Agentic Coding Test Drive

Last post I talked about what I'm building and the tech I picked. Here's what it actually was like when I sat down to try it out and get some work done.

I've settled on Claude Code. Cursor is the other big name out there. It's an AI-powered IDE and it's gotten a lot more capable recently — agent mode, MCP support, multi-file edits. But Claude Code runs in my terminal as a standalone agent, and that workflow clicked for me. The memory system especially clicked for me, and creating markdown files felt natural the way Claude guided me. More on that later.

Another big reason is that Claude is more geared towards my professional development at work. It's what we're moving toward and I already mentioned in The Stack that I favor that heavily.

All that said, I think the biggest reason I chose Claude is that all my friends use it. That's the real reason we gravitate towards a lot of technologies, isn't it after all?

I'm not sure what I expected at the time, but my expectations were about to be broken. Amaze me, Claude!

I felt like I wanted to do a test run, rather than begin a real project. Something small enough to get something working in a day but complex enough to teach me things. I picked a Texas Hold'em poker application thinking we could hammer out game logic, betting, and maybe an opponent. I was thinking more of backend logic I could verify easily.

I started out describing what the app should do and planning it out. Claude suggested TypeScript, React Native with Expo and Zustand for state. Claude Code built the core game logic first by starting with the deck, hand evaluator, game engine and betting in pure TypeScript with unit tests to validate all of it. There was no UI at this point, and this probably took less than thirty minutes total.

At this point Claude assured me that the application was working after it took a few passes to get the tests running and passing. I decided to just trust it for now and move on.

So now I had it build out the React Native layer on top so I could see the UI and test it. It built out the components, screens, and state management quickly. I pulled it up on Expo Go on my phone (Expo Go is an amazing tool for this) and I could sort of play poker as it was still quite buggy. But it was a real game running on my phone, and this was less than one hour total. One hour.

A few more iterations on the bugs for another hour and it was much more playable. The AI opponents still sucked as they just randomly chose actions, but I was pretty surprised at the progress in such a tiny timeframe. I'd say it was pretty surreal.

That was the build. Here's what made it work.

One of the first things I learned is that the context was everything. Claude reads markdown to understand what it's working on, what decisions were made, and most importantly how you want things done. I started with a few simple instructions but it grew quickly into a web of directories and a pretty complex system. Project documentation, decision logs, notes, rules.

If you're good at systems thinking, this part will come naturally and will even be a bit of fun at times. If you are the type that enjoys mentoring new developers or leading a team, working with Claude will be a blast. Setting up the system that allows someone else to succeed with code is a crucial skill to have. Claude was very good at helping this process along, automating even the organization of all of the markdown.

The better the context I set up, the better the output. I found this to be a general principle. I invested more into this since then, and it's paying off. Most of the journey so far has just been learning new skills and tweaking them.

It's been a bit of a jarring experience, but I think adapting to change always is. My mind's changed a lot since this experience, and so has my mindset around software development.

VS Code seems like it's much less important than I thought. It's open now, but it's used as a viewer more than I would have expected. Other tooling is more important. I picked up some MCP servers. Serena for language server protocol to stop the blind grepping around the codebase, Sequential Thinking for complex logic. I didn't use them much, and still don't. I need to pickup Playwright MCP for sure though. Claude seems better with CLI tooling in general.

I need to get better at breaking down work and decomposing it into sub-projects. If I'm not building the feature or system, I need to know how to break what I want into pieces the agent can execute on successfully. This probably would make me better in general with or without an agent.

One of the best takeaways is how much fun it is. I mean I started as a hobbyist more than 30 years ago, and I don't remember that being this much fun. Staying flexible and curious, being able to adapt, and just trying stuff has been amazing.

The barrier to trying something new has never been lower. I hope we all start building cool stuff.