AI / LLM

Development

My workflows change with frightening regularity, so this section is already out of date, and I haven’t even finished writing this sentence. Should I rewrite it before posting? NO! It will STILL be too late! It will ALWAYS BE TOO LATE!
…
Anyway, here’s what I’m working with at the time of this writing:

Antigravity IDE: I’m in the process of testing this as my daily driver. Gemini 3.8 is not bad at all for collaborative builds, and if your project is planned and scoped, it does a pretty good job, IMO. There’s also a lot of connected services to use and test in the Googlesphere, which my clients might use or want to use, so I want to stay familiar with them. In terms of cost, I was already paying for some of the bits in the AI Pro bundle, so it appealed to my frugal (cheap) side. If I keep my token use tight, it almost feels like the near-infinite inference I used to get from GitHub Copilot in the bill-by-request days.

VS Code / GitHub Copilot: I’m still using this for certain things like planning and security audits. I’ll use the more frontier models available in my GHCP subscription for specific tasks (whichever current “best” model from Anthropic or OpenAI). My CodeRabbit is hooked into this setup as well, so I do a lot of my GitHub work here.

Zed: I use this mainly with DeepSeek. This is for general advice, pre-planning, potential projects, quick spec planning, things like that.

DeepSeek has a new harness I’m looking to test, but I haven’t gotten around to it yet. It’s on my long list of stuff to test / goof around with.

Also, WordPress has a Studio I should be testing way more than I am, currently.

Below is the previous update I wrote. I’m just gonna leave this here until I figure out how to structure this whole section in the future:

---

GitHub Copilot

GitHub Copilot in VS Code is my daily driver. I’m using Claude Sonnet for anything involving design, writing, fact checking, etc. For anything involving infrastructure, security or refactoring, I use GPT-5. I’m experimenting with tiered strategy for token efficiency, due to the tokenocolypse over at GitHub Copilot.

I’m also testing Zed as an alternative to VS Code. I don’t think it’ll fully replace it as my main go-to, but it’s starting to fill a kind of niche in my workflow, and I find I have it open all day.

Cline

Before I moved to GitHub CoPilot, I was a Cline user. I still think it’s great and go back to it occasionally. I still love to test local models, and Cline is great for pulling those in from LM Studio or Ollama and making them agentic. I switch models all the time, so it’s kinda pointless to mention them here… although I’ve been getting a lot of mileage out of Qwen 3.5 9b @ 8q.

As an aside, I’ve recently been testing AI plugins for Obsidian. I don’t love them so far, but what I have had some luck with is opening my Obsidian Vault as a folder / project in VS Code, then running Cline connected to a small model through LM Studio to essentially chat with my Vault. This allows me to pull a lot of different data from my own notes without adding another plugin to Obsidian.

Gemini Cli

This is pretty new as of this writing, so I’ve only gotten to use it a little. It’s pretty good, I have to say. I think this fits into my workflow in an odd way. See, sometimes I want to work on an idea I’m having or quickly fix a thing I just noticed in a project I’m not really working on currently. I can open up a terminal in whatever directory I’m in, and immediately get to work on a thing without setting up a new project, environment or etc. That has value… I think.

Illustration

Okay, this one is tricky. There are a lot of very valid concerns when it comes to generative AI in creating artwork. I don’t have a lot of need for AI in my illustration currently, but I have been experimenting on using it for ideation, color balancing and tools like generative fill for better cropping… things like that. I’ve also been experimenting on training AI (through LoRAs) to use my own artwork to ink my own art using my own style. Is that ethical? I dunno man. It’s a complicated, messy world… but still amazing.

ComfyUI

This is my AI image generator / editor of choice for now. It works really well on my local machine with just 12g of VRAM. I chose this over Invoke on the strength of Qwen Image Edit, which as of this writing only works in Comfy (it’s pretty great). I still use Invoke for anything involving Pony.

Affinity

Affinity Photo and Designer are my current choices for illustration work. They’ve recently added a few AI features that I’ve played with a bit. I had thought that Affinity would be adding tools similar to what they have at LeonardoAI, since they were both acquired by Canva, but I haven’t seen that happen yet.

Update: Affinity combined their apps into one application and made it free… less than a year after I purchased the suite. Nice. Anyway, they’re doing a whole subscription thing through Canva for their AI features, which are not particularly compelling to me at the moment. We’ll see what they roll out in the coming weeks, though.