How I Learned to Build With AI

+ my AI tech stack 🥞

So many of you replied to my last post asking what I’m actually using, how I learned it, and how it all works. This is that post.

---

It started at Startup Striders.

A Saturday morning run. 🏃‍♂️ Jeff was raving about all the crazy things he’d been doing with Claude Code - dashboards on his phone, agents running in the background, shipping seven different projects at the same time.

He mentioned to swing by the office on Monday, and he’d show me the stuff.

I cleared my Monday. 💯

Two days later, I walked into Jeff’s office. It was just him and a mate. They were talking to their laptops. Not typing - talking. Wispr Flow dictating their prompts, Claude Code executing them, seven terminal windows open, jumping between projects like it was nothing. 👀

And I was sitting there. Typing. In my inbox. Like a monkey. 🐒

My mind was blown.

It was one of those moments where you go, “Ok, this is crazy and a whole new world.”

Jeff looked over and said, “Let’s get you set up on Claude Code.”

That was February 8th, 2026 - 9 weeks ago.

---

🔌 Week 1: Connect Everything, Build Your First Thing

My first Claude Codemessage was sent at 9:53am that day.

I spent the morning setting up MCP servers - Notion, GitHub, Airtable, Slack, Todoist, Gmail, Stripe, Bear Notes, Substack. MCP servers are basically how you give AI arms and legs to reach directly into your tools. No more copy-pasting between windows.

Then I said something like:

“Design me a website - batko.ai - website to coach founders. Here’s my LinkedIn. Here’s Substack for tone of voice and writing style.”

batko.ai was live in production by 6:44pm.

Not a template.
A fully designed, deployed, mobile-responsive, SEO-optimised website.

Built through conversation.

I said things like “make the whole site sound more like James Clear and Seth Godin” and “pull my top LinkedIn testimonials as quotes.” It just... did it.

Day 2, I started loading in my own content - 7 years of LinkedIn posts, Medium articles, podcast transcripts, Bear notes. I asked Claude to pull out my strongest frameworks and recurring themes across everything I’d ever written. It answered from my own words - quotes I’d forgotten, patterns I didn’t know I had. It stopped being a generic AI and started being *my* AI.

By Day 4, I discovered CLAUDE.md - a file where you give Claude persistent instructions about who you are, how you like things done, and what never to do. Think of it like hiring an assistant and handing them a one-page brief on day one. Every session after that started smarter.

Same day, I built a full Airtable system for a coaching client - properties, jobs, client records, checklist templates, category structures, all the right integrations wired up. The kind of thing that would have taken a junior analyst probably 40 hours full-time to set up properly. Done in a couple of hours through conversation.

By end of week 1: 665 messages. 10+ tool integrations. 1 production website. 2 side projects. Client systems built for a real business. Zero lines of code written by hand.

🚀 Week 2: Ideas Become Real in Hours

This is where things got wild.

GroceryBotko - just for fun, for me and my wife. An app where you upload a grocery list screenshot and AI parses the items, searches Woolworths, and adds them to your cart. From idea to working app in an afternoon.

Why? Honestly, my answer at the time: WHY NOT?! It was just fun ❤️

I built PitchMaster - an AI pitch deck analyser that scores your deck the same way I’ve evaluated hundreds of founder pitches. A co-founder matching platform. And started adding PostHog analytics, dynamic OG images, and a proper URL architecture to batko.ai. All in the same week. All through conversation.

Then my biggest time saver - I connected Granola (my AI meeting notes tool) to Claude Code. Now after every founder call, it automatically drafts a follow-up email. And from the pain points we discussed in the meeting, it starts a blog post suggestion that plugs straight into the Batko AI brain - my personal knowledge base of 11,000+ pieces of content - check them all out here.

That’s when it stopped being a building tool and started being an operating system. The shift from “I ask it to build things” to “it runs things for me in the background.”

🏗️ Weeks 3-4: The System Manages Itself

By week 3, I had so many things running that even I couldn’t keep track. So I literally had Claude summarise itself - every agent, every automation, every feedback loop - and published it at batko.ai/about/batko-ai-os.

31 active AI systems. 23 autonomous agents. 6 self-reinforcing feedback loops. Zero manual steps before morning coffee.

• A daily Chief of Staff agent prioritises my tasks.

• A Meeting Briefing agent researches attendees before every call.

• A Health Watchdog monitors whether all the other agents are healthy. The system watches itself.

I went through four distinct phases:

• Week 1 (Builder): Connecting tools, building through conversation. 0 agents.

• Week 2 (Builder+): Multiple products at once, first scheduled automations. 3 agents.

• Week 3 (Operator): Polishing, SEO, complex multi-source automations. 6 agents.

• Week 4 (Architect): Systems that build and monitor other systems. 12+ agents.

One of the most useful things I built along the way: tools that tell me when the work *isn’t* happening. Not just tools that do the work. My newsletter schedule warns me 7 days before a slot goes empty. My Monday standup flags when a number drops week-over-week. Feedback loops, not just automations.

🧰 The Stack

People always ask about costs. Here’s the honest breakdown:

• Claude Code ($300/month) - I can write code without reading a single line of code. That’s the whole pitch. You talk, it builds. and I’m building way too much so I’m on the Max plan, you probs don’t need to spend quite as much.

• Wispr Flow - voice dictation everywhere on Mac. Barely touch the keyboard.

• Granola - AI meeting notes that trigger automations

• MCP Servers - giving your AI arms and legs to reach into Notion, Gmail, Airtable, Slack, everything

• Vercel ($20/month) - website hosting and deployment

• HubSpot (free CRM) - pipeline tracking and client management

Total: ~$320/month.

For that, I built and run batko.ai (54 blog posts, 48,000+ words, 8,900+ visits), a client dashboard, multiple products, and 23 autonomous agents. Less than what most businesses pay for a single SaaS tool.

Claude can write whole strategies with you too. I wrote my entire product strategy through a back-and-forth conversation. Not a one-shot prompt. A real collaborative process where it challenged my assumptions and I challenged its suggestions.

🟠 Speaking of tech stacks - tools alone won’t build your GTM.

What really moves the needle is learning from peers and seeing what’s working right now in your region.

That’s why I’d recommend signing up for HubSpot’s upcoming GTM webinar.

It’s built for founders who want to create a GTM engine that actually scales.

The best part?

It unlocks access to a virtual bootcamp for APAC founders, with all the support you need to build.

Applications for the bootcamp close on April 24, so don’t miss it

-it’s totally FREE, and you’ll have a working GTM engine in just two weeks.

😅 The Mistakes

I’m defs still getting things wrong. Here’s what I wish someone told me:

I had no idea what git was. I was operating on the main branch with seven different builds happening at the same time. I’d build something, start building something else, and the first thing would just... disappear. Overwritten. Gone. Then I’d rebuild it. Then lose it again. I’m not a developer - I had no concept of branches, commits, or pull requests. I’m only just now learning to do this properly, 10 weeks in.

I let Claude run unsupervised for 12 hours. I asked it to build me a full Notion page with every step required to get an Austrian passport for my son. It went into loops on Austrian government websites overnight - 12 hours of churning, going in circles, parsing bureaucratic pages. First time I’d seen it properly stuck. Lesson: check on it. Set boundaries. Don’t let it run all night on government bureaucracy (the irony).

I didn’t set up error alerts early enough. When you’ve got 23 agents running, things break constantly. And I was way too slow finding out. Now every error gets pushed to my email or Telegram instantly so I can investigate. If you’re running more than 3 automations, wire up alerts on day one - not month two.

🧠 The Mindset Shift

You’re not learning to code. You’re learning to direct an AI that codes.

Think of it like being a film director - you don’t operate the camera, but you need to know what you want the shot to look like.

The 4-week arc:

• Week 1: ”Holy shit, it can build a website from a sentence”

• Week 2: ”Wait, it can build 5 things at once?”

• Week 3: ”It can run things for me while I sleep?”

• Week 4: ”The system is managing itself. I just steer.”

Today is the worst you’ll ever be.

I’m 10 weeks in now and I look back at week 1 me and cringe.

But week 1 me shipped a production website in 6 hours having never written code before.

So how bad could it have been?

---

🔥 If you want to learn how to do this yourself - not theory, hands-on, building your own project

- I’m running a course on exactly this:

>> Puddle Pod for AI is back and cracked :)

(website in production 😅 I’m building it as I learn + ran out of tokens yesterday 😭)

lmk if you have questions - just hit reply :)