AI Coding Tools I Use to Ship Faster - Q4 2025
Back in June I wrote about the AI coding tools that I was using to be productive. Fast forward through Q3, a lot has changed in the market. Newer tools and shinier LLMs, sometimes both in the same week.
I'm still using these tools very regularly. Success is still very mixed, but I'm still quite optimistic about things. Even with a sea of options out there, I'm finding that I still only gravitate to a couple of tools.
After dropping a few thoughts on the matter, I even stopped using one of the first generative AI coding tools I ever used.
Successful projects
Using these tools regularly means I'm using them on just about every project I work on. The most successful projects thus far have either been greenfield projects, small bug fixes, and less intense refactoring efforts.
Greenfield projects tend to be where generative AI shines. There's little to no context, so the robots can just do their thing.
A workflow we tried recently was taking a project that a non-engineer put together in Lovable, and getting it battle-ready for production. Things worked out surprisingly well, with most of the engineering effort going towards security hardening, and wiring up the more advanced parts of the backend.
Less successful projects
After shipping a fairly massive feature, produced near-exclusively by generative AI, it was time to iterate. Even though the robots put the code together, and we made sure things were well documented, the refactoring efforts were quite daunting.
I'm not going to put all the blame on the robots, as a lot of the slowdowns were related to design decisions in the app. Not just that, but for the sake of answering the question "can we even get a feature to production this way?" I purposefully floored it.
With the wheels barely staying on, we got the feature to production, but the time saved at the beginning of the project ultimately ended up getting tacked onto the end of the project, playing right into Tesler's Law.
What I'm using
I'm happy to report that with the release of cursor-agent, I have the tools
available to let me work exactly how I want to... from the command-line.
One of my biggest apprehensions about jumping on agentic coding is that there was a time when you needed to abandon your favorite editor. This is great if you already used VS Code, as that's what a lot of tools are built on. Not so much if you're a Vim/Neovim user, like myself.
Cursor
Cursor is still very much in my toolbox, but I couldn't tell you the last time I
opened up the actual IDE. cursor-agent has been more than sufficient, even
after I blew through my monthly budget in a few days before they let you use
auto from the CLI.
Claude Code
While Cursor continues to be the most cost-effective solution, it still falls short on some projects. When it does, Claude Code usually can swoop in and save the day, at 5 to 10 bucks a pop.
Even though Cursor has the same Anthropic models available, Claude Code is just a better tool in my opinion. It follows direction better, and tends to be better at talking through a problem rather than defaulting to brute-force efforts.
What I stopped using
Having moved all of my personal projects off GitHub, it was only a matter of time before I stopped using GitHub Copilot. Having moved most of my initial coding efforts to an agent, by the time I open a file, I actually don't want a robot giving any input. By that point, I've already exhausted what I can coax out of the agent and would much rather take the wheel and steer.
How I'm adapting
As the tools are evolving, so am I. Early on, I felt like ensuring the robots had the right overall context was the most important thing. At this point, I'm even more bullish on the idea that scoping of tasks, not context, is the way to win with the current landscape of tools.
Having been through a full feature, to production with users, and then refactoring efforts, I've learned a lot of lessons. The biggest one is that you need to slow the fuck down, if you want to speed up. The tools are getting better, so it's not as much of a hassle to coax better than okay code out of it.
I'm slowing my roll and reviewing code even more frequently, and not settling for as much slop as I used to let slide. Until that particular game loop gets eliminated, I'm going to continue to drag my feet on fully unassisted coding agents.