Bootstrap before acting

The rule

At the start of every session, before responding to the first request: read the project's index, check its open items, and read whatever entry-point file the project designates. Only then act.

The scar

Agent sessions have no memory of each other. Without a bootstrap step, every session begins by either (a) re-explaining the same architecture from scratch, or (b) worse, not re-explaining it and proceeding on stale assumptions from training data or a half-remembered convention.

The specific failure that made this a hard rule: a session picked up work on a screen that had already been investigated in a prior session, re-derived the same root cause from the codebase, and reached the same conclusion — hours of work, zero new information, because the existing write-up was never read. The knowledge existed. Nothing pointed the session at it.

How to apply

1. The project's AGENTS.md/CLAUDE.md names the entry points — read them first, not after the first tool call.
2. Check open items (grep -r "status: open\|status: in-progress" over the entry folder) before starting anything new — the thing being asked for may already be in flight.
3. Never skip this because the request "looks simple." The cheap requests are exactly the ones where a five-second index read saves an hour of re-derivation.
4. If bootstrap reveals the request is already answered somewhere, say so and cite it — don't redo it to be thorough.