Session Bootstrap
Before an AI collaborator does any work in this project, it loads three layers of context, in a fixed order. The goal is that every session — and every different AI tool that might ever sit in this seat — starts from the same shared understanding of what's forbidden, what already exists, and how to behave, rather than re-deriving it from scratch or trusting a stale memory of a previous session.
The three-layer read order, and the one question that decides whether to run it: has this already happened this session, or is there a reason to suspect drift?
The three layers
Layer 1 — the forbidden-pattern catalog. A living list of mistakes this project has actually made and corrected — not hypothetical risks, but real incidents with a wrong example, a right example, and a detection checklist, so the same mistake doesn't get made twice by a different session.
Layer 2 — the tooling and environment catalog. An inventory of every reusable procedure, script, and environment mapping already available. The rule this layer enforces: check what already exists before building something ad hoc. Re-inventing a tool that already exists is itself a forbidden pattern from Layer 1.
Layer 3 — the full rulebook. Read start to finish, this is where the approval gates, the quality bar, and the invocable-dispatch protocol actually live. It also states how to figure out what's currently in progress — from the branch name and the most recently modified progress file, never from a hand-maintained table, because hand-maintained "current state" tables reliably go stale.
When this runs
Automatically, once, at the start of a session. It can also be re-run explicitly mid-session — after a very long conversation, or whenever there's a reason to suspect the working context has drifted from what the rulebook actually says. Re-running it costs nothing but a few reads; not running it when drift is suspected costs correctness.