LLMs as Tools, Not Solutions

I started using ZMX to better manage my remote server access and took the opportunity to solve two lingering annoyances in my dev setup. The first was needing to choose the right IP address depending on whether I'm home or connecting through my WireGuard VPN. The second was password-based authentication. The first I solved with OpenSSH's Match directive. The second I solved with an app called Secretive.

I'm building a system that performs rhetorical analysis using LLMs, helping me evaluate whether a piece of writing's arguments actually support its conclusions. The first version works but chews through API tokens, so I need a second version that can run smaller models locally. In this thought, I discuss building a custom LLM inference harness in Zig, the debugging journey to get it working with llama.cpp, and why I chose to write all the code myself rather than relying on Claude Code.

In the original static publishing system for Skriptble Press, we didn't have to worry about the visibility of articles: anything that was published was visible to everyone. The next version of the publishing system needs to handle different visibility levels for article reviewers, subscribers, and the general public. In this thought, I discuss the design evolution from a simple Publish Time metadata field to a visibility state machine.