Your first useful AI result Repair an unsupported draft using supplied evidence. Reviewed: 2026-10-03 Illustrated teaching example • Fictional data Narration: Microsoft Mark · locally synthesized 00:00:00.000 — A useful draft starts with evidence. Your first useful AI result does not need a perfect prompt or an expensive model. It needs a clear task, a source you can inspect, and a check that catches unsupported claims. We will repair a fictional proposal together. Visual description: Source → Draft → Check → Repair 00:00:17.292 — Keep the source in view. Here are the supplied facts. Riley works with a six person team. They process forty requests each week, and each request currently takes eighteen minutes. A two week AI pilot is proposed. The pilot has not run. Savings have not been measured. Visual description: Team: 6 people Workload: 40 requests / week Current process: 18 minutes / request Pilot: Proposed · 2 weeks Measured savings: Unknown (warning) 00:00:38.042 — A polished draft can invent a result. This constructed draft says the twenty person team already saves forty percent and should roll out immediately. It is fluent, but the staffing figure contradicts the source. The savings and completed pilot are invented. The rollout recommendation has no supporting trial result. Visual description: Our 20-person team already saves 40% with AI. Roll out immediately. 20-person team: Source says 6 (fail) Already saves 40%: No measured result (fail) Roll out now: Trial not completed (fail) 00:00:59.083 — Which claim can you support? Pause here. Which statement follows from the supplied facts? Is it that AI saves twelve hours each week, or that current processing takes twelve hours each week? Explain your choice before continuing. You may calculate the baseline, but you cannot convert it into measured savings. Visual description: 40 × 18 minutes = 720 minutes. What does that establish? A. AI saves 12 hours each week. B. Current processing takes 12 hours each week. B · Baseline workload, not AI savings. 00:01:21.000 — Ask for the right artifact. Give the model the facts and ask for a proposal rather than a success story. Require it to separate what is known from what is proposed. Ask it to leave savings unknown and identify the measurement needed. The instructions help, but you must still check the output. Visual description: Context: use only Riley’s supplied facts. Task: draft a short pilot proposal. Standards: distinguish current facts from proposals. Evidence: do not invent savings or completed work. 00:01:40.333 — Repair the draft. A defensible draft says the six person team proposes a two week pilot. The current workload is twelve hours of processing weekly, assuming all forty requests take eighteen minutes. Measure the pilot before making savings or rollout claims. This repair preserves useful information without inventing success. Visual description: The six-person team proposes a two-week pilot. Current processing totals 720 minutes weekly if all 40 requests take 18 minutes. Measure time, quality and corrections before deciding whether to expand. 00:02:03.750 — Check before you use it. Compare each consequential claim with the source. Check numbers and their units. Separate proposals from completed actions. State missing evidence. If a claim cannot be supported, remove it, qualify it, or obtain the missing evidence. A confident tone is never a substitute for this check. Visual description: Facts: Match the source (pass) Calculation: 40 × 18 = 720 minutes (pass) Status: Pilot remains proposed (pass) Unknowns: Savings remain unknown (pass) 00:02:27.250 — Now audit a different brief. Return to the lesson exercise and inspect the new draft yourself. Mark supported and unsupported claims, write a corrected version, and explain one important correction. Use the supplied fictional material. No account, paid AI tool, or personal information is needed. Visual description: Produce a corrected brief and a claim-to-source check. 1. Mark unsupported claims. 2. Repair without inventing new facts. 3. Explain one correction. Your evidence: a draft another person can verify.