Begin with a task you already understand, not a tool you just discovered.

Choose a narrow repeatable task, define the safe input, place the human checkpoint, and write the stop condition before running a trial.

Choose the ordinary task

Pick work repeated at least monthly with an input you can identify and a draft output someone already reviews. Avoid hiring, firing, lending, diagnosis, legal conclusions, safety decisions, or automatic money and access changes as a first experiment.

What starts the task?
Which approved source does it use?
Who checks the draft?
What final action stays human-owned?

Shrink the first run

Use one document type, one ordinary scenario, one approved output format, and fictional or safely redacted input. A good trial is easy to stop, compare, and repeat.

Write the checkpoint before the prompt

Name the reviewer, facts to check, authority limits, and required evidence. Put the hold before sending, publishing, paying, changing access, recording a decision, or affecting a person.

Decide how the trial ends

Stop on prohibited data, missing sources, unsupported claims, unusual consequences, failed checks, or provider changes. Keep the original manual path usable.