Towards the end of 2024, my employer banned AI tools. All of them. Then, a few months into 2025, they did a complete 180 and mandated their use.

We’re an advertising and media company, so the focus was on generating copy and images. No one was thinking about the developers. I messed around with it a bit — personal evaluations, employee satisfaction questionnaires, even a resume — but nothing I’d call useful. Some co-workers were experimenting, but I was never impressed with what they were producing.

Then, about a month ago, we got premium access to Claude Code as an experiment. I decided to play with it, being very skeptical. I had heard that greenfield projects were where people had the most success, so I created a directory called game, cd’d into it, installed Claude, and typed init.

“What type of game would you like to create?”

Wow.

Within what seemed like minutes, I had a little car driving around a map. Not much longer after that, I had NPC cars driving around too. It was a multiplayer game, so I spun up a second browser and had two players interacting. The stack was Node.js and TypeScript. I stayed up past midnight tweaking the game, adding very rudimentary combat, agonizing for what seemed like forever getting keyboard input working just so.

I should mention: it felt like I was working with a junior co-worker, not typing into a text prompt. That feeling never really went away.

But some of the weaknesses started to show. We would fix a bug, fine-tune a feature, and the bug would come back. We would fix it again, fine-tune a different feature, and it would reappear. Early on I had given it instructions to add unit tests and end-to-end tests, and to make sure everything was committed to git before working on a new feature. It forgot these things time and again.

The next day, I continued working on the game, but this time focused on bundling and deployment. We created a GitHub Action to build a Docker image, deployed it through a local Portainer instance, and used Traefik to route external traffic to the container.

The container showed up in docker ps. It showed up in the Portainer UI. It showed up in the Traefik UI. It was even accessible on the LAN via an exposed port. But it wasn’t reachable through the public-facing URL. After about an hour of back and forth, we tracked it down — the container was on the wrong Docker network. Once that was sorted, we hit the next problem: it was being served with a self-signed certificate. The Let’s Encrypt configuration was wrong.

Both of these were caused by Claude making assumptions without asking me, and without looking at existing configs where I had already solved these exact problems. I would suggest something, Claude would suggest something. We got there eventually. But it was a preview of a pattern I’d see again.

What else can this do, I thought?

I had seen a lot of hype about AI being good at trading, and a friend had sent me a link to an algorithmic trading tool he had been working on. I made a new folder called algotrader, started Claude, and we started exploring ideas. We talked through crypto, forex, stocks, and futures. I had previous experience working at a forex firm that had an API available, so we decided to explore that.

Within 4 hours of creating the directory, we had connected to the API to download historic candles, built a backtesting engine, and constructed and tested 8 strategies. None of them were profitable in the end, but wow — that would have taken me weeks or months on my own.

But the next day, the cracks started to appear again. There were some major bugs in the backtesting engine. One strategy looked promising, so we built a paper trading engine using the broker’s practice endpoints. The actual results didn’t come close to what the backtests had predicted. Turns out there were some glaring errors in the code — fundamental problems. For example, it assumed that AUD/JPY trades were settled in USD. Something anyone with basic forex knowledge would instantly recognize as wrong. Claude had no way of knowing what it didn’t know.

This is where I had to learn how to slow it down. I had to learn to get it to question its own assumptions. I had to tell it to wait until I said I was ready before writing any code, and not just run off and make changes based on every passing comment. That meta-skill — learning how to work with it effectively — turned out to be just as important as anything else.

After enough time spent on side projects, I finally brought Claude into our company codebase. Hundreds of thousands of lines of legacy code, much of it from the early startup days when speed mattered more than maintainability. Multiple developers over many years, each with their own style. I wasn’t sure what to expect.

It’s a completely different experience from a greenfield project. We talked for about 30 minutes, but a large part of that was just orientation — pointing Claude at the right section of code, making sure it understood the context, and making sure any changes would be isolated to the new feature without breaking anything existing. That part took patience.

Once we had that foundation, the solution appeared — one that seemed obvious in retrospect. Inside an hour, start to finish, we had accomplished something I hadn’t had a clue how to approach at the start.

So, where does that leave me? It’s a useful tool. But there’s a learning curve, and you will be frustrated as you climb it. I’ve been at it for about a month, and I’m still figuring out how to get the best out of it.

What I can say is — there’s something here.

I don’t know exactly what yet. But I will keep coming back.

This post was drafted with the help of Claude.