Stop, save, inspect, narrow and retry.
Do now: Write the last known-good step for one task that recently failed.
Preserve originals before repairing outputs.Confident recovery turns errors into system improvements. Panic often creates the second mistake that makes the first one expensive.
Stop the chain
Do not keep publishing, deleting or generating while the failure is unclear. Save the prompt, output, source and error message.
If privacy, money, safety or public misinformation may be involved, escalate to the responsible person before continuing.
Find the failure layer
Check the goal, input, instruction, tool capability, connection, output format and human review. A vague goal needs a different fix from an expired login or a hallucinated source.
Reduce the task to the smallest reproducible step. Change one variable at a time so the result teaches you something.
Random retries produce motion; controlled retries produce diagnosis.
Repair the process
Recover from the original or last approved version. Add the missing checkpoint, permission, file copy or verification step.
Record the symptom, cause, repair and prevention in plain language. Future you is a collaborator who deserves decent notes.