Though I had dabbled with various AI agents like Google Gemini and ChatGPT I had yet to give a fully fledged agent like Claude a good try. Being a web developer for over 15 years, I thought it would be a great way to test its capabilities more fully in an area I was less familiar with. I've had an itch to scratch with game dev for a while so I finally downloaded the desktop client, setup a new repo and started working on my first game in Phaser.js.
Fast forward a few months, after the release of my first game, Hex Sweeper, I'm hooked on gaming and see there is no turning back with Claude Code. I've now also released the beginnings of a 4X game framework and a more complex turn based version of a DTD style game named Button Breach. Below are some thoughts on my first few months experience developing web games for the first time with Claude Code.
One big thing I find with Claude, at least for hobby programming, is that I can get a lot done in the evening after my professional work even when I'm tired. Normally, this would require a lot more will power. Don't get me wrong, I'm not auto pilot vibe coding here, I review every line of code it produces carefully and refactor when necessary, which admittedly can be quite often. I will occasionally even go manual, but I would say over 90% of the actual code is written by Claude. Most of my time is spent in prompt planning and code review once changes are made. Then instead of tediously typing away for hours at a time I can take a quick five minute break, walk around and start thinking about the next task.
I still do think about the code a lot though, and have Claude setup more as a code assist. I don't let it touch my Git repo and I have a lot of strict rules in the CLAUDE.md to try to prevent it from actually implementing any code without a key phrase like "go" or "do it" (though it still ignores these more than I would like it to). I code in chunks as I normally would and commit bits at a time. My general workflow is still task based, I work through my own thoughts via Claude summaries and feedback which may consist of quite a few corrections and alterations. When I'm satisfied I then finally implement and let Claude do its thing. Once implemented if it's working code that is at least 80% good enough, I'll commit at that point. If a bit of refactor is necessary, I find it's better to see the changes, rather than mess with what's already working code since Claude can really go haywire at times. I find it's best to be able to just revert and start with a fresh session if need be.
From the planning side of things it's great when you pretty much already know what you need to do. It can summarize the task really well, flag anything that needs attention and even ask great follow up questions for clarification. Where it tends to go nowhere is when I'm not sure about the task myself. For instance presenting a particular architectural issue I'm facing and asking for suggestions on how it may be done better. It will look at your existing code and some keywords in the prompt but not really be able to provide some out of the box thinking. It tends to just end up in a sycophantic loop contradicting what it previously said. But when it comes to implementing features and refactoring it really shines and can write an hours worth of manual coding for you in minutes.
Now how much of a boost has it actually been for me though, how many X's? I'm still not completely sure, I would say at least a 2X in terms of raw output. Rather than coding I spend more time in planning, code review and refactoring. Also, even though Claude can really frustrate me at times, I would say I have less mental exhaustion over the course of the day and week. My day is generally more relaxing since I'm not tediously typing away and tracking down typos. However, in comparison, I would say tools like Laravel, Vue/Vite are probably larger X's overall. There is already so much quality code working for you out of the box.
Another thing I noticed is that the sycophancy is inherent, it's like the code itself is sycophantic sometimes. In retrospect, as far as I understand LLMs this makes sense. AS I touched on earlier, when you have clear tasks it shines, when it's more ambiguous it can give you working code, but it can be quite sloppy. There were a few times where I ended up finding a better solution to a particular architectural issue I was facing that seemed like what should have been quite obvious suggestions for the all knowing AI to point out to me.
After a few months, even with its faults, I'm quite impressed with Claude Code. It's definitely a game changer even as it is now. There is a bit of a "feeling out curve" and the sycophancy can be an issue, but overall I've found it to be a useful tool. Considering I'm absolutely new to Phaser.js and game dev concepts, I feel I've had considerable progress.
It's definitely here to stay, and I feel it's a point of no return after getting the hang of it. My days are less mentally exhausting which is a win on its own. Though I do have to self monitor my frustration as Claude can feel so human. If I'm not paying attention, I can feel sparks of anger with some of the answers it gives me only to realize how silly it is to get emotional at the generated text it's outputting. It's just a tool and it's important to know where it shines and where its flaws are if you want to be more productive with it.