An AI assistant earns its keep when it stops handing you commands to run and starts doing the running around itself — against systems it can read, with records it cannot quietly rewrite. These guides are the working versions of that idea.
Have two language models of different families each propose a fix for a real bug, then argue with each other until they agree. Measured on a 44-case suite of real home-server bugs, the debate beat every single model we tried — including ones three and four times larger — by solving cases neither got right alone. The full method, the per-model scores, and the one thing it still cannot do.
Replace the ever-growing infrastructure document with a living library: every fact a statement with a stable id, history that names what replaced it, notes as the only write path, a generated website with search and a dashboard, and a coverage report that proves nothing fell out of the old documents. The page carries the complete specification — one copy button — written for an AI assistant to build the same system over your own corpus, whatever it is about.
Stop being the copy-and-paste link between an assistant and your servers: a ninety-line queue the AI drops script files into, run one at a time with every log kept, a Telegram message when the queue empties, and a systemd user service that survives reboots. Two lanes keep quick jobs from waiting behind long ones, a Prometheus gauge tells your alerts a measurement run from a fault, and one convention paragraph teaches any assistant the protocol. All four scripts in full.
The subtitle generator built on the AI appliance, released as open source in October 2026. It hears which language a video speaks and where, transcribes with two speech engines that check each other and a language model as referee, uses any full subtitle track already in the file before it listens, and writes one subtitle file per language — 45 to choose from — beside the video, unattended, from a queue with a browser front end. Runs on anything from an 8 GB NVIDIA card up, with Docker or on the host; nothing leaves the building. Includes the measurement that showed the voice-activity detector was throwing away whole conversations, and what replaced it.