The primary goal of this project is to gain exposure, experience, and maybe some muscle memory in the realm of large language models, especially with the CLI-based agentic model orchestrator tools: Claude Code, Gemini CLI, and OpenAI Codex. I may retroactively add GitHub Copilot into this, but that's something of a different beast, being a preschooler-targeted "click the pictures and make it go" type of product. (My disdain for the WIMP paradigm is legendary.)
revised opinion GitHub Copilot has matured quite a bit. Additionally, they have also introduced Copilot CLI. The VSCode elements of it are still VSCode and pretty heavily crayon-oriented, but they're better than they were.
The overall goal is to end up with a laptop that suits some of my computing needs (gaming, coding, tinkering with technology, AI, and maybe some music production). I have a System76 Kudu laptop that I'm not really using; it was most recently running Windows 11 to facilitate gaming. However, I couldn't tolerate Wintendo for very long and rage-wiped the disk. Currently it has a barebones Gentoo Linux installation after some iterations on an earlier variant of this experiment to guide the installation. Gentoo turned out not to be a great choice because of the ridiculous source compile times of many core system packages (gcc, llvm, kernel, other suff). Claude Code could not get past the hour+ compile times without timing out and crashing. It turns out that the "build everything from source" is not compatible with "rapid iteration" style work, which is where the conversational nature of the LLM chatbots and agentic tools really shines.
update The System76 device has been put out of its misery. It is no longer a player.
Back to the drawing board, I decided to use this LLM-consensus approach to help me decide on the "best" Linux distribution and the "best" window manager and environment to use, optimizing for how easily the system configuration and other data can be parsed by a LLM.
I hadn't even considered NixOS until Gemini threw me that outside-the-box curve ball (and I didn't even change any model parameters like temperature!) making the case for the declarative immutable system configuration paradigm being extremely LLM-friendly and automation-friendly.
As soon as I modified my general LLM prompt and my system prompt "Linux expert" (Tinus Lorvalds is his name) to consider NixOS alongside Arch, Gentoo, and other more "traditional" distributions, NixOS became a top contender with almost all models.
I have held for decades that the way we interact with computers is twisted and unnatural. Typing is not a natural human skill. Information presented on a screen in a very rough "desktop" metaphor is also extremely limiting and unnatural: hence my disdain for the WIMP (Windows, Icons, Menu, Pointer) paradigm. The most "natural" human information exchange channels are voice/hearing and writing (with a handwriting implement, not a typewriter)/reading.
The explosion of NLP and LLM's (as well as speech recognition technology such as Whisper) in the past couple of years has changed things dramatically. While I'm still typing, I can now begin to integrate "natural language" or "plain English" in my typed commands and inquiries, rather than having to memorize a ton of archaic text things. With some additional work, rather than typing, I can speak; this is fairly commonplace with smart phones and increasingly with tablet and laptop form factors.
Using LLM agent technology as much as possible (use plain English to issue commands), build a customized laptop that is optimized for automation/scriptability and LLM parsability. The system must be usable as a daily driver for software development, AI/ML/LLM/MLOps work, gaming (Steam/WINE/Proton), and possibly music production.
another revision the long-term goal is the same, and I think actually inevitable, but I am going to let it happen more organically, rather than try to build some big thing that does all this right up front.
Prompt several different LLM's with the same goal and guidelines, gather their output, and have my agent orchestrator synthesize a "most suitable" plan and configuration from a consensus of the other models, providing me the opportunity to override any decision.
I've been working a lot with Claude Code in my freelance/personal/pet project space. It is extremely powerful, and is the best demonstration of what I think is going to be how software is going to be developed from now on. My day job limits which AI models we have access to (none, outside of whatever they make available through the corporate Azure AI contract), has involved more use of GitHub Copilot, increasingly in Agent mode. I've also tinkered a little bit with Gemini CLI, and for fairness, I also want to assess OpenAI's contribution to the space: Codex CLI.
At the same time, I am learning how the top LLM's (at the time of this writing, 03 August 2025) behave with different inputs. A few of them provide dedicated chat interfaces, while others were accessed through a ML model gateway provider. I generated a "Linux Expert" system prompt (one that I've been using for many months for Linux expertise), documented the target system's hardware specifications, and wrote a prompt to guide the model to choose the top 5 Linux distributions and the top 5 graphical environments (window managers or desktop environments) given my parameters, and to provide a high-level implementation plan. I then instructed the agentic orchestrator (Claude Code, Gemini CLI, Codex CLI) to analyze and synthesize these into a "best of" recommendation and a more detailed implementation plan. Each of these is in its own branch in this repository.
These are the models I used to contribute to the consensus. The selection was not arbitrary, but came from a general sense of the "best" models in common use at this time.
- ChatGPT-o3 - "Deep Research" enabled. I used the consumer-facing ChatGPT UI, not the playground, so I did not have an opportunity to provide a system prompt. I entered the text of my system prompt as the start of the user prompt, then input the content of the user prompt, and attached the system specifications.
- Claude Opus 4 - Enabled "Research" mode. I did not have an opportunity to provide a system prompt. I entered the text of my system prompt as the start of the user prompt, then input the content of the user prompt, and attached the system specifications.
- Cohere Command-A - Cohere's playground provides an input field for a system prompt/message, where I input the text of my system prompt. The UI does not provide a facility for attaching files, so I pasted the contents of the system specifications at the start of the user prompt, and then pasted the contents of the actual prompt.
- Deepseek R1 05/28 - My interface to this model was via Fireworks AI which is somewhat limited, even in the Playground. The Fireworks UI does not provide either a system prompt or attachment capability, so I created a user prompt that simulates the system prompt and includes the system specifications and the prompt text.
- Gemini 2.5 Pro - "Deep Research" mode enabled. While the Gemini UI does provide a facility for activating system prompts (they call it "Gems") I did not use that for this project because it disables access to Deep Research mode. Instead, I attached the target system specifications, entered the system prompt as the first part of the user prompt, and then entered the content of the prompt.
- Kimi K2 Instruct - via the Fireworks.AI playground. Same caveats as described with the Deepseek model.
- Llama 3.1 405B Instruct - via the Fireworks.AI playground. Same caveats as previously described models used via that interface.
- Mistral Le Chat - via Mistral's chat UI with "Research" activated. Activating "Research" disables the file attachment ability, so I used the generic chatbot input that collates system prompt, target system specifications, and user prompt into one user prompt.
- Perplexity - via the Perplexity web UI; enabled "Research" mode, attached target system specifications, entered "system prompt" as first portion of user prompt.
- Qwen3 Coder 480B A35B Instruct - via Fireworks.AI playground. Used generic chatbot combined prompt.
This has been a nearly unqualified success, especially considering the primary goal: to learn the capabilities and limitations of an agentic AI orchestration tool and the underlying models.
The first attempt used Gentoo Linux, which it turns out is not a good choice for managing with something that's essentially a tool to facilitate iterative changes and thinking. Gentoo's flexibility, ordinarily a benefit, turned out to be something of a liability with the LLM agent guided experiment. Additionally, the very long compile times of many core packages caused the agent tool (Claude Code) to time out and crash. (Aside: I think Anthropic could do a little better than outright crashing the app when it times out waiting for tool response.)
Back to the drawing board, I asked the models "hey, what Linux distribution do you think is suitable?" However, I asked in something of a leading way, providing a few examples (Arch, Gentoo, Debian-based, RPM-based). Only the Gemini model ventured outside the box a bit and suggested NixOS as suitable for LLM parsing: a consistent text-based configuration approach: "Operating System as Code". Once I added that into the examples for the other models, it became nearly unanimous.
One of the things I wanted to accomplish with this experiment was to compare the capabilities of three agent orchestrators: Claude Code, Gemini CLI, and OpenAI Codex CLI.
I started with Codex, but that shit the bed very quickly. The UX for that tool is hostile and inscrutable, not even offering any sort of internal "help" command.
I ditched that fairly quickly and moved on to Gemini CLI. The UX is much more approachable, very similar to Claude Code's UX (I would say "a direct rip-off" but a good idea is a good idea). While that was substantially better than Codex, it fumbled a lot. It devised a comprehensive plan, but once it started executing and iterating, it made lots of mistakes. Worse, it is extremely sycophantic. I am certain that I could provide some instructions to compel it to be less of an ass-kisser, but it was very off-putting. "Oh, my! You're absolutely right, I am so stupid! What a dumb mistake and you are next to God for having pointed it out to me. I bow to your greatness."
So, back to Claude Code. What an absolute banger. There were a couple of minor course corrections provided along the way, but once I got the base system installed and instrumented with Claude Code, it essentially ran with it and left me with a running and functional Steam/AI/Coding laptop. Oddly enough, it didn't install a browser!
The initial build used Hyprland as the window manager, which provided tremendous flexibility and customization. However, the maintenance overhead and constant configuration tweaks ("yak shaving") became a distraction from actual work. After several iterations, the system was migrated to GNOME, which provides a comprehensive desktop environment with sensible defaults and less manual configuration.
Key takeaway: Sometimes the "simplest" or most "automation-friendly" solution on paper (minimal window manager with text configs) can actually create more friction than a full-featured desktop environment with battle-tested defaults.
I'm intrigued by NixOS, and will probably replace Arch on my other devices with that. Unfortunately it's not an option on my VPS. (Turns out, ssdnodes kinda sucks. You get what you pay for.)