ChatGPT Just Became Garbage

I want to start this post by stating no AI was used in the creation of this blog post (I did use Grammarly, but I'm not counting that one). 

Now, ChatGPT was a solid go-to for many tasks: python scripts, ingesting PDFs, giving me BLUFs, and basically making me habitually more lazy and reliant on an AI assistant to do various tasks. I’m not saying to walk away from AI, but I’ve come to a realization that there is a correct way to leverage these new tools and an incorrect way. 

My personal problems aside, I first became interested in productive prototyping after a few of our engineers made this interface to manipulate C++ code for Unreal Engine's physics simulation using as dumb a language as possible. Here is the video if you want to see what I mean. 

So a couple of months roll by and I’m tracking dumb new developments from ChatGPT like Operator. I found one useful application for it: 

“Find me a site online that sells ZYN in all their flavors and have it in stock”. 

To my disappointment, Operator wouldn’t do this task for me since ZYN is a regulated product. This makes no sense to me because I could spend hours and hours doing this myself with no legal consequences. In fact, this is the stupid crap an AI should be doing for me. It’s also interesting that I can give my passwords to Operator, and that’s kosher, but some open-source intel on where I can secure my nicotine is off-limits. 

Since I did still have things to do (believe it or not), I decided to give Bolt a try. This arose from a customer request for a prototype dashboard application. I’ve managed web development projects in the past so my understanding of terminology and components of modern web development and design is not awful. I also really like Stackblitz, the company behind Webcontainers. 

Bolt was amazing. I was able to exceed my own expectations in coding a proper front-end web application and set up a pipeline for me to deploy it so the customer could see it as it progressed. The quality of the code was good, and the outputs were predictable, so I had hope. I even began our company’s website as a Bolt project before exporting that code and managing it in-house. 

It was not until Anthropic’s Claude 3.7 hit my radar that I began to see the real potential of AI to explore the creative side of development – the side I am truly not skilled enough to typically explore. A colleague of mine recommended I try a VSCode extension called Roo Code. Now, I'll be honest: the trustworthiness of this extension is by no means vetted, but in the name of experimentation, I decided to see what it could do with my Dashboard application and the Claude 3.7 API provided by Anthropic. The results were incredible. Previous misunderstandings and errors that would not get resolved using Bolt were suddenly child's play for this local agent and Claude 3.7. I was able to expand that prototype into something robust, and it is now the framework for an application we intend to build for production.

Roo Code will also run headless Chrome to validate the UI changes it makes before convincing itself that it is finished. Something that was incredibly useful in debugging nuanced components of my front end. It sits right in the IDE and looks something like this:

It will also track the cost of my session so that I don't end up spending a fortune going down a fruitless rabbit hole. As you can see, this entire feature, which is now functional and part of the application, costs about $10. That's an hour of a potentially decent web developer's time from a place like Ukraine or India, although I haven't tapped those markets in some time, so the rates may be drastically different these days.

If you visit the link on Claude 3.7 above, you will notice that they released thecli application called Claude Code. If you haven't deciphered that I am a naturally curious and experimental person, this was something that I was destined to install and play with. So I did. I recommend anyone who wants to explore the utility of this model install it by running the following in your terminals (requires npm):

npm install -g @anthropic-ai/claude-code

You should also read the docs on Claude Code.

My First Run with Claude Code

Last night, I was watching the LA Kings continue their 4 game losing streak while casually interacting with Claude Code to get a sense of what it could do.

Background: I love terminal pastebins. termbin.com has been a go-to for some time; https://transfer.sh was good. The problem was that I never took the time to write the functions to make using transfer.sh useful without remembering all the curl commands to send files and text to it. Instead, I wanted to host my own, but I wanted to use Cloudflare Workers and Cloudflare R2 for object storage to store the pastes. So I had Claude Code give it a whirl. I gave it some features I wanted and It nailed it on the first go.

This resulted in me deciding to open-source the project, which I dubbed dedpaste. Within a few minutes of sharing this with some colleagues, one of our engineers recommended we add end-to-end encryption for pastes. That piqued the interest of another one of our engineers, and before the hockey game was over, we had the feature implemented, the package published to npmjs.com, and the repo open for contributions.

Bringing this thought to a close, what stands out to me is the level of collaboration that is enabled by starting with a quality foundation. OpenAI can't provide that in my experience, but Anthropic is accelerating creative teams with diverse talents to try new things, think about utility, and focus on features rather than inundating them with the burden of maintaining their ideas.

If you want to try out dedpaste you can install it with npm

npm install -g dedpaste

or you can clone it from my repo.

Check out Keerthi's upcoming post on using dedpaste and some of its features.