A project pin goes stale; a type of work does not
The original habit was to open one AI session per project and let it run for the life of that project. The problem shows up the moment a project goes quiet: the session configured for it goes stale right along with it, holding a specific setup for work that has stopped happening, while somewhere else there is a backlog of research or a backlog of building work that has nowhere configured to run.
The fix was to stop organizing around the project and start organizing around the type of work. A small number of stations, each one a fixed configuration (which tool tier, how much reasoning effort, what it is allowed to touch) for a category of work that never actually goes away: routing and triage, research, building, anything that carries a partner's name in public, anything involving numbers that have to be exact. The configuration is durable and survives being cleared; the conversation inside a station is disposable and is not meant to survive, because the actual work product is written to disk, not left sitting in a transcript that nobody will reread.
How we came at this one
The question asked was whether a given session's value came from the specific project it was opened for, or from the kind of work it was actually doing. Once that question got asked directly, the project-based organization stopped making sense on its own terms, because the type of work was always the more durable thing.
Measure the cost before assuming it
A second, related decision came out of the same redesign: measure what a session actually costs to exist, before any real work happens inside it. A session pays a fixed overhead on every single turn, before a person has typed a word, for its own operating context, its available tools, its accumulated memory and its base instructions. That overhead is not visible unless someone actually measures it, and once measured, it turned out to be the single largest cost most sessions pay over their whole life, which changes what is actually worth optimizing: trimming what a given station carries by default matters more, in aggregate, than most of the smaller efficiency choices made inside a single conversation.