Running Claude Code as a System, Not a Chat

Most of the value in a coding agent is lost to how it gets used: one enormous session that drifts, context that bloats until the model forgets the brief, and the same setup instructions retyped every Monday. None of that is a model problem.

This directory is the fix. It holds the operating doctrine, a decision tree that answers "where do I open this and in what shape", four reusable workflow scripts, a registry of everything that runs on a schedule, and a routing matrix for when to hand work to something other than Claude.

The two workflows that carry it are `iterate-until-good` (draft, critique with an independent panel, revise, and loop until a separate judge passes a fixed bar) and `deep-work`, which researches from several angles, synthesises a brief, then feeds that into the first one. The pattern in both is the same: the thing that generates is never the thing that decides it is finished.

The decision tree is the part that changed how I work most. It is three questions (which directory, which shape, which primitive) and the answers are mostly "the narrower one". Opening at the vault root loads every domain's context and every shared memory, and the model gets measurably worse at the actual job. Opening in the project directory loads one context file and the skills scoped to it.

The routines registry follows one rule that took a while to arrive at: anything disabled stays listed with `status: off` rather than being deleted. A registry that only lists what is currently running cannot tell you what you tried and turned off, which is the question you actually have six months later.