One Year of Building With AI

Introduction

This is How We Grow, where we write up what we're doing to grow Kai. This week was a bit different. Jim started at Morgen on 13 October 2025, and Kai launches on Product Hunt on 13 October 2026. So before launch week takes over, we looked back at one year of building with AI: the moments that changed how we work, the models that kept raising the bar, and what's left of a growth job when the work gets this fast.

Prefer to watch or listen? Here's the full episode

The first wow moments

Sonnet 4.5, and an extension in 2 hours

I opened with a quiz for Jim: which model were we using a year ago? He guessed Opus 4.1. It was Sonnet 4.5. Opus came a month later.

Jim's first real turning point came before Morgen, at Wooclap. He had a report to build from Metabase, the kind that meant pulling numbers from all over the place by hand. Perplexity had just released Comet, its agentic browser, so he asked it to find sign-ups by country in Metabase. It took over his browser, clicked around, found the right metrics and put the table together. For Jim that was the moment chatbots turned into something you could delegate to.

The second came in his first week at Morgen, when he met David in Prague. They went to dinner at 7 PM, talked growth loops, and came back with an idea: a small extension that saves anything you see on the web with a quick command, so you can come back to it later. Jim did the design with Claude, David did the infrastructure and the prompting, and within 2 hours they had a V1 working on their laptops. Two weeks later it was on Product Hunt.

Mine: Search Console, then the Morgen app

My first one was at Wooclap too, and Jim was the one who showed me. ChatGPT could now read our Google Search Console, so we asked it the boring questions, like how many people from one country land on one kind of page, broken down by month. It answered in one prompt. Before, that was a CSV export and an afternoon.

The second was at Morgen, in my first month. David told me I had good ideas for the product and asked why I wasn't building them. I asked if he meant me, building it myself. He said yes: just talk to Claude, explain what I had in mind and try. That evening I sent Marco my version of the Morgen app, and 2 weeks later my first change shipped: the referral program design, plus the calendar button moved onto the day's date.

A new model every few months

Jim's way of putting it is that every 3 or 4 months a new model comes out that raises all ships. What made each one count for us is that the system around it was already there. Our analytics, our documents and our code live in one repo, so when a better model arrives it gets to work on day one and the difference shows straight away.

Opus 4.6, February. It came out around our hackathon in Zurich, and that's when we started doing heavy product work. Jim cloned voices with ElevenLabs to build a voiced onboarding into the product.

Fable 5. Jim had tasks he'd been stuck on for days. He handed them to Fable 5 and got them back right on the first try. It's also when Lottie animations went from impossible for us to one prompt. For me it was a bit frustrating: the week it came out I had a thousand ideas and couldn't get to most of them.

Opus 5.5. Video. This episode's YouTube edit, from the cuts to the captions, was made with Claude on a pipeline I built this month.

clawdboard, Jim's usage leaderboard

At the Zurich hackathon, John showed us a terminal command that tells you how much Claude usage you have, and whether your $200 subscription pays for itself. Back then we were spending about $10 to $12 a day. Jim liked the command so much he built a leaderboard around it: clawdboard, where people who use Claude a lot compare their usage. It has a front end, user stats, syncing from each person's computer and a backend full of automations, and Opus built it in February.

Jim's own usage on clawdboard, year to date.

The chart tells the year. A few dollars a day in February, about $100 a day from the hackathon on, and now some days at $1,200. Jim's view is that clawdboard today would be one prompt with the effort turned up to max, with a cleaner codebase and fewer bugs than the February version.

How the growth job changed

From negotiating to shipping

Jim led growth at Wooclap, a scale-up, with a team of 4 to 8 people depending on when you look. He had 2 engineers, a dedicated data analyst and a product manager for activation and retention, all of them shared with other projects. So most of his job was negotiating: writing business cases to show why his project would bring in more money, so it would get engineering and data time.

That negotiation is gone. Today, as he puts it, we are the data analyst, the engineer, the PM and the growth marketer. A project now runs like this:

1. Scope and size it. Check there's a real opportunity before anything else.
2. Build it straight away, with the analytics built in at the same time.
3. Ship it that night.
4. Automate the read. The metrics land on Slack after 7 days.
5. Close it in 3 weeks if the numbers hold.

The building takes about half a day.

Seven projects at once

The other change is how much fits on one plate. A year ago Jim worked on one project at a time, 5 days on it, then the next. Now he can have 7 running in the background: one on ads, one on SEO, one on onboarding, one on a feature in the product. I'm the same, usually 4 or 5 at once.

Jim says he's never been more stimulated at work, even though he hands more of the decisions to AI. For me it's close to an addiction, and I don't know yet if that's good.

Why it works

Jim made a point I agree with. It isn't as easy as opening a chat and saying "write me an article." My SEO works because I did every step by hand for years first: the analysis, the keywords, the briefs. Each step is written down for Claude, including where to look things up and which Ahrefs data to pull, and because Claude follows instructions well, we can trust the output.

Video is the same. I've been editing my own videos for 10 years, so I knew which steps are boring and can be automated, and where I need to stay in the loop. The onboarding emails were the opposite. I'd never built an email journey, so it took me a long time, a lot of documentation and a lot of back and forth with Claude, and I'm still learning.

The other half is the knowledge base. When we started Kai, we put our docs, our website, our pipelines and our code in one repo, so we never have to explain ourselves to a model again. Jim calls it the best investment we made early on: in AI, context is king, and most people keep re-explaining themselves in a new chat every time. It's also the problem Kai solves for its users, carrying the context from your meetings, emails and tasks so you don't start from zero.

The messier parts

• We check less of the detail. Jim used to mock every landing page in Figma and write every line of copy himself. Today we keep a bird's-eye view and hand the micro-decisions to AI. Usually it's fine, but that care was worth something.
• Switching between projects costs a lot. When I come back to one of my 5 projects I have to remember where the conversation was. It needs structure, and projects close enough to each other that switching doesn't hurt.
• A pipeline without the craft behind it disappoints. If you've never done the work by hand, a model gives you something, but not something you'd ship. The emails taught me that.
• The bill grows with the habit. From $10 a day to days at $1,200, as models get better we use more of them. I've skipped Fable 5.1 so far because I didn't need it at its price.

What's next

Launch week. Jim thinks we could launch tomorrow and not much would change, which is a good sign. What we're curious about is what comes after the spike: whether we can keep finding scalable ways to bring users in and turn them into paying customers.

We'll record the next episode the morning after Product Hunt, with the results.

If you want to follow along, the whole series lives at How We Grow, and the changelog shows what we ship week to week.