Here are the ideas I keep coming back to from the day.

1. “Good” isn’t good enough anymore

Matthew said something on stage I’ll butcher in paraphrase, but the gist was: AI has raised the floor. Everyone can now write at a 7 out of 10. He’s a writer. He builds his career on words. And it bothers him that the bar for “good writing” has collapsed.

He’s right. If anyone can produce 7/10 work in seconds, 7/10 work no longer earns attention.

You’d think the answer is more polish. It’s not. The answer is being incredible at speaking to exactly who your audience is. AI raised the average. To matter now, you have to be specific, sharp, and unmistakably you.

Those were his words, these are mine: “The only way to compete is you + AI. Not AI alone. Not you alone.”

That doesn’t mean flashy AI graphics in every video, or agents running your whole business in the background. It means: figure out what people actually value from you — or what they could value but haven’t seen yet — and use AI to give them more of it, or a deeper version of it.

AI is a tool. Better: a coworker. And the thing most business operators are missing is protecting the craft they’re known for while handing off everything around it.

2. The two ways I’ve learned the most about AI

Too many people treat learning AI as the objective. The best learning is a byproduct of chasing a real result. Two things have moved the needle for me.

Carve out dedicated time. I went to the Slow hackathon in San Francisco earlier this year. I’d already been vibe coding on nights and weekends (vibe coding = building software with AI). But I wanted a retreat — somewhere away from Julia, away from the kids, with experts, on one objective. That weekend produced vibeshop.juliaberolzheimer.com, a shoppable experience built from years of Julia’s daily looks. I would not have built that in nights-and-weekends mode. The container made it possible.

You can build a mini version of this for yourself. Block a weekend. Set one goal: “By Sunday night I’ll have X launched” or “I’ll have Y problem solved.” Treat it like a sprint, not a study session.

Use it in moments of desperation. Major unlocks for me have come at 9pm when I’m slammed hours of work ahead of me and don’t want to do the busy work, to do the meaningful work.

The clearest example: I was badly behind on campaign emails. dozens of threads, hundreds of total emails. I was about to grind through the inbox the slow way. As a last-ditch move I asked Claude to build me a Kanban priority board of every campaign I had to catch up on. I expected a list.

It came back with a full board — and then went two moves further than I’d asked. It anticipated which campaigns would need follow-ups, flagged dependencies between them, and pre-drafted the next actions. It predicted moves ahead of where I was.

That was the boundary moment. I’d been asking AI to do the task. The shift was asking it to think like it had my job. Once you push past one boundary, the next is easier. A few weeks of that and you’re doing things you didn’t think were possible.

3. Stop looking for the perfect tool. Use a coworker.

The OpenAI VP closed with how their trying to help the creator economy. They’ve thought hard about how to support creators — dashboards, skills, custom tools. His takeaway after all of it: just start using Codex, their general-purpose coding-and-thinking coworker. The equivalent of Cowork or Claude Code.

Teach a person to fish.

He could have pitched a creator-specific product. He pitched the Swiss Army knife instead. He believes — and I do too — that the dashboard approach assumes someone else knows what you need. The coworker approach assumes you do. You bring the problem. AI brings the labor. That’s the relationship that scales with you.

These tools can dig into your analytics, run your operations, help you frame a script, or build something custom for your business. Most creators are still using them like a fancier Google search. There’s so much more on the other side of treating them like a colleague.

What I’d tell a creator starting from zero

1. Pick one annoying back-office task this week (email triage, invoicing, scheduling, analytics, contracts). Hand it to AI. See how close it gets.

2. Block a Saturday. One project, start to finish. No phone, no kids, no partner. Just you and the model. Set a finish line.

3. Stop chasing tools. Pick one general-purpose coworker — Claude, Codex, whatever — and go deep. Depth beats breadth.

The business who’ll matter in the future aren’t the ones with the flashiest AI gimmicks. They’re the ones who used AI to free up time for the work only they can do.

Reply and tell me: What are you incredible at that an AI coworker can support you with? Curious what comes up.

There’s a narrative that vibe-coded apps are trash. And yes — there’s a lot of slop out there. But I don’t think that’s the whole story. When I started publishing content 15 years ago, my photography was genuinely bad. It took two years before I even liked taking pictures. The slop phase is part of the process.

What I’ve actually stopped doing is calling what I build “vibe coding.” The term I keep coming back to is agentic engineering. I’m not sitting in Replit or Lovable making things for fun. I’m building very specific tools to solve very specific problems — and if I can’t articulate the problem clearly, I don’t build it.

The limitation now isn’t whether you can build something. It’s whether you should.

Simple. Then simpler.

The hardest lesson I keep relearning is this: just because you can add a feature doesn’t mean you should.

I went to an AI hackathon earlier this year and got completely carried away. Every five minutes I was adding something new. “Oh, what if it also did this?” The product we eventually launched — Vibe Shop — ended up with eight different paths, endless ways to explore. And once I had analytics running, I could see exactly what was happening.

People were only using one or two features.

All that work. All those features. And people went straight to the same one or two things every time.

If you’re building a multi-page experience, spend most of your time on the homepage. Be specific about what you want people to do. One action. One outcome. Everything else is noise until you’ve proven that works.

The problem has to be yours.

The best person to solve a problem is usually someone who lives inside it every day.

That’s how we’re building Coreli. I spend a lot of time thinking about how creators who compete on taste — people whose whole brand is aesthetic — should have one place that shows off everything they are. All their content, all their shopping links, all in one spot that actually reflects who they are.

The Bio Rewriter we launched came from watching people write nonchalant bios while their content was doing something completely different. The bio is sitting right next to the follow button. Right next to the link in bio click. And most people haven’t thought about it seriously in months.

That’s the kind of problem worth building for. Not “what can I build” but “what am I actually bothered by.”

Research is not optional.

Before I write a line of code now — or ask Claude to — I do real research. UI/UX patterns. Consumer behavior. I talk to people. I run interviews.

That’s not vibe coding. That’s traditional product work. Agentic engineering is just the build layer on top of it.

QA will take longer than you think.

I underestimated this badly in the beginning.

You can build tools that do an automatic first pass and catch the obvious stuff. But actually flowing through your app — from point A to point Z, on mobile, in another browser, with your partner sitting next to you watching — there’s no substitute for it.

What I do now: while I’m going through the app, I voice record the whole session using Wispr Flow or Granola. I narrate everything out loud. “I was trying to click the hero button and got an error.” Then I drop that recording straight into Claude Code and we work through the issues together.

It works. But I won’t pretend I enjoy it. I’ve tried to cut corners more than once. Don’t.

Track everything from day one.

Whatever you launch is not the final version. You need data before you make any changes — and you need to be tracking the right things to get that data.

The best tool I’ve found for this is PostHog. Incredibly easy to set up. And here’s the part I didn’t expect: I almost never log into the PostHog dashboard directly. I do all my analytics work inside Claude Code.

I’ll say: “Go look at how people are flowing through the app. Is anything broken? Is there anything worth optimizing?” Claude Code goes and runs the queries, comes back 5 to 10 minutes later with a fleshed-out answer. I pick two or three things I want to act on, it creates the tickets in Linear, and then I say “go ahead and make those fixes.” Boom.

Bio Rewriter started with 20 tracked events. Now it’s up to 40. You’ll keep finding new questions you want to answer.

Start with an internal tool.

If you haven’t built anything yet, don’t start with a public product. Build something for yourself or your team first.

The first real thing I built was an internal tool for the JB business. My team is more forgiving than strangers. They’ll tell me when something’s broken. They’ll tell me when something’s confusing. I can fix it fast and ship the next version before it becomes a real problem.

It was only after watching Julia actually use the app — and hearing her say “why can’t I click there?” — that I understood what was actually missing. You can’t get that feedback from analytics alone. Sit with someone and watch them use your product. Have them talk out loud. It’s uncomfortable and it’s worth it.

When you get stuck, just ask Claude.

You will get stuck. You won’t know what an API key is. You won’t know how to integrate PostHog. You won’t know what half the error messages mean.

Just ask. If the answer doesn’t make sense, take a screenshot and ask again. You’ll get there.

The best thing that’s happened in the last year is that most integrations are getting so simple that you literally copy a snippet, drop it into Claude Code, and you’re done.

The limitation is not the tools anymore. It’s knowing what to build and being disciplined enough to build the simple version first.

I can’t wait to see what you ship.

And don’t ask me any more questions. Just ask Claude.

infrequent AI is for operators and business owners who want AI to actually work for their business, not just in theory. No hype. Only when it matters.

If you know someone who keeps saying they want to build something, send this to them.

Share

I spend most of my waking hours building with and using this stuff — products, switching tools, dragging the team along — and I’m overwhelmed. So I can only imagine how it feels for everyone else.

So I started something I’m calling the AI Time Capsule.

Here’s the rule: every time I get excited about a new tool or feature, I drop it in. 30 to 60 days later, I open the capsule and ask myself.

Am I actually still using this? Or was it just a honeymoon?

This is the first one. The video walks through it. This post is the cheat sheet — what each tool is, whether it survived, and where to try it.

1. Nano Banana Pro 2[01:27]

Google’s image gen model.

I don’t use it daily. I thought I would.

The use case that stuck: web design. When I’m describing a layout to AI, generating a reference image and dropping it in as a screenshot beats describing it in words.

I built a Pokémon app for my daughter using Claude. I had Nano Banana generate the design first, dropped the image into the build, and the app came out looking exactly like the reference.

Try it: gemini.google/overview/image-generation

2. Canva Magic Layers [02:31]

I love when Canva drops new features but it always bugged me that they used to bury bury them. This one they pinned to the top — and that tells you all you need to hear.

Drop in a flat PNG. Magic Layers scans it and rebuilds the design as editable layers. Text becomes editable text. Objects become draggable. Backgrounds split out clean.

I use this a few times a week. Every time I’m in Canva.

Try it: canva.com/magic-layers

3. Claude Skills [03:50]

I have 40 of them.

If you use Claude and you don’t have a Skill yet, build one today.

A Skill is a reusable instruction pack. You build it once. You reuse it forever.

What do you do every week? Write a certain kind of email. Format a Substack post. Run a brand check on a draft. Each of those is a Skill waiting to happen.

There’s a skill-builder skill that walks you through making your first one. Use that to start. Or take a long chat where you worked through a common problem step by step until you got it right then ask Claude to turn that into a skill.

Try it: Anthropic on Skills

4. Wispr Flow [05:00]

I’m in the top 1% of Wispr Flow users.

Voice in. Clean text out. Works in any app on desktop or phone.

The fastest way to get thoughts out of your head and into AI is your voice. Typing is the bottleneck. Wispr removes it.

It also saves a history of everything you dictate. So if a text field gets wiped, you go back and copy it. That has saved me more than once.

Try it: wisprflow.ai

5. Granola [05:44]

I haven’t been late to a meeting since I installed Granola.

It pings me a minute before each call. It writes the notes. It keeps the transcript.

I use it for casual conversations too. “Hold that thought, let me start Granola.” Then we keep talking and I have a transcript I can reference later.

I was reminded again this week how much I love it when onboarding a new employee and I was able to quickly pull up notes from past conversations to help her.

Try it: granola.ai

6. Claude Code [06:49]

If Claude Chat answers, Claude Code acts.

It’s an agent. Plans. Executes. Builds apps. Analyzes data. Strings dozens of steps together without coming up for air.

This is what I built the Pokémon app in and Coreli Bio. Most of my data work runs through here too.

Try it: claude.com/product/claude-code

7. Last 30 Days [07:30]

I run this 3 to 10 times a day.

It’s a Claude Skill (open source on GitHub) that researches any topic across Reddit, X, YouTube, Hacker News, GitHub, and Polymarket — but only posts from the last 30 days, and only posts with engagement.

That filter is the reason why I use it more than any other research tool.

Default web search pulls 2-year-old blog posts. Last 30 Days pulls what people are actually saying right now. The signal it uses: if 100 people upvoted it, it probably matters.

For AI specifically — where things shift weekly — this changes how usable the research is.

Try it: github.com/mvanhorn/last30days-skill

8. Claude Cowork [08:44]

Cowork is the middle child.

Claude Chat is too simple. Claude Code is too technical for most people. Cowork sits between them.

You give it a goal. It breaks the work into 3 to 9 steps. It runs them.

I use Cowork a dozen times a day.

Ask it to write a blog post and it researches, drafts, runs an editor pass, and turns it into a presentation. Work that used to take me 30 minutes to half a day.

The name fits. It feels like a coworker, not a chatbot.

Try it: download the Claude desktop app and turn on Cowork mode.

9. MCPs [10:00]

This is the layer that makes everything else useful.

MCP stands for Model Context Protocol. Think of it as the AI version of an API — the wiring that lets Claude talk to all your other tools.

My Cowork and Claude Code are wired into Gmail, Slack, Granola, and Notion. So I can say:

“Pull the 5 most important emails I haven’t answered.”

“Summarize today’s transcripts from Granola and drop the action items in a Notion doc.”

Claude becomes the layer between every app you use.

Try it: Anthropic’s MCP intro

What I learned from running this

I’m surprised how many of these survived 60 days. Most things I get excited about don’t.

The pattern: the tools that stuck are the ones that integrated into how I already work. Not the ones that lived in their own tab.

Wispr Flow runs inside every other app. Granola pops up before every meeting. MCPs let Claude reach into the tools I already use.

If a tool needs me to remember to use it, I won’t.

You don’t have to be at the bleeding edge

If you’re feeling overwhelmed, hear this.

You don’t have to be at the bleeding edge of AI to get the benefit of it. You can be 3 months behind. 6 months behind. A year behind.

Most of these 9 tools will still be useful in 2027.

Pick one. Use it for a week. Then pick the next.

I’ll do the next time capsule in 30-60 days. Obsidian is at the top of the list.

They all build things on the internet and in the real world. They all create content. And they all just spent two days building AI products from scratch — side by side, in a room in San Francisco.

I was technically the fashion blogger. The fisherman was Kyle Lee. The woodworker was Jonathan Katz-Moses. The whole thing was hosted by Sam Lessin and Megan Lightcap at Slow Ventures.

I always wanted to do a hackathon. But I was never a technical person, and for a long time, that felt like a wall I couldn’t get past. When the invitation came, I just knew. The tools had gotten good enough. I wanted to be around people. I wanted to do something IRL. So I made it happen, even though the timing was terrible.

I needed it more than I realized.

Back home in Charleston, I felt like I was building in a bubble. I’d gone to a Claude meetup recently — that was great — but I still felt like I wasn’t on the same level as other builders. The only place I was really connecting was online. Watching. Listening. Then going off and trying things by myself.

This was different. This was actually getting in a room with people. And it went way beyond AI. I talked to entrepreneurs about what they’re working on. I heard about paths they’ve taken and mistakes they’ve made. A lot of the creators there were YouTubers, and just hearing how they think about their craft — how they think about adding value to the world — made me rethink how I approach mine.

The morning the hackathon started, I called Julia. “I don’t know what to build,” I told her. I had too many ideas. I wrote everything down, talked out loud, tried to analyze it all. Probably over-analyzed it. Then Julia cut through it: “Do that thing you had the idea for weeks ago. Just make it happen.” It had been sitting on the back burner way too long.

So I did.

I started building something called VibeShop — a way to shop with emotion. Use the beauty of imagery to meet people where they’re at. I’m a big believer that people get caught up in their emotions, and I wanted to lean into that instead of fighting it. Let the feeling guide the experience.

I had no idea how I was going to pull it off. But here’s what I learned: it’s all about the start. Start, start, start, start. Once I forced myself to begin, the ball got rolling and I learned faster than I thought possible. The scope started growing — not in a bad way. “Oh, I could do this. And because I’m doing this, I can do that too.”

By the end of the weekend, here’s what I had:

• 835 outfit photos turned into 19 shoppable vibes. AI analyzed every Daily Look for garments, colors, mood, and season, then clustered them into aesthetic worlds you can browse by feeling — not by date.

• A taste model that knows Julia’s style. VibeShop can score any product against her actual aesthetic. The travel capsule planner uses this to build a trip wardrobe from the archive in seconds.

• Personalization that goes deep. A Style Match quiz surfaces your top 3 Julia aesthetics in 60 seconds. A travel capsule planner builds a trip wardrobe from the archive. And a Most Worn tracker shows which clothing, bags, shoes, and jewelry Julia actually reaches for — cost-per-wear storytelling that beats any product review.

• 4,800 product links audited automatically. Dead affiliate links found, followed through redirect chains, and replaced — plus stock checks so users don’t click sold-out products.

• 100+ tickets from idea to production in one session. Claude Code as the only engineer.

In the past, when I tried building with tools like Replit, each thing I added made the whole project worse. This time, the quality held up as the ambition grew. That’s what’s different about this moment with AI tools.

I brought a teammate — Lolo, who does social content for our Julia Berolzheimer brand and lives in San Francisco. I believe everyone should be a builder, and I wanted to see what she’d come up with. She built her own project, learned a ton, and what she’s making is going to change how our team works internally. But she also saved me a couple of times. There were moments where I got lost or needed someone else’s brain on a problem. She’d sit down, we’d talk through a big feature, and we’d knock it out. Just having that feedback — and a record of the conversation — made all the difference. When you build alone, you don’t get that.

The thing that surprised me most was what happened to my own belief. I’ve had a big vision for a while now. It always felt far away. But each step I took at the hackathon made it feel closer. I talked people through what I was trying to do, and every time I explained it, I believed it a little more.

When I looked around the room, I noticed most people were building internal tools. That gave me more confidence. I was building something consumer-facing, something different from the rest of the room. Going in I’d already launched multiple internal tools and I wanted to lean into the inspiration from my peers, expertise of the teachers and the free time I had to do nothing but this.

And the way we connected was nothing like a networking event. We were typing next to each other. Sharing screens. Debugging together. All of us are creators — we have a shared understanding of what it’s like to put yourself and your work out into the world. We also get what it’s like to run small companies with no lack of ambition and big ideas. You don’t have to explain why you care so much. Everyone already gets it.

I came home with a product which I was able to finish Sunday night, well technically 330am Monday morning. It’s not perfect — not even close. But it’s feature-rich, and I think people are going to love parts of it.

More than that, I came home with relationships. Real ones. The kind that come from building something hard together in a short amount of time. I have a list of people I want to follow up with, keep building with, keep learning from.

If you’re on the fence about doing something like this — a hackathon, a build weekend, whatever your version of it is — just do it. Maybe it’s a 20% commitment. Maybe the timing is bad. That’s fine.

Commit to it anyway. And if you’re building something too, I want to hear about it.

Chat vs. Agent — What’s the Difference?

You probably already use AI chat. Ask it to review a document, write a draft, brainstorm recipe ideas — it’s good at that. But that’s a single step. You tell it what to do, it does that one thing, and you move on.

Cowork is different. It’s an agent. You describe what you’re trying to accomplish, and it figures out the steps to get there. It connects to your tools, opens your browser, reads your apps, and executes a multi-step workflow on your behalf.

That distinction — single step vs. multi-step — is the whole reason agents have been a buzzword for two years. The problem was that until now, they didn’t really work. I wasn’t going to let an AI book a flight for me. Cowork is the first one I actually trust with real tasks.

The Use Case That Sold Me

Every Monday, I need our team aligned on the week ahead. What are the post dates? What sponsors are we working with? What deadlines are coming? Are we traveling? What tasks are outstanding?

I used to either run a team call or cobble together an email manually. Now I have Cowork do it.

I connected it to four different calendars (sponsor posts, non-sponsored posts, preview dates, shoot dates), our travel schedule, Asana for task tracking, Slack for team messages, and the most important email threads about campaigns and projects.

Then I built a skill — basically a saved workflow — that pulls all of that together into a formatted weekly briefing. It runs automatically at 6 AM every Monday (or on demand) and drops the report into our team Slack channel. Each task is a nested reply so team members can check things off with an emoji as they finish them.

It took me maybe an hour to set up. Now it runs itself.

The Small Wins Add Up

The briefing is the headline example, but the smaller stuff is what made me a believer.

Caught a missed request. The weekly briefing surfaced a Slack message from 10 days ago where a team member had asked if we wanted to launch an Easter shop. We completely missed it. I texted Julia, she said yes, and we had it moving within minutes. Without Cowork flagging it, that would have slipped through.

Updated my LinkedIn. I was already logged into LinkedIn in Chrome. I told Cowork to go update my profile. It researched my publicly available info, opened the browser, navigated to the right pages, and made the changes. All I had to do was be logged in.

Cancelled a subscription. Got an email and realized I was still paying for something I didn’t need. Told Cowork to go cancel it. Done.

Reviewed a Substack post before publish. Had a draft open in the browser and asked Cowork to read through it — text and images — and flag anything I should fix before sending. It caught something obvious we’d missed because we were too deep in the weeds. With Substack, once you hit publish, that email is gone to everyone’s inbox. Having a second set of eyes that can actually look at the pre-live draft is valuable.

Pulled analytics for a blog post. I was writing something about Instagram performance and wanted real numbers, not fluff. I was logged into Meta Business Suite, so I asked Cowork to go pull some interesting metrics. It navigated in, found the data, and brought it back.

Try This in 5 Minutes

If you want to see what this feels like without committing to anything:

1. Connect Cowork to your email

2. Have it identify an important email it thinks you should get back to.

3. Have it draft a response automatically.

It’ll put it in your drafts for you to review before sending. Nothing goes out without you approving it. But you’ll see the difference between “AI that answers questions” and “AI that does work for you” immediately.

How I Think About It

The mental model that works: Cowork is an employee. Not a chatbot, not a search engine — an employee.

I talk to it the way I’d talk to someone on my team. I give it context on what I’m trying to do, tell it what good looks like, and let it figure out the steps. The more specific I am up front, the better the output.

Our whole team has switched to Claude, and now I’m pushing everyone to move from chat into Cowork. The framing I use with them: “We just hired you a personal assistant. How would you use that assistant to do your job better?”

Where I Draw the Line

I’m not giving it access to financial accounts. I’m being deliberate about what I let it touch. But for the things I’ve described — pulling data from apps, managing tasks across tools, doing browser-based work on pages where I’m already logged in — it’s been solid.

The Weird Part

I’ve been telling everyone I know about this in person. Entrepreneurs, smart people, people who are generally on top of tech AND AI. Almost no one has heard of Cowork. When I explain it, most of them say some version of “I just don’t get it.” or “I thought that was for collaborating with my team.”.

That gap between how useful this is and how few people know about it is exactly why I’m writing this. If you’ve been using Claude chat and thought that was the ceiling, it’s not. Open the desktop app, go into Cowork, and start treating it like an employee.

You’ll figure out the use cases fast.

One More Thing

If you’re reading this because you watched my video on Instagram — that video was edited by Cowork. All I did was film a very rough 10 minute video about this topic and upload to Descript, my editing software. I didn’t touch Descript again until I asked it to add captions. Cowork managed Descript for me. This blog post was also written and placed into Substack by Cowork. The only thing it couldn’t do was upload the images — it ran into browser security restrictions and I had to drag those in myself. Everything else you’re looking at was done by an AI agent on my desktop.

I talk to a lot of business owners and founders about AI.

The conversation always starts the same way: “I tried it, but the outputs are just so boring and bland. It doesn’t get what I need. I can’t use it.”

45% of my recent survey respondents say they don’t trust their AI’s output.

And my first question is always: “Have you set up custom instructions yet?”

The answer is usually no.

Here’s what’s happening: AI doesn’t know your context, your standards, or what “good” looks like in your business. So it fills the gaps with generic fluff and polite agreement. That’s not your fault — that’s just how these tools work out of the box.

The takeaway

AI feels generic when it’s guessing. Give it a working profile that’s specific to how you work — and it starts behaving like an assistant instead of a chatbot.

Why this isn’t the default: I’m honestly shocked that AI companies don’t walk people through this setup flow. Custom instructions have been around for over a two years, but most people never touch them. It’s a buried setting that could solve half the complaints about “AI doesn’t work for me.”

If you’ve been frustrated with bland outputs, it’s not because you’re bad at AI. It’s because nobody told you to do this one-time setup that makes everything better.

Our readers survey (more responses welcome) revealed that AI would be worth it if you could save time during the week, get better outputs and reduce annoying admin tasks.

When you do this today. It takes 10 minutes and you’ll immediately see the difference in these areas.

The 10-minute setup (do this once)

Step 1: Decide what to include in your custom instructions

This is the foundation. Custom instructions tell the AI who you are, what you care about, and what “good” looks like in your specific context.

What to include:

• Your role and context: “I’m a [specific role] working on [specific type of business/project]”

• Your output preferences: Do you want bullet points? Paragraphs? Examples? Templates?

• Your constraints: Time, budget, team size, technical ability

• Your definition of “good”: What does a useful answer look like for you?

• How you want the AI to handle unknowns: Should it ask questions? Make assumptions? Admit when it doesn’t know?

What NOT to include:

• Generic platitudes (”I value quality and efficiency”) — everyone values those

• Lists of tools you use — unless they’re directly relevant to how you want answers formatted

• Personal details that don’t affect the output (your hobbies, where you live, etc.)

The test: If someone else in your industry could copy/paste your custom instructions and get the same results, they’re too generic. Make them specific to how you work.

Step 2: Choose your approach

You have two options for creating your custom instructions:

Option A: Have the AI interview you (recommended)

Open a chat and say:

Help me write custom instructions for [your role] working on [your business]. Ask me 3-5 questions to understand my context, then draft the instructions for me.

The AI will ask clarifying questions you wouldn’t have thought of. Answer them using voice-to-text (I use Wispr Flow — free trial here) so you can talk naturally instead of typing.

If you want multiple-choice options: Add this to your prompt:

Ask me 3-5 multiple-choice questions about my preferences, then draft custom instructions based on my answers.

If you want open-ended questions: Use the first prompt as-is. The AI will ask open-ended questions and you can respond conversationally.

Option B: Start with a template (faster but less personalized)

If you just want to get started fast, use the template at the bottom of this post. Copy it, fill in the brackets, and paste it into your AI settings.

Step 3: Add your custom instructions to your AI tool

Where to add this:

• ChatGPT: Settings → Personalization → Custom Instructions

• Claude: General → What best describes your work? and What personal preferences should Claude consider in responses?. Slightly more advanced you can make skills or use projects (I’ll expand later in another post)

• Gemini: (depending on account type, this is how I can do custom instructions) Create a GEM with these instructions

• Perplexity: Settings → AI Profile (applies globally) or create a Space with custom instructions for specific projects

Step 4: Test and refine

Custom instructions aren’t “set it and forget it.” You’ll need to test them and refine based on what’s working.

Quick test:

1. Set up your custom instructions

2. Ask the same question you’ve asked before (one that gave you generic output)

3. Compare the results

If it’s noticeably better — more specific, more actionable, more in your voice — you’re on the right track.

If it’s still generic, go back and add:

• More constraints (”I don’t want options, I want your recommendation”)

• More specificity about what “good” looks like (”Include a real example, not a hypothetical”)

• More clarity on your role (”I’m a solo founder, not a team of 10”)

Common refinements after testing:

• “Too long” → Add: “Maximum 3 sentences per point. No preamble.”

• “Too cautious” → Add: “Be direct. I’d rather you make a call and be wrong than hedge everything.”

• “Too many options” → Add: “Give me your top recommendation. Then, if relevant, include 1-2 alternatives.”

• “Too formal” → Add: “Write like you’re talking to a colleague, not a client.”

Quick tip break (don’t skip this)

Use voice-to-text when you’re setting up your custom instructions. If you talk like you talk, you’ll get output that sounds like you. It’s the fastest way to “personalize” results without overthinking prompts.

I’m a big fan of Wispr Flow for this — it doesn’t just transcribe, it saves recordings alongside text so you can review what you said later.

Copy/paste templates (if you just want to start fast)

Basic template

What I do: I'm a [role] working on [type of business]. I make decisions fast and under time pressure.

What I want: Short, practical answers. Bullets > paragraphs. Minimal fluff.

Constraints: Optimize for highest return on time. I don't want 10 options — I want the best 1–3.

Definition of good: Give me a clear recommendation, then a checklist I can do in 30–60 minutes. Include one example/template.

If you need context: Ask up to 3 questions max. Otherwise make reasonable assumptions and state them.

Real examples from actual users

Example 1: Founder juggling content + product

What I do: I'm a founder of a creative tech company. I'm building product, managing a small team, and trying to post content regularly without it feeling like a full-time job.

What I want: Concise, actionable advice. No preamble. Get to the point fast.

Constraints: I have about 30 minutes to review and implement. I don't want research reports — I want "do this now" actions.

Definition of good: Give me one clear recommendation with 2-3 bullets on how to execute. If it's content, include a hook example. If it's product, include a quick implementation path.

If you need context: Ask 2 questions max. If unclear, make an assumption and tell me what you're assuming so I can correct it.

Example 2: Consultant managing multiple clients

What I do: I'm a freelance consultant working across 3-5 clients at once. My calendar is chaos. I need help staying organized and delivering quality work without burning out.

What I want: Systems, not advice. Show me the workflow, not the theory.

Constraints: I work in 60-90 minute blocks between meetings. Optimize for speed and consistency.

Definition of good: Give me a template, checklist, or step-by-step process I can reuse. One example is worth more than three paragraphs of explanation.

If you need context: Ask 3 questions max, then give me something I can use today.

Example 3: Business owner focused on reducing admin load

What I do: I run a small service business. My biggest bottleneck is admin work — emails, scheduling, client follow-ups, invoicing.

What I want: Help me automate or delegate. I don't need perfect, I need "good enough" that saves me 5+ hours a week.

Constraints: I'm not technical. If your answer requires code or complex setup, it won't work for me.

Definition of good: Tell me what tool to use, what to do first, and what I can hand off to someone else. Assume I'll hire a VA or use no-code tools.

If you need context: Ask me what's taking the most time. Otherwise assume standard service business operations.

Quick Intro (skip if you read this post)

I’m Thomas Berolzheimer. I help run a small team behind my wife Julia’s media + commerce business, and AI has landed at the perfect time: it’s helped us add real value to our business without adding headcount.

I’ve also spent countless hours testing new models, tools, and workflows—mostly because I’m curious, and partly because I don’t want our team (or our friends) to waste time chasing shiny things.

And after hundreds of conversations with operators, business owners, decision-makers, and creatives, I’ve noticed a split:

Some people dove in and figured out real solutions for themselves and their business’s using AI.
Everyone else knows AI is a major shift… but hasn’t cracked the code yet.

This newsletter is for that second group. No over hyping. No breaking news. Just the handful of things worth knowing—only when it matters.

Alright—first memo.

The Content Problem: Everyone Wants More. Nobody Gives You Time.

If you’re a serious operator, you’ve probably heard some version of this lately:

“Post on LinkedIn.”
“Start a Substack.”
“Do more Reels.”
“Be on YouTube.”
“Create content.”

That advice is usually directionally right… and practically useless.

Because the hard part isn’t ideas.
The hard part is turning what you know into something publishable without it consuming your life.

The best shortcut I’ve found is simple:

Stop trying to “create content.” Start recording what you already know.

If you have domain expertise—sales, hiring, real estate, skincare, restaurants, investing, law, fitness, agency work—the highest ROI move is to talk first and write later.

Long-form video (or audio) is raw material.
AI turns it into output.

Quick AI Tip (30 seconds)

Your first goal is not “good content.” It’s “reusable raw material.”
If you have a clean 20–40 minute recording, you can turn it into weeks of posts.

The workflow we use: one recording → multiple assets

Step 1: Record a “podcast-style interview” with yourself (20–40 minutes)

Don’t overthink it. Open your phone camera and answer questions.

A great starter format is FAQ + opinions, like:

• The top 10 questions customers always ask

• Common myths in your industry

• What you’d do if you were starting today

• The most expensive mistake you see people make

• “If I only had 30 minutes a week…”

Where to get prompts if you’re stuck:

• Your inbox (real customer questions)

• Your sales calls / discovery calls

• Reddit / forums in your industry (“What should I do about…?”)

• Your team: “What do customers always get confused about?”

Important: You’re not trying to be polished. You’re trying to be useful.

Quick AI Tip (the “anti-perfection” rule)

If you’re hesitating because you don’t like how you look/sound on camera:
Record audio-only first. Same process, same outputs. You can “graduate” to video later.

Step 2: Turn the recording into text automatically

We use Descript and love it. The killer feature is that your video becomes a transcript, and editing the words edits the video.

Even if you never publish the video, the transcript alone is gold.

And if you do plan to publish the video or audio (as a podcast) they have an AI agent called Underlord which will make the edits you request via prompt. This 14 minutes of video was easily cut by almost 1/3. The whole editing process only took about 30 minutes including the review and exporting.

You can also record meetings/calls (Zoom / Google Meet / etc.). We often use Google’s recording + Gemini notes when it makes sense.

Quick note (the 20% threat bucket): if you’re recording calls/meetings, get consent and be thoughtful about confidentiality.

The tool stack

Our simple stack for this workflow:

• Record: iPhone camera / Google Meet / Zoom

• Edit + transcript: Descript

• Repurpose: ChatGPT / Claude / Gemini

• Publish: LinkedIn / Substack / your site blog

Nothing exotic. The point is the system.

Step 3: Let AI do the repurposing (without losing your voice)

Once you have a transcript, you can turn it into:

• a Substack post

• 2–3 LinkedIn posts

• a short email to your list

• a web FAQ section (SEO)

• 3–5 short-form video clips (if you want)

This is where LLMs shine—not as a “writer,” but as a packager. I use ChatGPT 5.2 Thinking and found it has really excelled with this.

Prompts that actually work

Paste your transcript and try:

Prompt A: The Substack post
“Turn this transcript into a 700–1,000 word newsletter post. Keep my tone. Use short paragraphs and clear headings. Preserve any strong opinions. Add a ‘Do this now in 20 minutes’ section at the end.”

Prompt B: LinkedIn posts (that don’t sound like AI)
“Create 3 LinkedIn posts from this transcript. Each should start with a strong first line. No hashtags. No hype. Keep it practical and specific. Include one concrete example per post.”

Prompt C: Short clips
“Pull the 8 strongest ‘clip moments’ from this transcript. For each: give a title, a 10–20 second excerpt, and why it will hold attention.”

Quick AI Tip (voice preservation)

Add this line to every prompt:
“Do not use generic marketing language. Keep it blunt, specific, and human.”
It improves output instantly.

Why this works (even if you hate “content”)

1. You talk faster than you type.

2. You stay in your lane (expertise), and let AI handle formatting.

3. Long-form creates depth, and depth creates trust.

4. Transcripts create SEO (and increasingly “AI search” visibility) without you becoming a full-time writer.

Do this now (20 minutes)

1. Open your phone camera (or voice memos).

2. Answer this question for 10 minutes:
   “What do people in my industry get wrong—and what should they do instead?”

3. Upload to Descript. Export the transcript.

4. Paste into your LLM and run Prompt A.

You’ll have a real draft today—even if you’ve never “created content” before.

A quick ask: share this with a friend

If this made you think of a colleague, a business owner friend, or the person in your company who’s been told “we need content” but doesn’t know where to start—forward this to them.

Share

That’s the whole point of infrequent AI: practical leverage, without the noise.

Hi—I’m Thomas Berolzheimer. I help run a lean team behind a social media business, and we’ve always had to win with an intense focus on our brand and our content. AI showed up at the perfect time: it’s helped us add real leverage without adding headcount—moving faster, staying sharper, and automating the busywork that used to steal hours.

I’ve also spent countless hours testing the newest models, tools, and workflows—so you don’t have to. And after hundreds of conversations with operators, business owners, decision-makers, and creatives, I’ve noticed a clear split:

There are the people who dove in—curiosity first—and found real leverage.

And then there’s everyone else.

They know AI is a major shift, but they haven’t fully jumped in yet. Maybe the first experience was underwhelming. Maybe it felt confusing or gimmicky. Maybe they just haven’t found the right application that fits their day-to-day work.

That’s who infrequent AI is for.

We don’t cover breaking AI news. We skip the noise—launches, hot takes, jargon—and send only when something is truly worth your attention: the few changes that can materially save time, make money, improve how you operate, or reduce real risk.

Our rule is simple: if it doesn’t change decisions, it doesn’t get shared.

We focus 80% on opportunities—practical, low-friction ways to build leverage—and 20% on threats that actually matter: the risks, traps, and quiet shifts that can cost you if you ignore them.

No hype. No doom. No daily noise.

Just the memo you’d want from a sharp friend who already did the homework.

If it hits your inbox, it matters.

Subscribe now