LLMs Are Not Necessarily About Speed

Written at 2026-02-11

Tags: Ai

Around 1.5 years ago, I developed a web extension called Mark Scroll Positions both to solve my own need and to experiment. If you are interested in how it looked and the original idea, you can check Introducing: Mark Scroll Positions. But long story short, even though I did not have high expectations for it, other people also started using it after I published it. I received some comments on both Firefox and Chrome. A few users even went to GitHub and opened issues.

Most of the issues were simple problems that I could have fixed if I just had the time, energy and motivation for it. For example, adding a dark theme option or fixing the drag and drop feature that was not working properly in Firefox. However, I was a bit tired at that time and focused on other things, so I could not find the motivation to handle these requests. I left the repository as it was.

Recently, I spent a some time with my friend Cem. He showed me how he uses different agents and models in his flow. My perspective changed a lot. Seeing a good developer friend of mine working very efficiently with different agents in front of me was something I could not ignore.

So, I decided to try using LLMs on the extension that I had already stopped developing. I thought it would be better to offload some of the trivial tasks to an AI agent instead of not solving them at all.

Over three nights, I was able to improve the repository significantly. I migrated the project from Parcel to Vite and moved from raw CSS to Tailwind (because I thought AI would work more easily with it, and I also like Tailwind). I removed the SVG icons and started using the Font Awesome library. I also improved the overall look of the UI through deliberate design decisions. I had to take most of the initiative on the design because Claude did not implement things exactly how I wanted at first. But overall, I was able to try different design choices without the feeling of “churn” that I probably would feel if I had been writing all the frontend code by hand. I added dark and light theme support, fixed drag-and-drop problems in Firefox popups, added a “sort by latest changes” feature and a settings page, and fixed the broken auto-jump feature, where the extension would open the site but fail to scroll to the correct position. If you are curious about the difference, you can watch the previous version in this video and explore the latest release on the Firefox or Chrome stores.

Now, I still think that AI might be slower than an experienced developer who is fully focused and writing everything manually. But I think this way of thinking misses the main point. I think that for many people, the main benefit of using AI agents is not just speed, but being able to offload tasks that you don’t have the energy to do (especially those that require manual labor rather than deep/architectural thinking). Developer velocity alone does not mean much if you get exhausted and stop. It does not matter if you start fast but then get stuck on boring tasks and lose motivation. With AI, I feel that this mental barrier becomes lower. That is very valuable to me, even when it does not make me faster, although in many cases it does.