Overview

Recently I’ve been experimenting with Claude/Copilot with the mindset that it’s an accelerator and not a magic tool. I’ve asked it to help me create features for both personal and work projects and it’s been a hit and miss at times.

One thing that’s difficult and time consuming for me to do is provide enough context to the AI agent so it would create it as if I would create it. This takes a lot of time when the features themselves are complicated. On reflection, with simple features, there’s really no wrong way of doing things and I can definitely ask the agent to help with that. However, when it comes to more complex features, it can get things wrong and there are times I feel I can do a better job and then not have to worry about getting the prompt right.

This got me wondering, maybe I can use it to learn instead of asking it to fix things that I could do by myself?

Can I figure out how agents work with Claude?

I’ve always wanted to learn more about the inner workings of an agent and how the LLM, prompts, and everything ties together.

Well, I have a Claude subscription and I don’t max it out currently. The new design feature also just came out a few days ago. Time to use Claude to accelerate my learning

I started off with a simple prompt to help me understand how agents and LLMs work using animation. That created a simple slide deck with animation, which turned out to be pretty good. To be frank though, while it was pretty and filled with animation, I am not a designer nor a front end developer. Anything short of simple HTML is already guaranteed to wow me.

I ended up burning my tokens because Claude design is quite expensive. But, I’m planning to keep iterating over the deck and adding more information to see if I can build it out to something.

It turned out to be decent: How LLM works

Reflection

Reflecting on the times I’ve asked an agent to help me implement a feature, I’ve been more aware of its failings if I know what needs to be done. With .NET and C#, I can easily point out times where the agent decides to use something unorthodox.

A simple example is where I asked an agent to add a simple feature, but instead of using a List, it chose to use some weird AsyncLocal<List> as it got worried about some threading issue that would never happen.

If I wasn’t a .NET developer, I probably wouldn’t pick that up. It does leave me wondering whether I’m also missing things in a FE project or slide deck that someone more experienced would pick up. If AI did something unorthodox in a react project that I asked to create, I probably wouldn’t pick it up right away.

I think that’s one mindset that I need to continue. A project created by AI might look good in the surface, but underneath it all, it could be quite messy unless you understand it properly.