Not everything I build works out. These are the ones where knowing when to stop, or when to switch to something better, mattered more than the code.
An AI card-set generator that brute-forced its way into a wall
I spent several months trying to generate a complete, playable custom set of about 250 trading cards using only local language models running on my GPU host. The design was ambitious:
- "Squads" of agents, each responsible for a part of the set
- A loop controller to keep them iterating
- A wiki as the shared source of truth
It didn't work. The output was never coherent enough to call a finished set. Adding more agents and more loops made it slower and more expensive to run, not better.
What I learned:
- More agents is not a strategy. When the output was bad, my instinct was to add another reviewer or another pass. The real problem was that the task wasn't broken down well enough for any single step to succeed.
- Small local models have limits. The models I could run locally at the time weren't capable enough for long, consistent creative work, and no amount of orchestration made up for that.
- Evaluate early. I didn't have a clear, checkable definition of "good" until late. Without one, I couldn't tell whether a change helped.
I'd like to try it again with a much narrower scope, a clear evaluation step, and better tooling. A related card-creator app I rebuilt was also shelved.
GetContext.sh → Sigmus-Agents → Claude Code
When AI coding assistants first became useful, the hard part was giving them enough context about a project. I wrote GetContext.sh, a Bash tool that builds a Markdown snapshot of a project from templates: file contents, directory trees, command output, even nested templates. I released it as v1.0.0 and used it across many projects.
I thought I was onto something with Markdown files as context. Then it turned out that's exactly where the industry was heading, and the tools that came out of it did it far better than I had.
I took the idea further first. Sigmus-Agents was an autonomous multi-agent system:
- Agents claimed issues from my self-hosted Gitea, wrote code and committed it
- A supervisor agent managed them
- A real-time chat UI let me watch and steer
It worked well enough to teach me a lot about orchestration. But once Claude Code and similar tools had good agents built in, maintaining my own made no sense. I switched and abandoned it.
TrueNAS → a NAS I control
I ran TrueNAS (both CORE and SCALE) for a while. It's a capable product, but I didn't like how much it decided for me, or how proprietary its management layer felt for a box holding all my data.
I replaced it with a Kubuntu VM on my hypervisor:
- Samba shares for my Windows machines
- NFS exports for the Linux machines and the other VMs
- Storage on a hardware RAID array, carved into shares for the different kinds of data
It's simpler, and I understand every part of it. If I did it again I'd use Ubuntu Server without a desktop environment. The desktop doesn't cost much on that VM, but it's still something I don't need. More detail is in the homelab case study.