Claude vs ChatGPT

Over the last year I've been increasingly using the two leading LLMs to assist me with various tasks and experiments. I say experiments because in some cases I've done what everyone does and created some generative art, but in several cases I've used the LLM as a sounding board, help with organizing thoughts, a coder or experiments with some writing assistance. After trying both products, I've found I prefer Claude over ChatGPT, with one exception, and as newer models and new features are released, I've used Claude more and more. Overall the LLM has greatly sped up many of the tasks with my side project not related to design, although in some cases I have ceded the design.

Since Claude doesn't have the same image generating features as ChatGPT the comparison is not exactly like for like—they both write and code fairly similarly, though each chat session would reach the chat length limit at different times. I tried building several games with ChatGPT, as well as a screensaver/Mac app for displaying information widgets, but since I wasn't regularly generating art, I found Claude's features more useful.

In most cases, I simply don't know how to do something. I did land on an idea that was very much within my experience, but previously I had used Claude and ChatGPT to help with the javascript for a few smaller ideas. For example I came across Chiptune music and found a lot of the existing apps a little rough (ironic given I didn't have a design for this idea). Initially Claude created something very similar to the existing tools, but with a bit of feedback I arrived at something more appealing, ChipTune Composer. It is a work in progress—assuming I ever return to the idea—and some functionality doesn't unfortunately work (the example tracks don't currently load). Overall the UX/UI is still a bit rough, but the core functionality it there.

Recently I started building out an idea which is normally what my side projects are only this time I let Claude do the heavy lifting. In fact the original idea was broad and I first used a chat to narrow down to a simple idea—a site for sharing LEGO deals. We mapped out what was needed and began building the two sites, the front-end and the CMS to power the deals. There were a few issues throughout, but the process allowed me to build a rough MVP in less time though at the time it still wasn't ready to launch. And most importantly, I didn't wireframe out a design ahead of time—I let Claude build the initial static HTML pages as a proof of concept, and it's serviceable (it may not be the most cutting edge design, but it's a bit better than most off the shelf templates, and a lot better than similar LEGO themed sites).

I did hit a number of snags throughout the process (along with some issues with Claude fetching from the wrong source): data and the scope to complete the missing CMS components. One frustration was the Claude would regularly grab images from other sites when I had asked it not to, and more specifically I asked it to use images I had already downloaded, but I'd regularly see image URLs to third-party sites. (At the time I was stubbornly aiming to build an entire CMS from scratch so building on my local didn't require basic auth or accounts).

After a break, I decided to pivot and use Wordpress as the CMS. (Wordpress has in the past been my CMS of choice to rapidly built out an idea so I can get to validating the idea faster, once I've solved the content and front-end). Claude has been excellent at building out custom plugins to leverage custom post types in a way that's less of a hack. I was easily able to apply the existing design and Claude helped with a few of the templates, though it did start to deviate from the original design and layout. (I'll share this latest product once I've finished the final template, and inputed a few hundred LEGO sets—a limitation of the current LLMs is this task can't yet be automated).

There are still limits with Claude executing feedback, not to mention I've found even minor fixes require a full rewrite which takes time (and a lot of usage). One benefit of Wordpress is that I'm fairly comfortable with PHP. With most coding projects with Claude I've either asked for HTML/CSS & vanilla JS, or PHP for more complicated products, mainly because it means I can read and understand what Claude has written. Therefore once I have the basic structure built, I've found it's a lot easier to tweak a few lines of CSS or move around some HTML than going through the process of asking Claude to generate new code that may not be correct.

As someone who can code but I don't do it every day, I tend to be rusty, and that's not to mention how I may be a little unfamiliar with standard or established coding patterns. Claude has helped steer me towards writing better code. Aside from starting with the well structured templates and CSS, as I mentioned above, my style of modifying Wordpress is to "hack" a feature into the theme files, specifically the functions.php file, but Claude's more "professional" plugin included helper functions for retrieving the meta fields. Where previously I would include a call to the meta field in the header of the template, Claude gave me an object created as a helper function so I only need to instantiate the object in a much cleaner way. So Claude is actually making me a better developer.

Overall my experience using Claude to help me build my ideas has mostly worked. I have many other ideas, and I'm still experimenting with the process. A few ideas may need to be designed first, but there's something nice about the conversational approach to working with Claude to first flesh out the idea before experimenting with what is even possible to create in code. A lot rests on how I write the prompt and honestly for many of these experiments the prompts have been conversational instead of a prompt engineered with a detailed spec document so there generated results may not always be perfect, and the process of getting some feedback made has been trying.