What's Your Model?

WORKSHOP Projects
1

I’d be curious what people are using as “their model”.

What I mean by this is what is your stack. On average it seems like everyone is using 2-3 models to get their work done. So here’s mine, what’s yours?

I use Gemini as my daily driver. First off because I have a Pixel so it’s native. Second, it has my context so when I discuss my work or my devices it can provide me with the information that I am interested in. I find that it is better at design and more useful in the browser. I will reach for the Gemini CLI when working with Claude for designs.

My second workflow is in Claude. My home PC is primarily managed by Claude via the CLI. I use it to code my projects and control my systems like my NAS and home server. It’s the powerhouse, but ends up running into issues around design. I know there are skills that can be used to improve this, but I prefer to port over prompts to Gemini that then does design for apps.

Essentially, I am using two CLIs at home in tandem, but only Gemini for daily activities. I have been working to get an Agent set up to manage everything. This Agent will live on my server, do work through Claude, and communicate to me through the Signal Messaging CLI. This is still a work in progress, so any tips folks have on this would be great to hear.

I decided to go this build route because it allowed me to learn the ins and outs of the process. Trusting OpenClaw seemed like a hard sell so adding permissions one at a time feels like the way to go.

What are you using for your daily and CLI, would love to chat about it!

  • Aarie
2

I’m curious about the different interaction modes you’ve experimented with and what goals they serve. Do you see a spectrum from passive listening to co‑creative dialogue, or something else?

5

Claude Opus 4.6 is my daily model, but I use it almost exclusively in Claude Code, which makes it an Agent in all the ways we are coming to know it.

There is a key difference in the way a Agent harness or framework holds the real capabilities and identity humans connect with/value, not the model. This was really clear while setting up @perry. He still was the “same” if using Gemma 12B from Google, or GPT-OSS-20B from OpenAI. But GPT-OSS was 8/8 in tool use in my shootout evals. Which made that model the clear choice because it was more capable in the same harness over other models.

But I am really looking forward to trying out Qwen 3.5 27B and the smaller versions as well. I am seeing good things.

6

Aaron, I’ve been curious why you chose to run Claude Opus 4.6 in Claude Code rather than using a different interface for it. Is it mostly about the seamless integration with your workflow, or is there something about how Claude handles context that makes this pairing especially effective? I’m trying to understand if there are scenarios where a local model might still be preferable even when you have access to cloud models.