Recurring Loops

Automate a stable decision process, not unresolved strategy. A loop must be observable, bounded, interruptible, and safe under duplicate or partial execution.

If .agents/marketing-context.md exists, read it first; it is the truth file every skill in this hub shares, and its evidence states are constraints. Read loop-system.md for qualification, the specification, state, cadence, approval layers, the catalog, orchestration, failure handling, dry run, and the loop report.

Operating contract

1. Prove the task recurs and has stable inputs, decision rules, useful outputs, and a real owner.
2. Define state, freshness, idempotency, deduplication, concurrency, timeouts, retry limits, partial failure, reconciliation, and retention.
3. Separate read and analyze, draft, approve, mutate external state, and verify; bound spend, messages, records, runtime, fan-out, and data.
4. Define no-action, pause, escalation, stop, and retirement before activation.
5. Run a fixed-snapshot dry run and one credible failure path before scheduling, then recommendation-only mode before external mutation.

Do not schedule, send, publish, spend, or change live accounts without explicit authorization for that action and scope; fail closed when authority or state is unclear.

Return

The loop specification, state design, approval layers, dry-run evidence, monitoring and loop report format, and pause, stop, and retirement rules.