blake (00:00)
Hello, hello. I'm Blake Nielson Head of Accounts at 97th Floor, and this is the campaign.

Thanks for joining us today for another episode of the campaign. I sat down with 97th Floor's own Rachel Bascom

Head of Content and Jasmin Bennett, Enterprise Account Director. In this episode, we discuss when and how to use agentic AI, work with Claud Code, and how to automate content briefs. Let's get into it.

blake (00:27)
Welcome, welcome. Yeah. Today we're gonna talk about Agentic AI and the things we've been doing and building in 97th Floor with Agentic. let's first just talk about what Agentic means for us right now, we've seen this shift where a year ago we were focused heavily on

AI search and and we still are. Everything we're doing is AEO, GEO, AI search, whatever you want to call it. We're still hyper focused in that,

but now we're starting to include more agentic elements to better our campaigns, better our processes, make things more efficient, and ultimately just raise the quality of everything we're doing. yeah, what is what does agentic mean to you and like how is that different than just

AI chatbots and how we were using those previously?

Jasmin (01:11)
I think agentic is really powerful because it actually gives you the ability to do something. It's essentially an actually literally an agent that is deploying work for you.

Whereas with a lot of the stuff that we were previously doing within Chat, I think that there's obvious limitations to how far it can go. It can extract information from the web for you and it can help you

pull together pieces of information. But the agentic side of things is what actually allows you to automate and streamline your processes. So with the agentic side of a lot of these AI tools, it can actually empower your work where it's writing messages for you. If you connect your, you know, Claude to Slack, for example, or it can actually deploy different tasks within Monday.com if you connect your

You know, Zoom transcripts, it can take all of your action items from your meetings and directly input those into Monday.com, whereas the actual Chat functions or chatbots aren't able to do that. So it really does allow you to actually have a second arm to your work and make it a lot more streamlined in terms of deploying and executing for you.

blake (02:22)
Yeah, we jumped on AI chatbots early on, like with Chat GPT. We've been using that for years. There still is a lot of manual work that goes into that. And it never really got

like it never got that much better. But with agentic, now you can actually see those processes like end to end versus

asking one question, getting a really good response or getting it to do one skill for you, okay-ish, but still very manual.

Now we can actually see those happen end to end or eighty percent of the way where before it it was just one piece of the puzzle.

Jasmin (02:56)
Totally. And I think that with the agentic side too, you are able to mold it and refine it a lot more because you can give very specific sets of instructions or markdown files. Whereas a lot of the prompt-based conversations, again, you're kind of limited to how far those can go. And it is a lot more like in Chat iteration. And I feel like it's a lot more frustrating because you're not able to

Rachel (02:56)
Yeah.

blake (03:07)
Yeah, yeah.

Jasmin (03:22)
really refine or give as many sets of instructions versus what you can do with very specific skills documents.

blake (03:30)
you good.

Rachel (03:30)
I think that's where the shift has been so beneficial, is that it's a shift from thinking or simulated thinking to action and doing. And that's where can really benefit us as experts in our field. is, know, intelligent people.

You don't necessarily need to outsource your thinking. You know, it can be still helpful to workshop things and we know all the benefits of Chat bots and things and LLMs to help you work through stuff. But when you talk about agentic, it allows you to actually do the things that you would take up your otherwise time that could be spent for strategy, for thinking. For those things, you you talked about like,

getting rid of the 80%. So then you can focus on the 20%, which is what your brain is actually needed for. And that's been our philosophy, I think, from the beginning, even when we still were only really had the chatbot format, everybody sort of jumped to like, all right, well, we can create content then, because that was a doing thing. But it still was so much of like, that's the most boring way to use AI, and it's not the most effective way to use AI.

And Agentic is really helping us tap into the way more effective ways to use it.

blake (04:42)
Yeah.

Totally. I I'm thinking about this time where early on in AI chatbots, it was, hey, we can write content with AI. And everybody in the industry panicked. It's like, we're gonna have to charge a lot less for content because AI will just write all of it for you. Well,

pretty soon everybody learned that like AI written content's not great. It's very surface level. It doesn't really people wanna hear from humans, right?

But we still need to create content at scale. That problem hasn't gone away. And so then

agentic is now we've been able to say, okay, we're not necessarily just writing content. Now we're building processes to help us write content. And that's gonna be in briefs, that's gonna be in research, that's gonna be in all the elements that we take to write a really good piece of content, and we can do that more efficiently. And then that 20% just having a human write it and give their

you know, perspective and POV and all that throughout and make a really high quality piece of content

let's shift gears to some of the things we've been working on. So we've been thinking a lot about just how do we make more processes

automated through agentic AI. And one of the biggest things is going to be this content arm where we write briefs. And our briefs at 97th Floor are really, really good. we have multiple APIs, multiple tools that we've spent years building to have really good research to help us rank well in AI search, help us rank well in

traditional search engines and bring all that information to make really good content briefs. They take a lot of time. There's a lot of work that goes into that. And so to be able to do this at scale, we've tried to figure out how we can add agentic elements to make that faster. That's something I know, Rachel, you've been spearheading. Do you want to talk about that process?

Rachel (06:27)
I think I had a lot of hesitations early on with this. And I think it's natural with AI in general. There's a lot of skepticism. And a lot of the time it's for good reason. But the thing is, when you talk about content briefs, especially when you're talking about adding research and optimizing for SEO,

automated content briefs have been a thing for a while, but the quality has not necessarily been there or really at all been there. And we have known for a long time as an agency that like, yeah, you can't really pump out something with a machine and have no human eyes touch it and then have it be good.

blake (06:53)
Right.

Rachel (07:10)
But AI, suddenly it's like, is that still true? Or is it just another version of that? it's not just about automation. There are automated elements that we want to streamline and save on valuable time and resources. But to us, the most important thing is quality.

And we've had a lot of pitfalls in the process, you know, early on, even when we only had sort of just basic Chat GPT, trying out things and learning as we go. And we had a lot of situations before we realized, you know, before it was known that the AI likes to hallucinate regularly, we would create something and then later like

blake (07:48)
Mm-hmm.

Rachel (07:51)
obviously verify it with real research, our own research, and you'd have to go back to Chat and be like, where did you come up with this? I made it up. And so we had a lot of moments like that where it was a learning curve.

blake (07:58)
What can you do?

Yeah.

there was a lot of that early on. And we use Claude and so we're using Claude and we'd build these really cool, you know, processes, procedures. And then yeah, exactly going in like, wait, where where did you get this this piece of it? And Claude would Claude just straight up say, I made that up. okay. well, we can't have that. So now that's another problem of how do we source that information? Like, what do we need to tap into?

So kind of working through the troubleshooting and the A B like process of or the QA testing of this, we were able to open up where the pitfalls are and then where do we where we fix those. So it's like, connect this API or build this separate skill that will package back into this to solve those things because a lot of what we were getting was just AI hallucinations early on. And the quality suffers because of that, even though it looks good on surface.

Rachel (08:50)
And by the way, just burning through tokens, insane usage for just like stuff that Claude made up.

blake (08:54)
Yes.

Yeah. Let's I think, you know, none of none of us are developers by trade, right? So we're all we're all learning this vibe coding world of building processes and procedures. We're all really excited. But I think that was one of the biggest breakthroughs for us was understanding how infrastructure works and how to package skills. And I think this is something that Jasmin was able to unlock for us at 97th Floor of not just building really tall.

processes that burn through tokens, but being able to like package those up and have individual skills. Jasmin, do you want talk about kind of your journey to that?

Jasmin (09:34)
Yeah, so the way that I was able to actually kind of piece all of this together was I started finding different like Reddit forums that would then take me to GitHub and I would see a lot of conversations around, here's how I am, you know, automating my entire marketing stack and I'm able to run my Google Ads, I'm run able to run my local SEO.

And a lot of what I was able to find on GitHub are these pre-built skills that you can upload directly into Claude and start working within those to actually run your functions. And when I was looking at the actual setup within GitHub, I was able to kind of extract and see how a lot of these other files were built. And it became a lot more clear to me that the way that we were trying to build our content marketing brief, for example.

was a lot different than how these setups were working. So the way that we originally went about trying to build our content marketing brief was we had one master MD file that was essentially asking it to do eight different sets of instructions. It was asking it to do everything from pulling the research, asking it to actually build the content framework for the brief, and then also packaging it into

the properly formatted file. And that structure is just very bloated, I would say. You're telling it to do so many different things that it becomes very inefficient. And when you actually chunk it up into different skills documents, what I was seeing is you essentially want to package each individual step of the process.

So if I have something that is very persona driven, I would have a persona-based skill. And then within that, I would have the actual content framing. What steps does it need to do in order to deploy or to be able to actually identify what's ranking well within the CERFs? That would be a second step of the process. And each of those different steps is essentially its own set of instructions. And you can pull on those independently as needed.

Or you can also kind of package them together in one overarching file. And then you would have the MD markdown file that tells you when to reference each of those skills. So the skills gives you more flexibility in the sense that if I'm creating a persona skill, that could be really valuable for a content brief.

Or it could also be valuable for something like a customer journey map. If I want to, you know, do a user experience report. I want to look at that on the website. So it's a lot more customizable where you can use these things interchangeably versus if you are creating this really bloated system, you're confusing the AI, but also really pigeonholing yourself into that specific set of processes.

Like if we would have continued down that route, essentially you would have had to create a unique document for each individual client in order for it to actually learn the proper information about that client in order to produce a piece of content that is educated or sophisticated enough because just by doing it in one standalone with the way that we were doing it, it wouldn't have been able to like produce quality content.

Rachel, you can probably chime in more based on some of the things that you ended up taking over. But that's just kind of the quick and dirty version of, you know, what I was able to figure out there.

Rachel (13:08)
Yeah, Jasmin's initial infrastructure changed the game and it had a massive impact on two main things. One was usage. We talked about how we were early on just burning through tokens, just crazy amounts of usage because essentially every time you ran a new...

brief, it would have to review the full skill, the full Chat, all the context every single time, and that was just not efficient at all. And then alongside that, there was also major implications for quality assurance.

in the ways that Jasmin just described, you know, with that big monolithic skill, it was hard to pinpoint where issues were happening. And then you could try and fix one issue and then something else would break. And then, you know, you're kind of depending on Claude actually reviewing everything every single time. And sometimes Claude just forgets or like you lose certain aspects of the memory.

blake (13:43)
Thank you.

Yeah.

It forgets all the time. Like I I see that where I'm like, Hey, I think

you didn't run this. It's like, you're right.

Rachel (14:11)
Yeah. So by having that like a skill pipeline, that really changed both of those things. And so for example, we were able to take things like a client profile section and really work on that and say, this is right now, like getting all that context for each client is using up a ton of tokens. How can we make it more efficient?

how could we create cached profiles and then run with that. And then similar with like quality concerns. This one thing is like not working every single time. Let's bake in some checkpoints and some more explicit instructions for that stage of the process. And it's just the one skill rather than having to redo like the whole thing each time. I think that was, that speaks to like a lot of the early

lessons that we learned like early on in experimenting with agentic and one big one was if there's already existing resources use them. I think every time we wanted to start a new process we'd be like all right let me talk to Claude about it but and then it would start doing immediately but if you workshop first and you can workshop with Claude or Gemini or Chat GPT whatever you're

blake (15:14)
Yeah.

Rachel (15:23)
your choices and then you can also, like Jasmin talked about finding resources on GitHub where people had already built things out. It's much more efficient and much better to lean on those resources and not have to start from square one each time and reinvent the wheel every single time you start something.

blake (15:40)
Yeah.

Jasmin (15:41)
I was just gonna say, I feel like one of the things that I learned is when you're talking to these models, if you don't take the time to brainstorm with it, it will just run and do its own thing. And I think we saw that with the first iteration of the content brief where, you know, we explained what we wanted it to do, but we didn't explicitly ask it, okay, like

How should I go about doing this? What's the most efficient way of doing this? And it just was kind of spiraling in the direction that we told it to go in. what I first started doing was I essentially gave it the original markdown file and had to reverse engineer that. And I would ask it questions like what in this process is using the most tokens? And it would be able to parse that down and find the specific.

Areas that were really slowing down the system or bloating it. And so I feel like even doing things like that, if you know that you have a problem, just kind of reverse engineering it, trying to actually talk with the you know AI platform to figure out where the breakage point is, that gave me a lot more clarity into figuring out what we needed to do differently, versus sometimes I feel like you can get into a trap of just trying to like.

redo different things or trying to, you know, keep evolving a broken system. And so I feel like it is a lot better to actually like slow down, ask questions, and do more of that brainstorming and workshopping with it in order to create something better because it will just execute what you're telling it to do. And as somebody that doesn't know anything about AI platforms or AI tools, it's very easy to send it down the completely wrong direction.

blake (17:26)
I think you're kind of touching on what are the biggest benefits of Claude and one of the like the most frustrating parts is that you really have to hand hold and you have to ask it questions and you have to problem solve with Claude. Otherwise, it will give you something very surface level. And so sometimes when I'm using Claude versus Gemini, like I'll get something that is faster and like quicker on a different platform, but it doesn't necessarily have all the intricacies that I need.

And I feel like I usually uncover those more when I'm using Claude because it kind of asks you questions or does things without asking, but then you know that it's wrong and it's very obvious. So you have to really workshop with it. But I do think the end result is better and like what you've put together and built typically is more usable and shareable.

Something I want to touch on or call out that I think both of you said there's really, three things with how we've built out this infrastructure. And so the again, the idea here is that we're taking skills within Cloud and packaging those into larger MD files for larger skills that are then transferable to different clients, to different processes and procedures. so, you know, one of the biggest benefits here is the the troubleshooting. If you have one issue with

This skill or this package you've just built, it's probably one of the individual skills. We can figure that out. We can fix that individual thing, take it out, rebuild it, put it back in. That and then there's the efficiency. We we're we're using less tokens as we have it kind of built in this package. We're not just doing everything all at once where it's long and bloated. But I think the most valuable part here, and the third thing is those skills are transferable.

to many different things. And so us being an audience first agency, we can build these individual skills and then we can and packages that we can take those and put those other things. So one might be for a content brief, one might be for a customer journey map or a user experience report, or we can take those say, hey, these are for meta descriptions we're writing. Here's all the persona, here's all the audience information you need to know. We're writing ads, here's the cop here's information we know about our users.

help us with the brainstorm process. We can use that for like ad agentic things. And just having those individual skills be transferable across all the disciplines we do as an agency is super valuable and more efficient. So

Jasmin (19:35)
Yeah, I think there's a lot that you can unlock with that, especially even if you just take a very simple skill, like a audience skill, for example, that can be applied across a written piece of content for a blog. You can apply that to your ad copy. And just having that skill alone allows you to actually create more quality, better work that you can, you know, really streamline across all of these different pieces.

And I do feel like there's also an element of you're able to cross compare insights a little bit more. you know, I know my one specialty is advertising. That's something that I know very well, but I don't always understand the back end implications of how a website is working. But with a lot of these skills, you know, I am able to pull things together to give me more,

generalized insights about website performance if something is slow or if I need to, you know, figure out how that might be affecting or harming conversion rates, you can find skills that will help you fill in some of those knowledge gaps that you might have within your own specialties.

blake (20:38)
Yeah, absolutely.

Rachel (20:39)
Yeah, I think a really

good example of this is early on in the process of creating these briefs, one thing that we really wanted to incorporate was LLM analysis. So like a SERP analysis, but for LLMs, basically answer analysis for prompts.

And what that looked like for our marketers was going into Chat GPT, Perplexity, all of the main LLMs and searching incognito, logged out and taking the results, analyzing and then using that to inform our content, putting in the brief. A lot of manual work there.

Blake separately created a skill that was a brand sentiment tool that would crawl those LLMs to analyze the sentiment of your brand, which is a really valuable tool that we've created and we still use. But when we saw

blake (21:27)
But

Rachel (21:31)
demo that tool, all the content marketers were like, wait, could we use that for our brief process? And so we adapted that tool.

and baked it into our brief process so that we can now crawl those LLMs with a single search. And it's really awesome.

blake (21:47)
Right. Yeah. Just brings all that data

and helps build that content better because we have something else running that then packages into that. And that brand sentiment was something we'd been doing for a while. And it's a very manual process. It was we were going in and asking a whole bunch of questions for each LLM about a brand and building this out. And it was really valuable because of the results we've seen from it. You can manipulate things in AI search pretty quickly if you can just again adjust the sentiment of your brand.

Which often comes down to content consolidation or just better content on your site. Like let's focus on the things that matter for our brand. And often there's a lot of bloat. You can fix that. You can identify those with brand sentiment. But yeah, the process we built is just, we're taking all these things, analyze it all together, package it all up, give us actionable insights, and then have that be like a recurring thing that we can run on a regular basis to make sure we're like all aligned there.

Anything else that you guys are excited about when it comes to agentic? Anything else guys yeah we should call out?

see.

Rachel (22:46)
fun to play around with and experiment. Like I've had all these like side little pet projects at the same time. You know, I used Replit to create a tool that will.

blake (22:54)
Yeah.

Rachel (23:00)
give me audience analysis based on Reddit and relevant subreddits to your audience. And I've had a lot of fun using that and it's been super valuable. I also created a Chrome extension. And that was a good example of really planning your work. Because I've mapped that out on Gemini, told it everything, all the tools that I had access to, told it to give me examples

of existing software that I could just sign up for that exists and let me try those too or give me a plan and it gave me a plan that I could build a Chrome extension using Claude and then and host it with Replit and it was really fun and was another example of just like how you should approach it rather than just like going on Claude, build this.

blake (23:52)
I think what you're touching on is we've talked about like infrastructure and how to make the tools we're building more efficient, but you're talking about prompt engineering and having the prompts that we're asking the AIs like are more efficient. So when you say I want to build this process, what I typically do when I'm building something in Claude is I'll say, like, hey, here is everything that I would do manually. I go to this, I open this, I look at this, I have

These this data source and I explain the whole process and that's one prompt. And you put that into a single prompt, and then it's like, okay, great, now it has something to work with, versus just going and be like, Hey, have this idea. Can you help me build this? Because now you're using more tokens. yeah, you're not very efficient with it. So that planning that goes into it, I think is a really critical piece.

Rachel (24:36)
Yeah, and always have it.

Tell it not to act yet. Tell it, yeah. Always say like, give me your plan. Like, what are you going to do? And then, and then...

blake (24:40)
yeah.

Yeah.

Jasmin (24:45)
I've also saved a lot of time just by telling it to create the prompts for me. And sometimes I do like ChatGPT better for prompt creation, but I'll say, okay, here's what I need to do. I'll explain that I need to audit an ads account, for example. the way that I would, you know, tell Claude to do that if I was just describing the instructions myself would probably be very

short and parsed, I would say, you know, look at the campaign performance, look at the keyword performance, tell me the overall trends that you're seeing the last seven days versus the previous seven days. I'll basically describe that, put that into Chat GPT and ask it to build a prompt for me. And it's like massive paragraphs. It's like you are a senior ad sp specialist and it will go into really in-depth detail about each individual step of the audit process that it should uncover.

And then it gives a lot more specific instructions too in terms of like one I see a lot is it will tell it to give me the top five areas for opportunities this week, the most immediate things that I can action on in order to improve the account efficiency. And those are just things that like I naturally probably wouldn't think of if I'm developing my own prompt. And that's where I do feel like it is helpful to kind of workshop and just describe what you want it to do.

These LLMs know how they think and function. And so if you can just give it that guidance, you know, it's probably going to be able to produce something a lot better than what you could produce.

blake (26:05)
Yeah.

I hadn't even thought about that. That's crazy. Cause yeah, you're right. the way that you write something and frame something, it obviously can understand that. But when you have an LLM, write that prompt out for you, it's writing it then for an LLM and it's gonna give all the information that it would want to do the thing. Wow, that's super cool. I hadn't even thought about that.

Rachel (26:30)
Yeah, I think one of my favorite ways to approach it is to say,

Give me like three to five options. Here are my main concerns and rank them by this, know, like price, ease, know, timeline, sort of, you know, all my main concerns. Here's what I needed to do. And then it will give me, you know, a few options and it will usually be like, this is your best one. And I might agree and I might not, but at least I can see like the pros and cons of each.

blake (26:56)
See it do it. Yeah. I think you hit on a good point about like telling it not to act, because it's so easy. It'll give you options and then it'll just start building something. And then you're like, and then you end up with something that's like a B minus, and you're like, well, it will get me there. And then you're like, well, is this actually saving me time? Is this even good? Is this more efficient? Is the quality even there? so having it stop, work through it, okay, yeah, now go. Like that's yeah, that's the move.

Well, thanks for the conversation today. This has been good. It's been fun to I've learned some things. It's interesting because I work with both of you so much. We we have this task force here at ninety seventh Floor. We're trying to solve all these AI problems and we meet weekly and I get so much value from those conversations. But even today, like I'm learning things from both of you. So thanks for the conversation today.

Rachel (27:40)
Yeah, no problem.

blake (27:41)
That's all for today. I had a great time talking with Rachel and Jasmin Connect with us on LinkedIn if you want to discuss how to use agentic AI in your marketing campaigns. You can find past episodes of the campaign and examples of our work at ninety seventhfloor.com. That's all for now. Thanks for listening.