Making My Own LLM Tools

I made a few tools using my FreeGPT API to test out a couple of my ideal LLM workflows- “snippet generator” and “wtf does this highlighted code do?”. Neither require fancy models, large contexts or anything besides one-shot prompting, so the free versions do just fine. I don’t really work with massive codebases, so I don’t need codebase exploration. If I did I would use one of the many off the shelf integrations from any LLM vendor. Likely the cheapest one.

Generating the first draft of a snippet that has been “solved” is such a time saver- things like basic CLI parsing, or sending an HTTP request with the cURL library, with followup drafts reworking what fell out of the model to fit my needs.

When looking through someone elses code, or when working in a language I’m not familiar with, having a box that I can direct my dumb questions to is also really helpful.

how it went

Meh.

Hats off to Cloudflare, I got detected a few more times, though not the reason I stopped using the tools I made.

After implementing the two cases, I realized it just wasn’t that useful to me personally. I “work” on my own time for two reasons,

  1. to learn, or
  2. for fun

It’s more fun if you consider these the same thing.

I found I was prompting more and more, more and more just sitting and waiting, then validating someone elses code. And turns out that leaning on LLMs more when working on my side projects meant I started enjoying it less. I learned this early on with the release of ChatGPT 3.5- I generated several graveyards worth of rust projects trying to “learn” it, not sure why I thought it would be different now.

The engineers at OpenAI got it right the first time with the basic chat window. I just have a hotkey in my window manager that takes me to a duck.ai browser tab (or one of the many free money furnaces hosted by nvidia). Building it into my editor seemed convenient, but I didn’t really notice any time savings. The majority of the time spent using LLMs for my usecases is taken up by generation and validation, not by how many milliseconds it took to prompt the LLM.

I am considering building some kind of project exploration tool using a cheap API like DeepSeek. Should be fun to make. Though I’m not sure on the actual value. For my personal small-ish codebases I could just grep it myself and then carry the knowledge forward with me- rather than have to ask a chatbot each time I need to trace some kind of flow. It’s not high on the list of projects.

where it’s going

In my experience, relying too much on LLMs to offload my thinking is both pretty boring and useless for any meaningful learning.

That, and every time I use these tools for learning areas that I am not familiar with I am reminded of the Gell-Mann Amnesia Effect.

“iF YoU dOn’T uSE tHeM YOu WilL bE lEfT bEhInD”. A few years ago it was “best practices” to threaten to kill yourself to get ChatGPT to reliably produce JSON output. Things like ClaudeCode, Codex, OpenCode etc. are really… simple? You can pick it up in an afternoon after watching a YouTube video if you’ve got even a little experience.

I would argue that atrophying your skills by delegating a significant amount of your thinking to a vendor-locked chatbot is a much greater concern than being left behind.

To be clear I still generate first drafts of boilerplate and consult the chatbot on strange new code. I just do it using the free version in the browser.