out, especially in this AI bot crawling world. Please give her a warm welcome.

Tory Gray: Mhmm.

Jojo Furnival: You are in fantastic hands. So I'm gonna hand over to Tory now, and I will see y'all later.

Tory Gray: Thanks, Jojo. Alright, everybody. Hello, and welcome, and thank you for coming out. And let's, you know, get into it.

Here we go. Okay. So first, I would like to address the elephant in the room. First, I am a millennial. So there will be memes. I am so sorry, but I promise to keep all the cringey emojis to a minimum.

But more seriously, I'm guessing most of you are here because you needed to. You were told that you had to, perhaps a boss or a client, or maybe it's just social pressure from the industry as a whole. Maybe you're here and you're not quite sure what a skill is or what it does or how it fits into the AI process tooling in general. Or maybe you do, but like most of us, you've been learning on the fly, and you don't really, maybe you just need, like, a foundational review, ensure your bases are covered, and so you don't have to be quite so anxious about, like, not knowing what you don't know. Or maybe you want inspiration for how to apply this existing knowledge into specifically AI processes. No matter what, we're here for you. Today's whole point is simply to get you out of your own way and get you using new AI tools, when appropriate, more comfortably. That is, if you can get past the big bad.

So the biggest, hardest barrier here for most folks is not the actual execution. It's the emotional one. It's the fear. And I think for a lot of us, it isn't a big scary fear so much as a simple, wow. I need time and space and, like, emotional energy to wrap my head around yet another new thing. And, realistically, when the heck will any of us ever have time for any of that? But, again, that's why we're here today, to shortcut that emotional anxiety or fear. And I can promise you very honestly that skills, in some of these AI processes today, we're talking foundational stuff. And it's definitely a really good example of one of those things that, like, feels emotionally like, ugh. I can't handle that right now. But as soon as you do it, it's much more simple than you think.

So let's get into it. First, we're gonna talk about what skills are and all of the AI process building blocks that you have in your AI pool. So, a skill in general is a standard operating procedure of the AI prompting world. It's an SOP for the modern age. It is a narrow, repeatable instruction set for one repeatable job.

And for the most part, the AI industry has decided skills as the name for this concept. You may hear these referred to as custom GPTs or gems. Those aren't exactly correct names. Those do analogous but also somewhat different things. But it's a Wild West of a world right now, and people are using names accurately and inaccurately, and we never know what they mean. So it's definitely one of those cases where I think it pays forward for all of us to ask the dumb question and say, like, hey. When you say GPT, when you say gem, do you mean skill? And make sure you're speaking the same language, essentially, just to check. In case.

What does a skill look like? Functionally, it is nothing more than a text file. It's saved in a markdown format, and it lives within your AI tool of choice, whether it's Claude or ChatGPT, or Gemini or what have you.

Much like a file even in your Drive. Right? You name it whatever custom name you want to, and that text is going to outline the written instructions for a narrow process. And like any SOP, it should be structured in a logical, sequential, step-by-step order. It can, when needed, refer to supporting files for additional instructions, much like a web page would link to another or a Google Doc would link to another supporting doc. So for example, a skill could have a step that has the AI refer to your brand guidelines in order to follow, you know, those voice and tone rules. There might be a separate file in support of that for things like, say, banned words or grammar conventions like, do you use an Oxford comma or not, or do you wanna ban the use of an em dash or always use an em dash?

More strategically, the process outlines the steps that the AI tool itself should take in order to finalize a task.

That can be completely autonomously, whether multi-step or just a one-shot. Hey. I tell you to do this. Go forth and do this. It can also be a very stepped back-and-forth process between yourself, or maybe another member of your team, and the AI tool. And that's what we call human in the loop, and that's generally what I prefer to do. Because I'm picky as heck. You can stack these skills one after another. They can be steps in a sequence. You can have one trigger another. You can also have a skill double-check the work completed by the first skill.

In Claude, you'll find the skills in your settings and then in the Customize > Skills menu, and you'll see what skills are already in place there, or you can add and edit new ones. In ChatGPT, you can go to the Plugins on the left nav, and then click on the Skills tab. I do have screenshots here, obviously, but keep in mind, these change, like, every three seconds, and I think they might have changed actually since I took these screenshots last week. So it's just to help you generally look around for where it is, but you know, buyer beware a little bit there.

Okay. Now that we know what skills are, let's talk about what the heck projects and MCPs are.

Projects are like a folder of information about a specific subject. And it gives context to variables. And it's not just a folder, it's memory across conversations. So it has the context of those conversations, of those files. That way, you don't have to provide it in a single prompt. So rather than, like, a mega long, crazy prompt that has all the details all the time that you're copying and pasting, a skill will have that handled for you, and a project can give it context that is specific to a client, which is very nice when you have multiple clients and you want different context but, say, the same skills.

And, again, client details are a great example here, stuff like brand guidelines. Tone, voice, do's, don'ts. Whatever is specific to that client that you want it to remember, and, again, you can run a general skill or a process generally in your account, or you can run it specifically in project A or project B against different clients.

Versus an MCP. So it is a tool. It is a connector that brings in outside data. If it is a Google Analytics MCP, for example, it would bring in GA data. If it's the Sitebulb MCP, it'll bring in the crawl and the audit data. Note that as a workaround, if you don't have access to the MCP or you haven't set it up or for whatever reason, you can always pull individual files into each of the chats.

I would flag that for the purposes of this presentation, I'm going to say MCP and connector a little bit interchangeably. They're not quite that, but frankly, for our purposes here, they are. So let's not be, you know, too on top of ourselves. But working together, these tools can really help you quietly unlock a lot of value from your existing AI tool subscription. Yes. I'm using danger quotes. Sorry. Lots of AI words right there. Quietly.

Okay. The last big kind of tooling option around the process stuff that I wanna mention is Cowork, or it's called Work in ChatGPT. We'll mostly call it Cowork because Work feels just like the word skill. It's just so general. What even am I talking about? But what a Cowork or a Work is, is it enables automation. Right? So a one-off chat has access to a specific file that it fetches on command. But in Cowork, it's persistent access to, say, the files on your computer. Or in your Drive, or in your email. That way you can automate tasks, and they can run on their own, scheduled ongoing stuff. So when you're talking about actual, like, automation without you having to trigger it manually, you're gonna wanna work with a Cowork.

Projects, again, are like folders a little bit functionally in your system. It's a place you'll navigate to within your tool. You can stick files, reference resources, in this kind of folder section that's at the bottom right side. Existing older conversations will go in that bottom section, and you'll start new chats at the top of the page within the project.

In order to find them in both tools, they live in the left sidebar. But in ChatGPT, you'll have individual projects listed there right now. Whereas in Claude, you just click Projects, and then you click on the project within the main nav, and then you go see it.

Now if you're like me and you've already started conversations asking about client prompts, perhaps even in, like, the sales process, if you're agency side, you're gonna have conversations about that client that aren't within the project. But don't stress. You can always click that little hamburger menu, the little three dots, and you can add any conversation to any project retroactively when that's the case.

You can find MCPs in Claude in Customize, and then Connectors, or in Plugins within ChatGPT. If you have a Claude or a ChatGPT team account, the owner of the account is gonna have to add that connection for you. But once it's installed and added in your account, you can tell the AI to use it in any particular conversation, like toggle it off and on by clicking the little plus in the bottom left corner of the chat. And then there's, like, a little, you know, slider sort of button that you can toggle off and on.

Okay. Let's shift a little bit more now into thinking about when and where we would use skills. A little bit later in the presentation, I'm gonna be talking you through specific skill ideas to get started. But in order for you to just brainstorm beyond then, there's a few places you could start. So you can do the classic thing and brainstorm yourself or with your team, perhaps. Talk to your team. You can also look to existing SOPs that your agency or in-house environment might already have. And you could see if any whole or parts of those SOPs could be tested to be taken over within AI. So here's a great use case where you can actually use your AI tool to help you get started. So you can upload your SOPs, for example, and say, hey. What part can you take over? You can also ask it, hey. Based on our conversation history, based on my brand, based on my industry, based on my website, what sorts of skills should I be thinking about making? So AI is a great sounding board here for generating these ideas when you want to move on that.

Okay. So now that we're thinking about what skills to make, I actually want us to pull back pretty just briefly and ask a very critical question, which is: is AI in general even the right tool for this job? If you're trying to check redirects, for example, are they 301s? Are they routing to the correct destinations? An LLM is not your tool. I hear too many people talking about using AI to, like, hey. Let's go just check our 301s. Like, that's not what the LLM does directly. If you wanna check that kind of stuff, you know, use a tool, maybe like Sitebulb. And then you can maybe build the skill to do it all in one.

Or if you need, say, a reminder to take a pill, you don't need to over-engineer it and talk to AI to set up a process to send you a ping and a reminder. Right? You can just use a phone alarm. You can just use a calendar reminder. I am quoting Britney Muller here, and that's where these images are from, when I say you don't need a blowtorch to light a birthday candle.

If instead you have a lot of messy, unstructured data, like client data in bulk, an audit output, a 150-page deck that you somehow need to, like, ingest and understand and speak intelligently about in your meeting that happens in ten minutes. Maybe AI can help.

But for more information, Britney has a really great LinkedIn post with a framework on how to kind of figure out, should I be using AI for this? Should I be using AI to build a tool in order to accomplish this? Should I be using AI with a skill connected to an MCP like Sitebulb in order to check this? Or do I just not use AI at all? It's a great resource to do that. This Bitly link will bring you to that social media post, or you can go to community.orangelabs.co to learn more.

Once you've determined that, no, no, no, this is a thing I actually do wanna use AI for. You need to determine the shape of the thing, again, to put it in an AI way. Essentially, all this means is what tool or what structure across those kind of four tool sets — the skills, the MCPs, the projects, and the Cowork — that you wanna use? Where do you wanna start?

The first two questions to ask yourself that matter the most are frequency and complexity. If it's a task you're gonna do once or twice, don't bother. Right? Don't set up a bunch of process. You just need to prompt in a one-off conversation and then do it again later. That's okay if you're only doing it once or twice.

If it is something, however, that you do repeatedly, especially with AI, then suddenly a skill makes sense because then you don't have to copy and paste that information every time or rewrite it every single time. If you need more specific, deeper context for that query, then you're gonna wanna think about using a project. If you wanna pull in outside data and import those files directly, then you'll leverage an MCP or the connector. If you need automation when you're ready for that step, use Cowork.

Okay. For the processes you wanna run, one other kind of more important question is: is this one skill or is this multiple skills? Don't get too hung up, but generally speaking, the more complex the ask, the more likely it is that you'll want to use one to two skills. Or two-plus skills, that is. If you aren't sure — I mean, frankly, that's okay. Sometimes the answer is an edge case. It's another case where just literally asking your AI tool, like, hey. What the heck do I do? Should this be one skill or two? It will tell you an answer that it thinks is correct. The answer is not always perfect, but sometimes it's just enough to get you out of your own way to get started, because learning is doing. And remember that this is pretty low stakes. It is reversible, and you can pretty easily change that later. Just because you start with one, you can break it into two later with the help of your AI tool. So get out of your own way by just doing something.

And like any process, remember that it's going to be improved over time. Inherently, so start somewhere.

Now once again, if you're like me, you might be questioning this advice, maybe for even good reason. I think many of us have heard the complaints on LinkedIn about, you know, SEO and GEO influencers calling for people to stop asking the LLM how the LLM works, which is frankly accurate in some ways and not in other ways. There's essentially a critical difference that I'm speaking to here.

AI tools are built to help you use them for maximum impact. They are quite literally trained on how to use their own public features and tool sets. They want you to use them effectively. AI tools don't want you to game their algorithms in order to make your brand show up more frequently. So that LLM that you're talking to in your conversations in your tool isn't inherently trained on even the things about how their algorithm works. So you know, they're borrowing that insight, and they're just quoting what other people on the internet say.

So spoiler alert, you should not just trust everything that everyone says on the internet. But I think when you're talking about how the tool works, in order to use it more effectively, because that is publicly available data that these tools are trained on. It's safe to ask them those kinds of questions. So if you're just like, hey. What can I ask the LLM about the LLM? Is it public data? Is it private data? The answers will be dramatically different about, you know, the results of what you'll see.

Okay. Let's talk a little bit about what makes a skill good. Now improvement, or dare I say optimization, of skills is totally normal, just like any normal process improvement is normal. But generally speaking, we like skills that meet your needs. Right? Whether that's your client or your boss, or your own nitpicky requirements. You want a skill that you can trust and that you're proud to show off to your team, a coworker, someone externally. Or maybe it's just good enough that it helps you draft something and get you 30% of the way there or 80% of the way there, but saved you time and it moved you forward and it got you out of your own way. I think either of those scenarios is really great.

Okay. So enough with general advice. We're gonna go through some very specific SEO skill examples as well as kind of the structure so you can see how they work together. These are things you can start making potentially this afternoon.

So first up, here is a very simple skill that helps you draft options for quotes. It could be used as a response, say, to a journalist inquiry. You know, when time is of the essence and you need to draft something very, very quickly, AI is good at being quick. So in this example of this structure, I have client context within their project. So we have stuff like which spokespeople for that brand are available for what topics. The brand values, voice and tone, do's and don'ts that are very specific to that brand.

And then within the prompt, you're going to provide the context of the inquiry, like the journalist, their publication, what the subject matter is. You could have that skill vet for fit. When you're moving really quick and you think something based on a keyword [unclear: possibly "maybe"?] alignment is a good fit, but maybe in that case, it doesn't make sense to actually provide a quote. You could have it first vet again for that fit. And then as a second step, it could potentially draft a quote or maybe multiple quotes from different angles — pros, cons, leaning into it, expanding it. Like, there's different ways you can quote, and you can have the skill provide you however many different variations accomplishing different needs that you want.

So maybe those are things you can use, and they're good enough for you to just send to the client and get approval or send right to the journalist and get going. Maybe they help you get your ideas turning. Maybe you look at it and be like, oh my god. I hate it. But now that makes me think of this, and then I move fast because I can deliver something. But in any case, hopefully, it helps you move more quick and respond to that inquiry again, more quickly — that increases your chances of getting it. And getting your client cited.

Okay. So here we have a skill for content refreshes. Let's say it's an older post that we all have. It's no longer performing as well. And you wanna know mainly what the competition is doing that your post is not doing. You have a project, again, that has client context. Maybe this time, it has their priority topics, their landing pages, their goals.

Here, we're suggesting the SEO Gets connector for access to GSC data so that you can use this skill to ask about, say, a specific URL. It'll maybe look up what keywords have dropped in rankings that used to drive meaningful impressions and clicks, but maybe aren't anymore. And then it can look at the actual search results to see who's ranking instead, and then what do they have on their pages. So what topics, angles, you know, cited studies? What do they have on their page that you don't necessarily already? So quick competitive research. Right?

I will call out that this is specifically a good example of something that's very likely a cluster of skills versus an individual single skill. Maybe one finds your blogs with performance drops. Maybe one looks to see what is the reason for that drop. Is it because something used to rank and isn't anymore, and therefore we have less clicks? Or is it a less controllable factor, like maybe an AI Overview where your ranking stayed the same? And maybe a different skill that reviews the SERPs and evaluates the competitive difference for those keywords that are pertinent for this query.

Just a quick callout for those of you especially in the EU and more globally. I think everyone's all a chatter about the new EU AI transparency rules about word mark— you know, con— marking the content, water— you know, god, what even is that word? Watermarking. For that content. But here I wanna point out, like, these tools aren't just good for copywriting, and maybe you don't even use it for that. Maybe you use it to help you do the competitive research so you can understand what's available out there, so you can decide what's right, and how you can differentiate to make your content that much better than what is already out there. It doesn't have to do the copywriting. Maybe it can do different parts that still help improve your process.

Okay. Here's another one. In this case, we're using Claude, and we have weekly or monthly crawls set up. You could make a skill to compare the crawls and then to bucket those hints maybe into meaningful groupings that help you communicate site changes quickly. So if your site experiences a lot of bugs, and you're in that situation where things keep breaking and therefore you're not growing your SEO results because you're busy chasing after things that broke in order to fix them, to regain the traffic, this sort of very simple summary that, you know, helps communicate to your team — hey, the site went backward, the site moved forward, we have stayed static, and we actually haven't deployed any changes to make any meaningful progress against what we're doing. These are all a great, very, very simple way to show you the frequency and the scale of whatever those issues are. So it's a very simple, but maybe more helpful than you think, way to communicate these.

Okay. In this skill, we're combining multiple MCP sources in order to analyze insights across sources simultaneously. So here, again, we're using the Sitebulb MCP as well as the SEO Gets one in order to get the GSC data. And here, we're trying to help investigate the source of a drop in traffic, map it to the problematic URLs maybe driving that loss, and then checking via the Sitebulb data to say, hey. Maybe there's been a change on the site. Maybe we accidentally, you know, broke the canonical on these pages or no-indexed these pages. These sorts of checks can get very streamlined and very quick when you're using multiple sources in order to do this.

Okay. Last one, for now. This time, I wanna pit one skill against another in order to verify its work. I used this example very specifically because I think we've all heard that horror story that was on Reddit about a report, about, you know, it had actual data from analytics and the QA or the AI tool was creating data and creating the report the company was basing itself off of. And several months later, after many decisions made based on this data, they realized the AI just didn't even bother to fetch the data. That is a problem. Right?

And so we can QA that manually every week, every month, every whatever the frequency is. But we can also potentially reduce the rate of errors by pitting one skill against another where this QA skill is going to look at the work already completed by that original skill and QA the work again. So are these metrics actually accurate? Is the analysis that it created based on that more accurate — again, to ensure that real data was used instead of hallucinated? So here's a way we can use AI to make AI better and reduce the likelihood that you're making tragic, horrible mistakes with your, you know, marketing decisions based on bad data.

Okay. Now that we have some example structures, we're gonna get into the logistics of building the thing.

In theory, making a skill is a pretty straightforward creation process. You write it, you save it, you test it, you refine it. But in reality, these steps aren't always quite so nicely ordered. And that's okay. But there's a lot more loop-de-looping around that happens. You're probably going to test your skill live within the conversation that you're making that skill in. There are likely multiple revision rounds. You might wanna bring in coworkers, for example, to help review the skill that you built. Review those steps for gaps, review the results and the output, and make sure that everyone likes it, because different people have different eyes and will bring different value to that process. Just like in the engineering world where pair programming is a thing, chances are your process is improved by having more than one person review it.

You can and should ask your AI tool for the issues that it finds during those revision rounds. So just like you might talk to a human to get feedback on it, Claude or ChatGPT can be a great place to start as a sounding board, filling in gaps you didn't notice. I generally find Claude to be better at this because ChatGPT tends to say yes to whatever it is that you say, and Claude tends to be somewhat more argumentative, which is somewhat annoying when you have to go six conversation threads deep to prove it wrong. But that's neither here nor there. But here are three different examples of ways that, like, we at Gray Dot Co have used Claude directly to make our skills better.

We ask it literally what it needs. Hey. I'm making a skill. What do you need in order to do this? That's a valid question to ask. You can also ask it once you have stuff filled out — hey. What steps am I missing? Where are my strategic blind spots? It's gonna give you a bunch of answers. You don't have to agree with them all. You don't have to implement them all, but it's good to have, again, a sounding board because then you can react to that and decide what you wanna put forward.

You can ask it to actually draft the skill. I don't think I've ever made a skill from scratch. I'm always working with the AI to write it for me. And, you know, get all the details. And then I get that, and then I can review and read through that and review the output and make sure it meets my own specs.

When you go to actually start building the thing, there's gonna be many kind of critical fields that you should consider supplying in order to make a good skill. You're gonna cover the obvious stuff. Right? The purpose or the goal of that skill, if there's a target audience for who's gonna receive, say, the deliverable of whatever thing you're creating with that skill. You can tell it what you'll provide each and every time in order for it to move. And do its thing.

What it should do if you don't provide it that information? Or what if that information is contradictory? Do you want it to be skipping steps? Do you wanna let it allow you to do that? You can tell it those instructions. What specific steps should be taken as a part of that process? When a human-in-the-loop step — right, if you want, like, a, hey, once you draft this, I want to review that output before you then, you know, draft a blog post, for example. You'll wanna include details on what the output format should be. So should it be a blog post? Is it long-form written text content? Is it a PDF? Is it a client deliverable on nice letterhead? Is it a spreadsheet? Right?

I would also very much link to example work that meets your specs. So if you've already done this process many times and you already know it works and you have example deliverables across clients, import them — the more examples, the better, frankly. You know, seven, eight plus. Then it has lots to work with, to compare against, to say this is what makes it successful, and I think it'll improve your results.

You can also set instructions for how to call or trigger that skill to kick off. The very simplest answer to this — funny, because people tend to, like, Google this and they're, like, scared to ask, but the simple answer is you just tell Claude or ChatGPT to do it. Hey. Go run this skill. I want you to go run X on Y information. The name doesn't need to be perfectly exact. It figures it out. If it's ambiguous, like, if you have, say, two skills with similar names because you've gotten good at this and you're building more and more, it will ask you to disambiguate. So it's pretty easy to do.

You can also, within the skill, build instructions for when the skill should move again without you. So if one skill finishing should trigger another skill to start, then you can build that step into the actual steps of the skill. But, again, this is another good case of, like, don't overthink it, and if there are dumb questions — I don't think there are dumb questions — but you can Google it or you can ask, in this case, how you call that skill.

If you want to see actual real-life skills that are in the wild that people have spent lots of time and thought and effort into building, good news: Sitebulb has a few available. They're on their GitHub account. The Bitly link that I've included here is the one they recommend starting with, the "what to fix first." So this works in conjunction with the MCP, but you could also import audit data and say, get started. So literally read through those skills and see what elements they included and think about why they did it and how that's going to improve the results. And even think about how would you tailor that to your use case and make it even better specifically for your clients.

Okay. Finally, we're gonna end with some tips and considerations and, really honestly, things to watch out for as you build these skills.

The first critical one is the obvious. Make sure it's doing what you think it's doing. Check your facts and sources, because a one-off issue in a single conversation is significantly easier to fix, versus a skill can do something wrong 50 times in a row, because it seems to be working, but nobody bothered to check. So I'm very much encouraging you to check it. Nobody wants SEO or AI slop. So take the time that you've saved in automating this work and at least initially put it towards QA-ing and ensuring that it builds what you want, and it's going to add value and make your clients or your boss happy.

Just remember it's AI. Mistakes happen. Hallucinations happen. Think of it like a junior SEO or an intern. Like, having a layer of checking, judgment, tone, relationship history, context — all that stuff matters. And robots, just like any of us, are not perfect, and they're going to make mistakes. So stop depending on it to be magical and build QA ongoing into your process.

Skills will evolve over time. Hopefully, they're gonna be getting better. But sometimes that evolution means that stuff accidentally gets overwritten or erased, or lost on accident. You'll want to keep them up to date, and you wanna minimize the chances of that erasure happening. Right? And you can do that by following these steps.

You'll wanna QA the output on some regular basis, maybe considering building skills that help you do that more frequently.

I think it's really important to define the owner of a skill. If you're a single freelancer, that's gonna be obvious. Right? But if you work at an agency, if you work in-house, what if the person who created that skill leaves? They go on maternity leave or they get a great new job. Who owns that now? What happens if it breaks? Who's responsible for going and resolving that? Like, literally just declaring an owner of that skill goes a long way in ensuring that things are streamlined.

You can also keep a change log of what changes were made to that skill by whom for what purpose, and your AI tool will help you do this if you ask. So within the skill, one of the steps can be like, hey. As this skill evolves over time, let's keep a change log, and I want you to annotate all these details so that we have a marked history of what was added at what point. And as a complete failsafe, you can always export the text version of the skill steps, and you can store it somewhere as a backup. Right? Somewhere maybe the AI doesn't have access to. So you always have a version. If it gets overwritten, you can bring it back quickly.

Okay. So this is the biggest one. A quick word on privacy and security. Because let's face it, a connector or an MCP to live data is more powerful. And using a skill off the shelf that's publicly available and prebuilt for you is so much faster. And frankly, there's such huge pressure to move on all of this stuff now, now, now, now, now that I think too many of us are forgetting some of our security and privacy basics. There are risks here, but there's also simple questions you can ask yourself and steps you can take in order to limit these risks.

So the first big one is just: who made the thing? And maybe you don't wanna trust internet randos because yeah, it's the internet. People can lie. People can do mean and malicious things. So for example, there is no official MCP for Google Search Console. There are a few different community versions available. You can also use the SEO Gets one. But for any of the, you know, multiple community versions — who built that, what does it do, and what does it connect to? These are all really important questions, as well as: do you trust them? Do you trust them to be trustworthy, and do you trust them to know what they're doing and to understand what risks they're taking and telling you to go forth and do, and connecting you with, that you might not realize? Official MCPs and official skills can reduce the risk because those companies are literally incentivized to reduce the risk to ensure that it works well for you, because their reputation is on the line versus Joe Schmo on the internet.

This next one, what it's connected to, is highly related, because it measures how big the impact could be based on literally what it's touching. So the Sitebulb MCP, for example, it brings your crawl data into your AI tool. Depending on your brand, your site may be publicly crawlable. It might just be available. So porting in that data via the MCP into your instance of Claude or ChatGPT is pretty high reward and pretty low risk. That's not necessarily private information. Right? On the other hand, a public skill, a prompt, or an MCP that you don't know or you don't trust could include, say, a prompt injection attack.

Or it could link to a third-party codebase, to a GitHub account, that maybe right now is hosted by someone you trust, but maybe they don't maintain that over time. That is a thing on GitHub. What if someone takes over that account and suddenly someone you don't know and you don't trust has access or permissions to your data, your gateway through your connector here, or to your computer in general. So know what you're connecting to and what the implications are.

The next really big one — but, honestly, this is a big one, and I see people being too casual about this — is: do you have permission to share this data on behalf of this client? I don't at Gray Dot Co. Not for all of my clients. For some, I do and for some, I don't. And your agency or your company might not either. And generally speaking, don't share data that you don't have permission to share.

Why might you not have that data? So maybe you literally don't have property access. I have clients that don't want anyone outside the company, based on their IT rules. Like, I literally can't even get GSC access, for example. They have to export reports and send them to me because they're very privacy-minded, and they're very closed and locked down.

Maybe the tool that your company or agency uses isn't approved for use by — I'm sorry — by the IT team. So I have had cases where clients maybe are in the Microsoft ecosystem and Copilot is approved for use, but I'm using Claude. Claude is not approved for use by their AI team. So it could be about the tool itself versus the client's data. Or a privacy-minded client in general just doesn't want you specifically sharing their proprietary data with any AI system because they want to keep that on lockdown. So understand what you're allowed to share, when, and maybe don't just go and add a GSC connector to all your clients if that gives access to some clients who don't want you to do that. Yeah. Don't share access where you don't have permissions to do it.

Okay. So the next big one is to think about what could go wrong. And how quickly — like, can you fix it, and how quickly can you fix it? So the Google Drive connector to me is much more high risk. Because you're giving it access to potentially proprietary client data. Could that information be hacked? Could it be not hacked, but instead just deleted or accidentally overwritten? I mean, if you're like me, you've heard stories about accidentally erasing somebody's entire database of data when they were trying to do an update, or it deleted all of the files on their computer. Like, what would you do if that happened? And how can you set up a structure and permissions to ensure that it has access to the things that you want it to have access to, but it doesn't have access to the things you don't want it to, or that you have things like backups. Security protocols. Think through these scenarios in order to limit your risk.

Because each connection out promises — it promises the world in terms of saved time and effort. But all of those things come with risk. So be thoughtful and think about how to reduce or eliminate those risks in order to protect yourself and your brand or your clients. But mostly, overall, just, like, breathe and go try some stuff and try and go a little bit more slowly than everyone's pushing you to do.

And also, I want to call out that Jojo very nicely has allowed us access to their MCP, which is still in beta, for everyone that is joining us today. So this Bitly link will bring you to that link so you can set that up for yourself if you have a Sitebulb account. And I would also encourage you to check out the skills in conjunction with that. So it can be great to start with these things off the shelf, again, from a trusted company that you trust that they're giving you good information, and that's a great starting point for experimentation.

Thanks again for joining, and hopefully this has helped you get kind of more unblocked, started on making skills and AI processes for your business, for SEO. But now I think we're done, and we can switch to some questions if there are any.

Jojo Furnival: Hey. Thank you so much, Tory. That was very, very cool. Thank you. So at the start, I was quite curious to see out of the audience who had already been building a skill and who hadn't. And we're actually roughly fifty-fifty.

Tory Gray: Hi, Jojo. Mhmm. Okay. That's a good mix. I like it.

Jojo Furnival: Yeah. So some people who have definitely had a go maybe not feeling super confident or want to sense-check their process — or, you know what I mean? So that's quite interesting. I'm going to close that poll now, and we can move on to the questions because we do have some questions in. Here we go. So the first one is: is a skill connected to a certain model? Or does it change with the model? How do I make sure that the skill still works when a new model is launched?

Tory Gray: Great question. In short answer, it's gonna use — you know, like in a chat where you can toggle to a new model, and you can, at least in my Claude instance, you can tell it I wanna use, you know, a more expensive but more in-depth model. So as you call that skill, you can just toggle it when you start that conversation just like you would in any new chat. If it's automatically triggered, then it's gonna use whatever you defaulted to or whatever you used the last time you used it. So it will automatically update for you when you change that model, when that becomes available, and you've toggled to use that.

Jojo Furnival: Yeah. And just to add to that, I think we have experienced internally that if you create a skill with one model, and then it gets used by a different model at another time, it doesn't necessarily behave exactly as you're expecting. So something to keep in mind and another good reason to — every time there is a new model that comes out or there's a model update, make sure that you are evaluating your skills to ensure that they're still fit for purpose for whatever model you want to use with it.

Tory Gray: Yeah. That's a great callout because models are, you know, better and worse at different things. So you might need to give it extra instructions in order to avoid something that annoys you, or breaks your specs.

Jojo Furnival: Okay. And another one is: have you used slash created gems in the same way as skills? On the basis that they're not quite the same? And if so, what differences did you notice in terms of setup, implementation, and results?

Tory Gray: I haven't played too much with gems. I live in the Claude ecosystem right now. Gems are a little bit — they add in some agentic capabilities, which are super fun. So they're on my list to do, but I think the biggest change is — I assume, that I can't speak to, would be mostly about what it can do based on agentic capabilities. Right? Can bots go and crawl the web and do some other things a little more autonomously? Jojo, can you speak more to that?

Jojo Furnival: Yeah. No. Well, similarly to you, I live in the worlds of Claude and also ChatGPT. I haven't experimented with gems yet. However, quick shameless plug. So I've started this podcast called Search Bar Crawl, and I recorded an episode — go look it up on Spotify and YouTube — I recorded an episode recently with a guest who was speaking about gems, actually. And in their view and in the evidence that they've seen, they would describe gems as a bit more creative. Certainly in process and output, certainly in comparison to — like, maybe Claude is just a bit more technical, perhaps a bit, I don't know, drier.

Tory Gray: Mhmm.

Jojo Furnival: And so as a result of that, they use gems for content creation processes. And, you know, how they differ, how they compare to ChatGPT on that — not sure, because this person just was like, I personally hate ChatGPT, and I'm not gonna use it. So I think the bottom line is, if you can, try it yourself.

Tory Gray: Mhmm.

Jojo Furnival: Yeah. We don't have the answer to that, but I'm sorry. Yeah.

Tory Gray: Yeah. Liz chimed in in the chat and said they felt that they were more limited than skills. So that's an interesting takeaway because in theory, if it can do agentic things, it should be broader. But you know, things are evolving.

Jojo Furnival: Thank you. Thanks for that, Liz. But okay. So if there are no more questions — well, actually, maybe I have some questions. So I'm kind of interested in, like, in an agency because it's very, very different. Obviously, I'm in-house.

Tory Gray: Mhmm.

Jojo Furnival: You're agency. So who do you think should own skills? Is that, like, in an agency? I mean, is that kind of an ops thing? Is that the team leader's job? Or should every SEO be building their own skills and trying stuff out for what their needs are?

Tory Gray: I think it's valuable for everyone to experiment in making skills because, again, learning is doing, and you want to see does it improve. I think everyone will also benefit from reading any of the existing skills. So I used your guys' as an example. But if you're a junior SEO in an agency, I think reading through what you're already doing, what are the rules, what are the considerations, how do you QA — like, those are helpful learning mechanisms.

In terms of ownership, though, I think there needs to be sort of a subject matter expert. You could have a different person on QA. And I think if you're a big agency, it's gonna be more multifaceted. Right? You can have someone who's more responsible for QA-ing the delivery output, versus someone who's in charge of maintaining the skill, versus someone — to your point, Jojo — who can look at the model. And if there's a new model update, like, let's actually dig in deep and see if it still accomplishes goals. I'd say it's often like the director or it's someone that they've appointed to own that, that's like, the innovation person on the team, that's, you know, well suited to doing that.

I also just wanna reemphasize here, like, how much value I've found in that sort of pair programming kind of aspect. Like, I'm a pretty senior SEO, and I'm very picky, and I can human. And when Anne and I, a member of my team, go forth and make skills together — just chef's kiss. The things that we can make together, because our brains work in very different ways and consider different aspects and can improve the output of the things that we make. I just think that's really meaningful.

Jojo Furnival: Okay. I think we've just had another question. Which does not seem to want to display on my screen, but that's fun. Basically, it's: we have an AI group across all channels in IT so we can evaluate things — oh, that's not a question. Sorry. I read, should this be a skill?

Tory Gray: It's a good structure suggestion for how to do this, for sure. Because I think you're right. Sometimes it's an app. Sometimes it's a skill. Sometimes it's a skill within a project. Sometimes something moves from a classic skill, and then you wanna automate it because it's really working, and, you know, getting it in Cowork and all that kind of jazz. It evolves.

Jojo Furnival: Yeah. So we've got a fifty-fifty here audience in terms of who have already tried building a skill and who hasn't. Someone in this room is gonna open their laptop tomorrow morning and try building a skill. What is the one skill that you would tell them to start with?

Tory Gray: I think the easy one to start with is just, like, fix the things that bug you. Create brand guidelines. Create tone and voice. None of these things are perfect, but the rate at which — like, if I don't want em dashes, stop showing me em dashes. Stop showing me double dashes — that's a thing it really tries to do instead of an em dash when I won't let it use an em dash. Like, build your brand guideline, tone, voice, do's and don'ts in terms of the language. And see if you can improve the rate at which things sound more like you.

It is, I think, an easy place to start to say, or, like, a social media posting skill. Right? Like, I have an idea. Help me copywrite and craft, in a more smooth language, what this is gonna be, and then look at the output and see, does it meet your specs? Does it sound like you? Because usually what it outputs by default is very masculine. Very kind of tech-broy. And if you don't want — I'm not a tech bro. I'm not a bro at all, so I don't want it to sound like that. I want it to sound like me, and that's just a very easy place to start to say, hey. Do this. Don't do this. Always do this. You know, having it do an Oxford comma all the time because I love an Oxford comma. I know that's one less thing I have to edit.

Jojo Furnival: Yeah. And from my perspective, like, I think, you know, that I have to juggle a lot of things. We all have to juggle a lot in our day-to-day work. And some of the stuff that irritates me that I have to do is, for example, we have, on Fridays, we have a show and tell in our team, in our company. And I can't remember what I did that week. I have — like, literally every night when I go to sleep, my entire memory resets. And then I'm like a baby, and then I wake up the next day.

So I have built a skill, and it took no time at all, to, you know, run automatically on a Friday morning to tell me what was cool, what was interesting that I did that week that I can share with the team. For example. So just as, like, another little example that you could maybe think about what you do, yeah, in your day-to-day that — ugh, bugs you that you have to do it.

Tory Gray: That's sexy. I like that.

Jojo Furnival: Yeah. Okay. Thank you so much. Thanks, everyone, for joining. Thank you, Tory, for your brain and your time and your insights. That's it for today. Our next webinar is gonna be a panel discussion about the power of MCP in your tool stack. Keep your eyes peeled for that. And, yes, like I say, I'm going to send out the slides and the recording via email tomorrow, and you will get all of those delicious links and things and resources so that you can build your own skills and try out the Sitebulb skills. Thanks again, guys, and bye.

Tory Gray: Yep. That's a good one. Thanks for coming, everybody.

The speakers

Tory Gray is CEO of Gray Dot Co, a technical SEO and research consultancy. She's spent recent months building and testing concrete example skills across different SEO disciplines, and brings that hands-on build process directly into this session.