As someone that’s been working in the tech industry for quite awhile, I have mixed feelings about the introduction of AI into our workflows. It started so innocuously with the introduction of more useful tab completion when writing a line of code, to a tool that could write entire functions or modules based on a comment in a file, to “ooh, I can use this tool to write an entire codebase from a plan that I create with some back-and-forth with an agent!”. It’s been quite a couple of years.
And here we are now, with some people having full doomsday predictions, saying entire industries will be wiped out by AI at some point. I cannot predict the future, and I think we are still quite a bit away from that point, with a lot of unknowns (and a lot of AI slop being generated).
With all that said, today I want to call out some areas where LLMs have been actually helpful in my day-to-day work:
Migrations
I used just a couple of chats with Claude Opus 5.5 last month to migrate my entire blog and my markdown-based blog posts from Gatsby (ancient at this point), styled-components (CSS-in-JS was the new shiny thing for a long time!), and an old version of React to a fully modern stack based on modern React, Astro and scoped styles. If I had to do all that work by hand, it would have taken likely over a few weeks of weekend dabbling to get all the pieces in place, which is why it was on the back burner for over 2 years 🙂. The styling migration away from CSS-in-JS to Astro’s default of scoped styles alone would have been a lot of work.
This removes most of the friction from changing stacks entirely, helping people to easily move off of old or deprecated tech stacks and frameworks, many of which have security vulnerabilities along with code rot.
Code Reviews
Yes, the world of agentic development means engineers are often reviewing a lot more code, since the general pace of projects tends to be faster, and the timelines shorter. Where LLMs have become very useful for me is in doing an initial code review, either just prior or just after a pull request is opened. Based on the AGENTS.md or whatever other rules are set in the repo, agents can call out all the foundational issues with a pull request up front, saving time for human code reviewers. Human reviewers can then focus on the subsequent reviews, flagging any issues that agents missed that may be more nuanced.
Accessibility
This one is huge. It’s hard to get accessibility right in the first place, no less keep accessibility best practices in play as component libraries and applications evolve. It’s very useful to have agents who have been fed in accessibility guidance in the form of a skill, rules, etc, to comb through code changes in a pull request and flag any issues.
Not every team has access to accessibility experts for reviews, so having at least a first check done by agents is a big win.
Debugging
Debugging code can be very tedious. I’ve found that LLMs are an excellent debugging partner, helping to come up with ideas for where to start when looking at a bug, to creating entire separate code modules or components in an isolated environment to help recreate an issue, to doing live in-browser debugging using browser tools like the Chrome devtools MCP server.
This makes the whole debugging process often shorter and less frustrating, even for experienced developers.
Idea Generation and Exploration
When the cost of creating and testing a bunch of different approaches for something like a new component API, or even the color scheme of my blog, is so low (well, for now, at least), LLMs make me feel like, “hey, why not — let’s explore.” In the past, I often had to come up with a whole bunch of ideas or have a brainstorming session with coworkers, but now LLMs can make the initial brainstorming feel less overwhelming.
I’m not saying at all the ideas are good, but that’s the whole point of brainstorming! I’ve also found LLMs useful for generating ideas for naming schemes of things like design system motion tokens. I find it’s much better than just starting fully from scratch.
I also find LLMs useful when first exploring newer tech, like new CSS features that I haven’t used. While I love reading blog posts and sites like MDN and CSS-Tricks, sometimes it’s nice to use an agent to help me build a first demo using a new feature, like sibling-index. Side note: CodePen’s AI integration, a pro feature, has been fun to test out for this. Then I can dig in and explore the edges of it, building on top of it, versus having to start from scratch every time.
Featured image by Brad Switzer via Unsplash
