Idempotent writes
ATL agents and ceremonies re-run constantly — on a restart, a retry, a resumed sprint. This core rule makes those re-runs safe: a write to any durable store must converge on the same end-state rather than duplicate work or clobber a newer edit. This page is the user side of that discipline.
What's happening under the hood
The idempotent-writes rule auto-loads in every session (and into the autonomous claude -p workers the delivery-team spawns, via the same global rule reflection). It codifies one habit: check-before-create by a stable key, overwrite current-truth in place, never blind-write.
It was distilled from the corpus itself — the same principle was independently re-derived in the delivery loop (its cross-backend "concept #10"), in /brainstorm, in /create-code-diagram, and in /profile-restore. When a discipline recurs in that many unrelated places, it belongs in one rule.
What it means in practice
Check-before-create by a stable key. Before creating a durable item — a work-item, a page, a config — the agent searches for it by a key derived from stable inputs (a parent + ordinal, a content hash), never a per-run id. Found → it reuses and updates; not-found → it creates and stamps the key. A collision resolves to the existing item instead of erroring.
Overwrite current-truth in place. A store that holds "what is true now" — a wiki page, a generated diagram, a review report, a config file — is replaced on re-run, not appended to. Running the same ceremony twice leaves one converged result, not two.
Never blind-clobber a newer edit. When an overwrite could destroy data written since the agent last read the target, it guards the write — comparing timestamps/versions, or diffing and confirming — because losing a newer edit is worse than pausing. This is exactly what /profile-restore does before it touches your global memory.
Why it's a core rule
ATL's autonomous delivery depends on it: a resumed sprint must converge on the durable state, not re-create it — otherwise a restart means a duplicate work-item, a doubled PR, or a lost edit. Making convergence the default (not a special recovery mode) is what lets an unattended claude -p worker retry and resume safely. The one intentional exception is append-only stores — a journal or audit log, where adding a new dated entry each run is the point, not a duplicate.