Kit 4 · Train the AI, Check the AI · deliverable · designed and built by I Wayan Usada — instructional design portfolio
Support agents run the 4-rung Verify check before sending any AI-drafted customer reply, on every send, without a supervisor prompt.
Agents send AI drafts after a glance; hallucinated details reach customers because checking feels like extra work, not part of the work.
| # | Behavior | What it looks like | Where the artifact trains it |
|---|---|---|---|
| 1 | Spot the risk hotspots in a draft before editing | 3-of-5 hotspots found, safe decoys left unflagged | Scene A-B hotspot hunt |
| 2 | Run the Verify ladder on flagged lines | four checks executed in order: source, fact, tone, commitment | Scene C Verify ladder |
| 3 | Make the send decision and own its consequence | send / edit-and-send / reject decision with customer outcome | Scene D branching send |
It will not ban AI drafts or teach prompt-writing — the deliverable is the moment-of-send check, nothing upstream.
Debrief scores hotspot discrimination (not volume), ladder completion, and the send decision — naming the check as the habit.
Prefer the source? Download the raw markdown.