Classify failures first
Invalid parameters, permissions, timeouts, rate limits, and parsing errors require different responses. Do not retry every failure indiscriminately.
Design idempotency and recovery
Use stable identifiers for writes. When a result is unknown, read actual state before retrying. Track completed steps to avoid repeated side effects.
Prepare for a live change
Consider reduced budget, human handoff, or repeated execution. Explain how you would change, verify, and document the workflow.
Updated: 2026-10-10 · Verify these methods against actual projects.