How I Run A 0-Employee Marketing Agency With AI Tools

Hey everyone. I just had an incredibly mind-blowing experience. I just had a solo entrepreneur who’s running a six-figure business by themselves with AI walk me through exactly what they did — the tools they use, their workflow, their exact prompts — and they’re sharing it all with you. That is today’s show. You’re going to want to stick around. We’ve got an amazing guest: Barbara from Startup Cookie is going to walk us through everything. If you’re looking for inspiration for starting an AI business, this is the perfect show for you. Let’s get to today’s show.

I guess, to kick us off — where did you get started? If you’re somebody out there and you’re like, “Hey, I think I want to run a solo business with AI,” how do you even get started doing that? Where do you start from a tooling and process perspective?

When I started this agency, we just got ChatGPT. It had just launched. I was working for a YC startup as head of content, and I had been struggling for a few years managing an entire team of people just to sustain, honestly, a pretty mid content schedule — your average once-a-week blog post, social media posts, creating all of these different resources.

I wasn’t doing any of the content creation. I was just managing people, communicating with people. It cost a lot of money, it took a lot of time, and it wasn’t the most fun work. It was just managing a lot of people. And when I saw what AI could do — even in the early stages — if you prompted correctly, I was like, “Wait, what if I can just do this myself without hiring anyone?”

I started off pretty small, but already with our first client I realized this is completely doable. And now, when you look at the last three months and the improvements — what OpenAI launched and everything that’s progressing in the AI world — you realize you can effectively scale your agency exponentially just by using these tools.

I love that you started the company because you had the realization that managing people is horrible because it takes a lot of time. Now I go into ChatGPT, and when it doesn’t do what I say, I turn caps lock on and just absolutely…

Exactly. You don’t have all of the messiness that comes from managing humans. The only thing that might be worse is your holiday parties might be a bit odd.

They’re great. It’s just me here in the hot tub.

Mark Andreessen was on Lex Fridman’s podcast, and he talked a lot about the fact that he’s really excited to see the first companies where the CEO is just managing a bunch of agents, or the agent is managing a bunch of agents and the human is supporting those agents. Your version is like, “Hey, I’m going to create a company and just have all of these agents do these use cases.”

One of the interesting things is Sequoia had a report recently that showed the stark differences in AI usage if it’s built into services versus AI usage if it’s a consumer app. It’s actually been much more successful in services businesses.

What was the very first use case — content-wise — that worked really well? What was the first content agent you built where you were like, “This is actually better than a human”? Because I think most people’s experience with content and AI is the opposite — it’s much worse than a human. And I think that comes down to prompting intelligently versus the AI capabilities.

Hot take incoming. I worked with a lot of freelance writers while I was working for startups full time. Unfortunately, most startups don’t have the budget to pay for exceptional writers — the kind I would still fully pay today. Most are around $200–$300 per article.

My experience was: you give someone a brief with extensive detail and context to write a piece. That piece comes back and you essentially have to rewrite 60–70% of it. That was just the process. And we called that “efficient” because we weren’t willing to pay for each piece at a higher level. If you’re trying to sustain a certain schedule, it’s difficult to set up $100K a year just for your blog to be active.

But it was a lot of work on my end, and it didn’t feel creative. It didn’t feel innovative. It was just the best I could get out of that quality of writer.

Then as soon as I tried it with ChatGPT, I was like: not only is this better than the usual draft I get delivered, I can work with it, and I don’t have to wait two or three weeks for it to be delivered. I could cut out the waiting and the communicating and all of that work up front — without getting a result I was remotely happy with.

Like you mentioned, it’s all about the prompt. The first time it clicked for me was when I started using our webinar work — recordings and transcripts — as a starting point for blog posts. I was like, “Let me see what kind of blog post we get if I have this transcript, and I explain our brand guidelines and writing style.”

It was exponentially better because you have a genuine piece of content — insights and knowledge you gathered first — and then you use it to create something, instead of going into ChatGPT and prompting “write an article on this topic.” We all know that doesn’t work anymore.

That was the first time it clicked: if I have something to work with, I can repurpose it across channels — social media, blog, email — and try new growth channels pretty easily. It’s kind of endless if you have that good starting piece. And honestly, that’s still what I do today.

We never start by prompting AI from scratch to do anything — even a simple social media post. We always have something to start with, whether it’s a transcript from a webinar, a podcast episode like this, or anything similar.

I actually started my agency with a co-founder. Now he’s here only to interview founders and CEOs, because if they don’t have a podcast or a webinar series, he gets on a call with them and talks about what they’re excited about — what they’re building, their views, hot takes, anything they want to communicate but can’t do themselves. We manage their personal accounts as well, but we always start with that transcript so we have someone’s actual insights and knowledge. We also get tone of voice, writing style, phrasing — that’s what you need for the result to be good.

Of course, you still have to fiddle with it and play with it to get it to the quality you want. Each founder gets one interview a month — about an hour. That’s totally fine. Other companies have webinars and things like that we use.

That was the first time it clicked that this is not only producing great results — it’s fun for both us and the founders. The founders love it. They’re like, “Oh my God, this was like a therapy session for us — a monthly one.” They get to explore topics they don’t get to explore in product meetings, because sometimes you just need to talk through concepts. You test messaging and see how it performs.

It’s been working really well so far. As AI gets better, it’s getting easier and easier.

So if you were going to suggest how people follow this, it’s: get context, because context really matters. And one thing Kipp and I were just lacking on is having someone interview people — that’s a great way to get additional context that’s not available elsewhere.

Exactly. And especially if you work with a company for a while — our average client has stayed with us for a year and a half already — you end up with a lot of content. Then teams reach out — especially for busy founders — and you can’t always ask them, “What’s your take on this thing going around right now?” Teams message us and they’re like, “What do you think he would say? What would be his phrasing?”

Then I put in 10–20 transcripts we have from him and ask ChatGPT, “Did he mention this? Did we talk about this topic? Let me know.” And we work with the team.

It’s like creating a persona of someone and having that available in ChatGPT.

Yeah, we do that both in Claude and ChatGPT. You can have a project with all that context, and you can interact with it to get clarity.

So if step one is getting context — either publicly available transcripts or recording conversations — then step two is doing stuff with it: prompting and engaging with that.

We were hoping you could show some prompts and walk us through how you’re doing this so someone at home — a marketer or agency owner — knows what to do.

Do you want everything Barbara is talking about in the show? We’ve taken Barbara’s insights from this episode and created an incredible full-stack AI Marketing Toolkit. It’s packed with things you can copy and paste right away — all the AI prompts we talked about are loaded into this guide. It includes use cases, customization tips, and quality control so you can copy and paste everything into one killer AI workflow. You’re going to want to get it right now. It’s at the link in the description below, or you can scan the QR code. Now let’s get right back to today’s show.

Okay, so I created this a while ago just to have an overview for myself. The “before” is basically everyone I had to hire when I was working as head of content full time. I calculated how much that would be per year based on different salaries.

Even with freelancers, it’s more money at the end of the year than you expect. You always have reviews, another request for the writer, and it all adds up. Then you introduce SEO copywriters, video editors, audio editors — they’re expensive. Their contract periods are often at least six months, so you have to commit. That’s difficult when you’re trying to test different channels and test what content works.

If you have a proven channel, it’s worth investing. But when you’re still testing — and I’ve only been working for startups — my mindset is scrappy. How do we test as fast as possible, with good quality, without spending a bunch of budget?

So this is the whole list of people we needed: copywriter, marketing designer, all these roles you either need full time or freelancers. Then you need someone like me as head of content — but instead of creating content, you’re just managing all these people. It’s not efficient.

Now I have that team — and AI. I put my entire tool stack here. None of these tools are unheard of, but people usually don’t know about Granola. Granola is excellent for meeting transcripts, and you can chat with the meeting transcript afterward. Instead of uploading it into ChatGPT and asking it questions, you can immediately ask, “Wait, what was talked about? What did he mention?” right there. It works really well.

And then: Riverside, Midjourney, Descript, Superwhisper, OpenAI, Claude — that replaces all these roles, even some more technical ones, like a data analyst. For marketing, you can get away with social analytics. But we had a data analyst on the team — a guy whose main thing was Tableau. Full time. You’d ask, “Can you give me these numbers?” and he would.

One of our clients built a data analytics and visualization tool where you upload a dataset and you chat with it. You ask, “Tell me these numbers for this quarter and how it compares to last year.” Non-technical people couldn’t do that before. I couldn’t do that before. And it can cost less than $1K per year. That’s my cost of running an agency. That’s crazy.

What you really have here is a stack to go from idea to final piece of content. You still don’t have the email service provider or distribution costs like ads. But this is a 400x cheaper way to go from “I have an idea” to “I have a finished piece of content.”

Yeah. Sometimes you need to go through all of this to test a new channel like email. The biggest friction before AI was: you get energy, you get an idea, you want to test it — and then when you delegate it, you’re testing it two months later. That’s not ideal. And you spend money to test an idea that might flop.

It prevented a lot of companies from testing different channels because of lack of budget or resources.

So I have our workflow with AI — what we talked about earlier. You start with obtaining authentic insights. Once you have that unique insight — from a webinar, podcast episode, founder — then you can build from there.

I often tell founders: if you don’t have the money to pay an agency, turn on Superwhisper and record yourself talking. You don’t need someone to interview you. If you’re a founder starting out and nobody’s inviting you on podcasts, you can do it yourself. Talk about what happened during the week. What you learned this month. How the team is doing. How the product is doing. Challenges. How you overcame them. Then use that transcript. It doesn’t have to be an interview.

I’ll also link all my prompts in a sheet under the episode. People love going in and copy-pasting. It’s always public.

Our first prompts are usually: “Extract all of the insightful topics that we covered in this conversation.” It’s an hour. Sometimes I listen to it if it’s interesting. But the reality is: you don’t have time to do that every time. So you ask ChatGPT what topics were discussed, which ones were interesting, which ones were fully covered and fleshed out.

It gives you a list. You can ask for timestamps, then go to that part of the conversation. I still often listen manually a bit, because sometimes it misses the part where you’re like, “That was a great zinger.” You have to work with it. You have to be present. If you’re trying to prompt AI without thinking about the content piece, it’s not the best strategy.

Once we have a list of insights, I go one by one and tell it to summarize. I tell it to look at the entire transcript and fill in gaps, add context, and make sure it only takes information from the transcript. That’s important because sometimes it pulls in stuff online that’s relevant but wasn’t said. When a founder is talking about a product, you need accuracy and their phrasing — their features, their branding.

Then I instruct it to create a social post from it.

Is there any prompting you’re doing to extract interesting takeaways? For example, I’ve played around with mapping talking points to templates that perform well on different platforms. How do you have it understand what’s interesting?

It depends on what they want to talk about online and what their persona is. We help them with branding and recurring topics. But it’s about not expecting ChatGPT to deliver a perfect list from one question.

One thing that’s helped me: I very rarely type into ChatGPT. I use Superwhisper. I use my voice. The results are infinitely better because I add more context than I would when writing. When you talk, you naturally give more context, and ChatGPT performs better when you give it as much context as possible.

So I’ll talk through it: “This founder wants to talk more about the product. Their product is XYZ. We want something interesting for this target group. Extract insights from this transcript.” Then I read through and see what looks good.

A big part of creating content with AI is good judgment — knowing what quality looks like. That’s hard to cultivate and hard to put into a prompt. You have to be present. You have to look at the list and think critically: does this sound interesting? Can it become a bigger piece? If not, move on.

Sometimes from a transcript I get one or two insights. Other times there’s a lot. Sometimes it becomes a list of posts. Other times it becomes a bigger resource — like a full email newsletter series. You have to be strategic and sometimes harsh: “That’s not interesting. It’s been said before.”

So I prompt a lot. It’s not one long prompt — it’s several smaller ones. I’m having a conversation with ChatGPT. I highly recommend people talk to ChatGPT instead of typing, especially for bigger pieces like blog posts. It works much better.

You’re basically saying: talk to it about what you want to do and then have it create the prompt. It’s hard to understand without seeing it. Could you show some prompts?

And you mentioned you’d share the spreadsheet link in the description so people can copy-paste. But can you walk us through a couple of your core power prompts for co-creating content with AI?

Awesome. When you take a unique piece of content and download a transcript — or any written content, like a resource guide — this is the main starting prompt I use:

“The text below is a transcript from a conversation. Your job is to do the following…”

Then I ask it to list the main ideas from the conversation that I can repurpose into an insightful blog post. It doesn’t have to be a blog post — it can be anything. Focus on the most unique and insightful topics. Capture the real meat, not a vague, useless summary.

You can be a little mean in your prompts. Be direct.

Then once you see the list, think critically: if you saw this on social media or as a blog post, would you read it? Would you click? If not, it might not be worth creating.

People fall into the rabbit hole of “I can make a lot of content easily, so I should.” But AI can help you zero in on the most unique, interesting, insightful content — not just quantity. Especially on LinkedIn, it’s rough. I looked at the comments on a recent post and I’m like, “These are all AI-generated.”

Once you find an idea you like, tell it to focus on that. Go one by one. If you tell it to take the whole list and make posts, it gets messy. And it’s difficult to figure out what it hallucinated — even though it does that less now.

So you take one idea and tell it to craft an engaging, insightful social media post based on that topic. Use the full transcript, not just that part, because you might come back to the topic later and there are important points. You want to connect everything. ChatGPT has better pattern recognition than me — it connects points better than I could just listening.

Use the entire transcript. Craft a narrative. Follow a social post format. Tell it to be punchy. “Punchy” performs better than “concise.” I recommend using words like punchy, engaging — they’ve worked consistently for me.

Then I instruct it not to use metaphors, catchphrases, jargon — it makes it cringe. I’ll even write in caps: “DON’T BE CRINGE. DON’T BE CLICHÉ.” It works.

I keep a list of words updated: don’t use “ensure,” don’t use “in the realm of,” don’t use the “question sentence” segue. It does that a lot — “The goal?” and then it writes the rest. We all know that’s AI. If you prompt it upfront, you remove a few steps later.

Another thing that’s helped with recent updates is Projects. We have a project for each client and for each deliverable — blog posts, social media. It’s slightly different content, so I prefer separate projects.

Right now it misses about 50% of the time, which is not great, since that’s the main point. But when Projects came out, everyone was like, “Oh, neat feature,” and I was like, “You don’t understand. I can scale my agency because of this.”

For example, for one client where we write weekly blog posts, in the project files I have writing style samples for voice and tone. I asked the founder: who are your favorite writers? What blogs do you read? Whose style do you like? I added all of that. When we create a piece he loves, I add that too. Then general info about the company, the product, things discussed — a huge document of everything you need. So when we start writing, it has context.

Then in the instructions — arguably more important than the documents — I instruct how I want it to behave. I still have to remind it in chat to follow the project instructions, which is annoying, but it’s getting better.

This is the prompt that works best for me across blog posts and social posts — you just adjust based on the deliverable:

“Write in a direct, no-nonsense style. Use clear, simple language. Short sentences. Focus on concrete details. Skip startup buzzwords. Avoid cultural references. Keep it brief. Stay honest. Don’t remove parts of the transcript about challenges or personal things. Share specific examples and data points when relevant. Never use questions as transitions between topics. Avoid sentence structures that negate expectations (‘This isn’t just about…’).”

Then I list words I don’t want it to use.

Sometimes I have to instruct it further. You go through iterations and you start repeating yourself. That’s when you go caps lock and let the frustration out.

I’m also excited about Tasks. I have a task every morning to get the latest news. Most of our clients are in fintech and health tech, so I have one for each. It helps me stay up to date without scrolling Twitter endlessly. Before, I felt like I had to be on X every day because it’s my job. It was low-key ruining my mental health. Now I don’t, especially with Grok and search within X.

That’s my favorite use cases right now, apart from the other prompts.

Okay, that’s amazing. Thank you in advance for sharing the prompts with everyone.

I want to recap what you taught everybody. You basically taught a master class on going from a big, coordinated content team to a smaller, agile content operation. You’re doing it as an agency, but you could do it inside a company.

The first thing you did that’s really important is you have a clear workflow. You know the steps. You document it. You know what tools and process you use at each step. And you’re using the native functionality of those tools deeply. You’re using ChatGPT way more than the average person. People are falling down because they’re not going deep enough into the tools they use. You’re doing the opposite.

You’re deep in the features, the prompting, the language — and you’re getting remarkable output.

The key insight is: you never start from a blank page. You always have context. Context is everything. You either use existing context or have a conversation to create it.

As we close out: if you’re giving advice to people who want to make this transition and automate content creation and sharing stories online, anything we haven’t covered? Where should they start today?

I think it’s very important what you said: you never start with a blank page.

If you’re just starting out with ChatGPT for work, you don’t have to create each prompt yourself. All of our prompts for agency management — also in the sheet — I don’t write them. I tell ChatGPT what I need, and it writes the prompt for me. It’s easier and provides better results. I often cross-reference with Claude.

If it’s a workflow, if it’s a task you do often and want to automate: sit down and record yourself talking through what your process is now and what you want it to be. Sometimes you don’t even have a process — and AI is great because it makes you reassess processes regularly.

Before AI, I’d have a process and follow it until it broke — until I realized I didn’t have time. Now I revisit every week. If something takes long, I talk it out into ChatGPT: “Where can I be more efficient? What can I outsource? What integrations can I set up with Zapier?” And it creates the workflow and the prompt for you.

People think they need to find a prompt. Just tell it what you need to do and it will create the prompt. That’s more powerful than scavenging Twitter or X for prompts that worked for someone else, because you can create something uniquely suited to your use case.

I love it. It goes back to being deep in the tools you have. You don’t need more tools — you need to use the tools you already have more deeply.

I’m a Pro user of ChatGPT, and o1 Pro writes amazing prompts. I just did a show where o1 Pro wrote a prompt, Operator executed it, and then I came back and o1 Pro wrote the code necessary. Something that would’ve taken hours, we did in like 10 minutes. It’s a wild transformation.

Barbara, I really appreciate you joining us on Marketing Against the Grain today. Thank you for sharing the behind-the-scenes of how you’ve done all this. Thanks for joining us.

Thanks for having me.

All right, see everybody next time.