This is yet another AI post. I will try not to let this turn into a trend as I am just as exhausted with the ideas as everybody else. I’m writing this out publicly because I’m noticing a shift in the way I search for information and how I organize notes.

In My Current AI and LLM Usage post, I talked a about how I was dipping my toes into this AI mess. I was just as frustrated with the state of writing on the topic as I am now. That probably won’t change until this bubble bursts and we’re left with less nasty tooling.

This post is an update to that as well as a description on how my usage is shaped now. At the time that I wrote that post, I was being encouraged by a client to lean heavily into LLM usage for their project. I had the bottom tier paid Claude subscription and ended up exhausting my usage limits regularly on the project. Eventually I just decided to roll with the $100/mo plan. This expansion changed the way I interact with the models and I use it more frequently now.

I wrote every word of this post as I do with all my blog posts. Ain’t no LLM in my nvim…

My LLM Landscape

Since that post, have now consolidated to the following services:

• Kagi family plan
• Claude Max
• Proton Lumo

I have dropped the following services

• Kagi Assistant Ultimate upgrade
• Zed Pro
• LM Studio

After seeing Paralell Agents in Zed, I’m considering adding Codex or maybe another Zed Pro subscription but that’s a project for another day. I would like to use one of those to add to my review and search workflows.

I still don’t hit the limits of the Claude Max subscription and I wish there were a middle-tier option but I understand why there isn’t. I am hopeful for a day when I have enough hardware or the models are efficient enough for me to use them locally. LM Studio is just a bit too heavy for my hardware.

Use in Client Projects

The elephant in the room, the embarassing reality. I don’t like to admit that I use LLMs on client projects. I am a freelance engineer and I need to know that every line of code or configuration I deliver is solid. My work mostly consists of cleaning up and refactoring existing codebases or infrastructure. It would be disingenuous of me to just toss that work into an LLM and collect a paycheck.

Where it gets further weird for me is often the LLMs are substantially faster at implementing things than I am alone, even including my review time. This is a force multiplier for me, letting me get the day completed and not have to work late if I wield it properly. I feel deep embarassment or shame over using the tools, however I feel like I’m using them pragmatically so the ill feelings are likely unwarranted.

For software based projects that clients hand to me, I am often dropped into a large legacy codebase (or several codebases) where many people have contributed for over a decade or more. There is usually a lot of institutional knowledge involved and a lot of code that is very hard to follow. In these cases, I use Claude to help me understand the shape of various classes or modules while I am wrapping my head around everything manually in parallel.

When I’m digging through a new (to me) codebase, I will make a list of notes and then drop those into Claude to have it chase threads and see if I missed anything or if I am understanding data flow correctly. I will often have a couple of Claude sessions reading the code for me and giving me a sort of psuedo-UML diagram of how everything works so I can find other things that need fixing.

If the project is greenfield (a rarity for me), I will use Claude like a pair programming session where I am the one driving. I’ll tend to write code myself but I’ll have Claude look up information for me or write tests as I go. Often I’m using the LLM to avoid opening another window and context switching. I’ll have it run web searches for me or give me stackoverflow-esque code snippets with the help of context7.

A unique type of work I’ve landed on is reviewing LLM generated code for some clients before it is deployed. I’ll review the PR myself, make some notes, and then pass those notes to Claude for an additional review. This has helped me catch a few issues in large PRs that I might have missed due to fatigue or simply being focused on other sections of the code.

If the project happens to be infrastructure that is not IaC, I will write out all my notes and diagrams and then I’ll pass them to Claude asking for it to find any mistakes before I implement things. I’ll also often use it to look through patch notes nad update reports to see how safe it is to update machines. For this (and all of my work), I have a standing rule that it should provide references with each response and I read those references myself.

For IaC projects, I will use Claude more heavily than I would in code. I’ll let it generate OpenTofu blocks for me as well as have it spit out sections of YAML for Ansible playbooks. I am reviewing everything it returns but I find this helps me avoid running into silly whitespace errors. It’s also helped me find new things in OpenTofu/Terraform because it’ll grab the latest docs from context7 and I can go read them myself. Intellisense on steroids.

In all of this, I am constantly checking the work HITL-style. I know people who are currently running a bunch of parallel agents and using voice prompts to interact with them. They’ll go from idea to spec to test to implementation to deploy all without viewing or touching any code. At the moment, I am incapable of this workflow (and not just because I suck at voice interfaces). As I stated at the beginning of this section, I feel deeply responsible for every line of code or config I deliver to my clients and I verify everything before handing anything over.

Use in the Homelab

In my home network, I have two servers. One is mostly for boring network stuff and the other is a small slice of media-related things that you might expect to find in the common online homelab circles. I am indifferent if these two severs become smoking craters. I have backups and the services are totally non-critical.

This presents an interesting scenario. I put so much care and effort into my client’s networks and software that I often lack the energy to do everything myself at home. “The cobbler’s children have no shoes” and all that… If the infrastructure is disposable and I lack the energy, why couldn’t I let Claude run rampant on my hardware?

I’ve done exactly that. I have a repo set up with OpenTofu/Ansible configs and I have users set up on my servers with SSH keys that don’t require passwords. Included in this repo are documentation, incident reports, and runbooks. When I want to set up a new service at home or I want to trouble shoot an issue on one of my machines, I pop into this repo and prompt Claude with the issue, then let it run on its own with minimal interaction from me. It’s bolstered by linux-mcp-server so it makes more intelligent reads on system state.

I have rules for it to always update the journals, incident reports, laundry list, and documentation every session. When a session is over, it knows to commit the changes to the repo and then deploy the updated docs to one of my servers. I can review it all myself and Claude has what amounts to a detailed memory in cold-storage on disk.

This has worked shockingly well. I don’t know that I could ever trust it in production, I’m still a little uneasy about the fact that LLMs are not deterministic or idempotent. However codifying everything as IaC sort of resolves this issue to a degree.

A friend of mine is developing walterops and I am trying to find the time to tinker with it on a couple of disposable Linodes. It is an extreme version of this. I’m oversimplifying the work here but it amounts to a chatbot interface to ops work on the surface. It feels like it could be pretty powerful.

Use in Personal Software

Once again in home computing (not work related stuff), I am exhausted with programming. I work with code all day long and pay very close attention to it to produce quality work. When I have to write some scripts or I want a silly little program or game at home, I no longer write it myself. I am not ashamed to admit I straight up vibe code my personal tools that never see the world outside of my home network.

Programming, for some, is a craft. This is a very valid position and I am curious how it will evolve with the rise of LLMs. I still maintain a high level of code quality for work and that is where I maintain my craftsmanship. After work I lack the brainpower to keep making pretty code and my tools/games are small enough that I simply don’t care how clean they are.

I maintain a lot of frivolous data sets on books, movies, games, and music that I like to play with. I’ll fire up a Claude Code session and talk to it in plain English (or Portuguese) and ask it to sift the data, make graphs, make inferences, suggest ideas, etc. I don’t have to load up Python, I don’t need Jupyter notebooks, or any scientific libraries. I just dump my brain and provide the dataset. It’s fun.

Similarly, when I need to move a lot of files on my desktop or if I need to organize some media, I tell Claude to do it. I have backups. I am lazy from technical work all day so I let the AI do the work for me. I don’t need to remember options for rsync, awk, xargs, or whatever. I just let it ride.

Finally once in a while I’ll fire up Claude Code and have it make a game for me. I will never publish these games, nor will I share them with anyone. I often delete them after I’m done. It is a fun outlet because I am not a creative person in that way and I am not a game developer. Dumping some ideas into the chat hole and having it whip up a game for me is a lovely diversion.

Use in Information Finding and Recording

This, for me, has been the biggest change in the way I utilize LLMs. I am constantly looking up and recording information. I read all sorts of things on the web, lots of books/periodicals, and I dump absolute piles of text into markdown files (migrating to markdown from org-mode years ago has been useful).

I am really bad at organizing my diaries, personal notes, and work project notes. At the very least I’ve moved from just a flat directory that I used for about a decade to a few folders. I’ve tried things like evernote, joplin, obsidian, simplenotes, vimwiki, zimwiki, dokuwiki, fedwiki, cherrytree, and on and on… For whatever reason, other people’s structure just doesn’t work as well for me as a pile of plain text with rg (and its predecessors).

Claude has turned this process into a powertool. Because it reads and writes markdown so well, I can turn it loose on a directory and ask all sorts of questions. It’ll provide a response with references down to the line number. I can drop in anywhere and recall memories. I can link them together and I can have Claude generate even more text that tightens these connections. My ability to recall where I was on a project, what’s next, and what I did has increased massively.

For larger project work, like maintaining journals about my day-to-day work and planning future projects, I use compound-knowledge. This plugin helps me keep track of important arcs in my notes and the conversational aspect of it helps me think through ideas like a rubber duck scenario.

The tradeoff to this is that my real personal data has traversed Anthropic’s servers and I have no idea what they do with it. It sucks but the damage is done. There aren’t any secrets or bits of data in there that could yield identity theft but it’s probably rich with data for advertising.

This all hinges on Claude being local to my system and having access to my files so it renders Claude web as not very useful. I will occasionally fire up Claude web or Lumo web sessions and have them dump a markdown file that I can add to the heap and use for a next prompt. I’m not yet comfortable with the idea of an always on agent chillin’ in Discord or something like other folks.

Finally on this topic of information synthesis, I’ve noticed my Kagi search totals have declined rapidly since February and my Assistant usage is near zero. Kagi provides incredible results and is a lovely search engine but the LLM sessions have gotten really good at search (Kagi themselves lean into this).

An example of where this clicked for me was when I was looking up physical media trends for the past few years. I started by typing “physical media 2026” into Kagi and scrolled through results, opened 20 odd tabs, skimmed those, added some to Instapaper, skimmed more.

This was a pretty manual process and I thought about the same workflow in an LLM. I took my three word search and expanded it into a paragraph prompt that described the kind of articles I was actually looking for and wanted to read. Claude and Lumo both presented lists of links that were exactly the type of writing I wanted to read plus they provided references with counter-points and original sources for the articles. A couple more prompts later and I had a bigger picture about the trends that I was trying to understand. I added the relevant articles to Instapaper and read them on my Kobo. It felt like a more productive process.

Conclusion

I’m still tired of the LLM hype/doom cycles. I’m tired of all the writing and talking about it. Here I am contributing to it. I wanted to continue to present the middle ground that I find missing in all of this. I believe there are reasonable uses for these tools and they are just that: tools.

There is a lot of evil around them as there is in any capitalist pursuit. That doesn’t mean that there’s not any benefits to them. I respect those that opt not to use the tools and their reasoning. Yes, IP theft… yes energy use… yes AI psychosis… yes people getting dumber. I don’t deny those issues and I am not attempting to justify my usage. I am merely describing it.

I am hopeful that the next phase of this is a lot like the dot-bomb boom or some of the proprietary to open source software evolution. I hope that in the future I can run my own LLM server and have full control over my data. In the interim, I’m going to keep exploring these tools while maintaining a human touch for all of my work.

Perhaps in a later post, I’ll share my CLAUDE.md or MCP configurations but I am already adding more walls of text about AI to a world full of walls of text about AI… so maybe I won’t.