I run 22 AI personas against one life, not one assistant against everything

The obvious way to use an AI assistant is one assistant, one context window, every kind of task. It works, right up until the assistant has to hold "what's my startup's PMF thesis" and "what does my partner need to hear tonight" and "which Todoist label means what" all at once — and it starts answering all three a little worse.

The fix I landed on: route by domain, not by session. A task-triage persona. A relationship-reflection persona. A GTM-discovery persona. Each one reads its own narrow brief before doing anything, each one owns a slice of a shared knowledge base, and a routing table decides which one a given task belongs to.

The part that actually breaks

Routing isn't the hard part — matching a task to an owner is a lookup. The hard part is the unowned task: something that lands in a default inbox with no clear domain, so nobody's lookup table ever finds it. A task about "overwhelm" with no further detail doesn't match any persona's keyword list, so every persona skips past it on every sweep, forever, until someone adds an explicit check for tasks with no owner — a check that has nothing to do with being more careful, and everything to do with the structure having a blind spot by construction.

The second hard part is staleness: a dossier of "what's true about this person/project" goes stale the moment the real action happens somewhere else — an email gets sent, a task gets done — and the dossier doesn't know yet. The fix isn't "be more careful reading the dossier." It's a hierarchy: the system of record (the sent email, the checked-off task) always outranks the note that describes it.

Neither of these is really about AI. They're both about what happens any time a complex job gets split across narrow, well-intentioned specialists — the same failure mode a big company's org chart runs into, just compressed into a single afternoon's bug.