Creator of Tailwind CSS, Adam Wathan, tried using AI to help design some components for the first time. But he found that Claude Code slowed him down significantly. Here's the scoop.

On Monday, August 11th, 2025, I received an email newsletter from Tailwind CSS creator Adam Wathan, about the launch of dark mode support for all 600+ Tailwind Plus blocks in the component library.

Like many of us, Adam was intimidated by the overwhelming nature of the work: needing to revisit every single one of each of 600 components to add dark mode seemed like it was more trouble than it was worth.

Adam wrote in the newsletter, "Now, with LLMs, though, we thought we could probably build some tooling and tune some prompts to automate a ton of this work and actually get it done!"

Unfortunately, things didn't work out as expected.

Claude Code "AI" Turned Out To Be a Letdown

Daniel Hollick ("Dan") is a design engineer at Tailwind CSS, and he and Adam went through several stages of using Claude.

They began by iterating through a rules.md file, instructing the LLM in how to do the process.

That wasn't enough, so they created an internal GUI tool for converting and previewing each Tailwind Plus block to add dark mode.

It was a lot of work.

(It doesn't say exactly in the email whether the GUI was built on top of an LLM, but that was the sense I got from the newsletter.)

Adam wrote, "But even with all this effort, the results were honestly so inconsistent that I'm absolutely certain that reviewing + polishing all of the work done by the robots took more time than if we just did every single block by hand in the first place. 😆R&quo;

While Adam doesn't share anything beyond a partial screenshot of the rules.md file, it seems that he and Dan found the LLM (in this case, Claude Code) slowed them down more than it sped them up.

Wrapping Up: LLM Slowed Adam & Dan Down 50%

I wasn't especially surprised to read that Adam and Dan had so many problems using LLMs for this project.

After all, they were performing a task that demands 100% accuracy. And that's always going to be a tough fit for the stochastic (non-deterministic) nature of the output of LLMs.

I'd also be curious as to which prompts and models Adam and Dan tried. In my experience, those dramatically change the results you get.

I also found that I personally didn't see any coding speed increases with LLMs — except on pure greenfield (start-from-scratch) coding work — until I did a significant amount of context engineering.

Even then, I was really surprised to read how negative Adam's review of using Claude Code for this project was.

He concludes by saying, "We signed up for this project thinking that with AI we could do it in maybe about six weeks, and it turned out to be more like about 12 weeks. Way more time than I could ever justify on this!"

I take that to mean that they anticipated that the project would take 6 weeks without AI as well and that they were just hoping it would be easier (maybe less cognitively demanding or potentially quicker) using an LLM.

Instead, they found that the project took them 12 weeks, double their original estimate of 6 weeks and slower than their subjective estimate of their velocity without LLMs at all.

So what do you think?

Is a 50% speed decrease in your productivity what you found in using LLMs for your professional software development?

Or are you finding that you're much faster?

Personally, I found that I am much faster with certain well-defined tasks after significant context engineering.

But even then, I still end up wasting time due to gaslighting (confident hallucinations) that leave me wishing I had just done whatever work the LLM got wrong myself correctly by hand instead of delegating it.

So yeah, I'll be looking forward to reading your thoughts in the responses below. Especially given that OpenAI recommends Tailwind CSS as a recommended frontend language, such as in their GPT-5 cookbook.

Happy coding! 🐎🏇🎠