AI should help.
Not confuse.
Start with the job, not the technology. A useful result beats a longer conversation.
1. Name the actual job
“Draft a reply using these three facts” or “fix this one failing test” gives you something concrete to review. A general request to improve everything leaves too many decisions open.
2. Separate facts from hopes
Record what exists, what you observed and what remains uncertain. A prepared file is not a live website. A passing unit test is not a tested payment journey. Keep those differences visible.
3. Set a finish line
State the output, how you will check it and when the worker should stop. A written budget boundary is helpful guidance, but it is not an enforceable provider spending limit.
4. Review before acting
Check names, dates, prices, links and any promise made to a customer. For code, keep a backup and test a bounded change. For sensitive or consequential work, use appropriate professional advice.
5. Hand over the result, not the whole history
Keep one current note with completed work, evidence links, settled decisions, unresolved blockers and the next small action. Label old plans as old plans. Remove secrets before sharing.
“Here is the verified state. These decisions are settled. These files matter. Complete this one next step, show the evidence, and stop.”
Adapted from Rob’s earlier practical-AI service material. This is a general workflow, not a promise of accuracy, quota savings or business results.