Course contents
Capstone: Pilot One Real Workflow
Pick one workflow you own, run it through Work with approvals on, measure honestly, and decide like an operator.
OpenAI's own launch guidance is the capstone brief: start with one workflow you know intimately — not an organizational transformation. Classic pilots: lead triage, competitor benchmarking, monthly reporting. You'll run yours end-to-end with the guardrails on, measure it against your manual baseline, and make a keep/kill decision with real numbers.
Zapier, Shopify, and Virgin Atlantic on the workflows they actually delegated — calibration for choosing yours.
Choose and baseline the workflow
Pick a recurring workflow you personally own, with clear inputs and a deliverable someone actually consumes. Write down the baseline honestly: how long it takes you, and what "good" means.
VerifyOne workflow chosen; baseline hours and quality bar written down before the agent touches it.Run it with Plan mode and approvals on
Connect only the plugins this workflow needs. Commission the task with your best context — reference files, named deliverable, sources requirement — and review the plan before execution.
promptHere's my monthly competitor benchmark workflow: read benchmark-brief.md, pull the sources listed there, and produce the report in the attached template's format. Plan first — I'll approve before you run.VerifyYou approved (or corrected) a plan before the run; the artifact arrived with sources you can check.Review, revise, and re-run twice more
Score the output against your quality bar. Feed corrections back via scoped edits, capture the working recipe (template + brief), and run the workflow twice more on fresh data to see consistency, not luck.
VerifyThree runs; you can say specifically what the agent nails and where it needs your judgment.Decide like an operator
Compare total agent time (your review included) against baseline. If it wins, schedule it and keep the approval gates — OpenAI's guidance is explicit that gates are the operating model, not training wheels to remove. If it loses, write down why; that's your calibration for the next candidate workflow.
VerifyA keep/kill decision backed by measured numbers — and either a scheduled workflow with review gates, or a documented lesson.