My Usage Of LLMs
Learn when and how I use AI during my workflow.
- AI
- Workflow
- Notes
AI is the talk of everyone nowadays, so much so that it starts to feel annoying. But it is hard to ignore.
Do I use AI? The short answer is yes; the longer answer is also yes, but a bit more complicated.
My first approach
When I began to learn to code, I learnt it the somewhat traditional way, through online sources such as videos, books and documentation. I didn’t like using any sort of LLMs, because during those times they were often wrong and unreliable. It was only two years into my career that I used an LLM, ChatGPT in this case, to help me port Kotsoft’s Particle-based Viscoelastic Fluid Simulation into C#, the language I was learning and the one I was most confident in. I had little experience with JavaScript then.
Incorporating it to my workflow
Since then, I’ve used it more often. It didn’t do a very good job back then. it was buggy and unreliable, but LLMs as a whole have improved a lot since. I faced some problems when I further experimented with them. They often wrote difficult to read, unorganized, buggy messes of code, and relying on them limited how much I could innovate and improve upon my code. Since then, I’ve settled on exactly where and how they can be used for a positive impact on a project, which I explain below.
How I utilize them now
Today I mostly just use Claude, for the following purposes:
Planning and researching
One thing that really speeds up the development process is being able to research a practical topic or problem and get a helpful answer immediately. That answer can even be used to do more precise research on the topic on the internet. The other is having a template to work upon for a particular project, such as a project scaffold or boilerplate, so you can quickly get to work.
Designing and visualizing
On the frontend side, I find working with CSS the traditional way too time-consuming and, frankly, boring. More often than not, I use Claude for the CSS and elements. I make the color palette and theme, lay the foundations, give the ideas, and describe the kind of design I would like to experiment with, and through multiple iterations and many, many tweaks, get the final page done. In short, I direct the design and Claude executes it. It is also sometimes used for helping with JavaScript components such as the homepage animation sequence of this website.
Debugging and general optimization
Sometimes finding those bugs becomes a real nuisance. Claude comes in handy by writing tests in bulk. It is also useful for writing benchmarks, which I run myself, to find performance bottlenecks, and for proposing somewhat surface-level optimizations for simpler parts of the code.
Code cleanup, documentation, and minor questions
They can really help with skipping the most boring parts of a project. I use Claude for cleaning the codebase and documenting it for future me or others. Everything is still human-reviewed, of course. I read every diff, and necessary adjustments are made to the documentation. The other usage is finding the answer to a specific minor problem quickly. for example, I learnt Git and Linux commands by asking: How do I do x with Git/Linux?
Where I don’t use them
Whenever the code is for performance or security critical applications, or when innovation is required, everything is hand-written. The same goes for anywhere else where reliability and correctness are crucial and a bug cannot be afforded. Claude may then be used for testing and benchmarking them. Examples of this are most of the Vofluid project’s core logic and most of the backend of the Kazheh project.
If something is not clear, or you have a particular question about how I use AI, you can contact me.