Module 5 — Build your own

Office Hours: Session 16 — Thursday, 14 January 2027

Session goal: apply the same pattern to a job of your own, and know when it should come to the AI Team.

Bring: a repetitive job you'd like automated, and the answers to the worksheet below.

The pattern

Gather (code): fetch the data and decide every fact with one right answer.
Worker (skill): make the judgment calls on that data only — search off.
Checker (skill): check and correct the worker's output, with a list of changes.
Repeat worker/checker for each judgment-heavy stage (analysis, then writing).
Deliver (code): anything formatting-, link-, or rule-based.

What's the output, and who reads it? (Email, Doc, Chat message, Sheet row.)
Where does the data live? (Calendar, Gmail, Drive, Sheets, a system with an API.)
List every rule. Mark each one: one right answer (code) or judgment call (skill).
What would a wrong answer cost? That decides how strict the checker needs to be.
What should the AI never see or do? (URLs, other people's data, sending on its own.)
How will you test it on a small slice first?

Weekly project status from a tracker sheet and related email threads.
Intake triage: new requests → structured summary → routed to an owner.
Meeting prep: tomorrow's external meetings with the last related threads and notes.

It needs data you can't reach with your own account (another system, an API, Chat).
Others will depend on it daily, or it needs to run for a team.
It needs code beyond small edits to the build kit.
A wrong output would have real cost.

Bring the worksheet to an intake conversation — it's most of what a one-page PRD needs.