Pick a vertical you actually know. Interview one real person about the work. Count whether the problem is worth automating: volume, frequency, cost of the manual way, and have the nerve to conclude "no" if the numbers say so. Then write the requirement one-pager: the problem in the stakeholder's words, what "correct" means as a test, the one thing that must never happen, the budget. Finish with your prediction sheet: expected cost, expected failures, expected accuracy, on record before anything is built.
Day one includes a foundations debrief, so nobody starts underwater.
It ends with
A one-pager your agent can execute, your named expensive mistake, and your first cost estimate on record.