An instrument, not an oracle. 

Ollama runs on a Mac mini in the homelab. It drafts, summarises, and answers questions against my own runbooks and notes, and it does a real share of my day-to-day work. Nothing leaves the house: no prompt, no document, no log. The write-up of why and how — which models, for which tasks, what the machine can and cannot do — is in Writing.

Khyenpa, Kuralok and this site are built with Cursor and an agent in the editor. The model proposes; I read every line before it ships. The useful habit is the same one the bench taught: a result I cannot reproduce is not a result, and a fluent answer with no source is the dangerous kind.

• Retrieval over my SOPs — the runbooks are the corpus; the answer names the document it came from.
• A Copilot agent on the Microsoft side — instructions plus one uploaded document, built as an interview coach; it declines what the document cannot answer.
• Scripts drafted, then run by hand — the Graph SDK clean-up that finds duplicate Entra objects started as a draft; removal was a second, deliberate step.

• An agent over the homelab runbooks, offline, for the 2 a.m. question.
• First-pass triage of Splunk alerts — shape, source, policy that held — for a human to confirm.
• SOPs drafted from shell history and corrected by hand, so the record is written while the work is fresh.

Possibilities, labelled as such.

• Reproduce or discount. A result I cannot get twice is not a result.
• Draft, then decide. Models are good at first passes and bad at being trusted with the last one.
• Watch the confident failures. The dangerous outputs are the fluent ones.