The architect and the engineer never share a context window. Here’s the protocol that lets them work together anyway: a handoff comment in Linear, ADRs for anything durable, and a status field that says whose move it is.
The architect/engineer split from the last post became an actual CLI, and the engineer role moved off my laptop onto a machine of its own.
How splitting the work between a frontier model and a cheaper one cuts costs without a quality hit.
System prompts don’t work reliably for smaller models. Guidance injection delivers instructions at the exact moment they’re needed instead.
A growing collection of tool docs that agents read on demand
How I turned an idle Mac Studio into a private AI backend, and built purpose-built agents on top of it.