Workshop: The scope

This workshop turns your Module 1 opportunity into a scoped, feasibility-tested v1 — with the prototype-prompting instrument used for real.

1. 1Run the prototype test: assemble 15-20 realistic inputs for your feature (for the Dana case: delayed-order records with customer context — fabricate realistically or sanitize ruthlessly; if you sanitize real customer records, anonymize before the data reaches the chat window — the same privacy line Module 4 draws for episode logs applies the moment you paste a real record into a prompt, so flag the tension here where it first arises rather than discovering it two modules later), write the core prompt, and grade outputs against a 5-line quality sketch (correct? appropriate tone? would you accept this as the customer?). Record the finding honestly: proceed / kill / caveats — and list every caveat as a future eval case. If the finding is kill, run the workshop on your Module 1 runner-up; killing on evidence is a passing grade with distinction.
2. 2Score the triad for your (surviving) use case: value density with numbers, the workflow-position analysis (where does human verification naturally live? — write the v1 position explicitly), and the data reality table (each data need: exists? current? permitted? — with the one honest 'this needs checking' flagged for engineering).
3. 3Write the scope contract: v1 in the tolerant position with the vision noted as roadmap; 4-6 non-goals each carrying its reason; the miss-behavior table (low confidence / missing data / unexpected reply → designed path each, with forecast percentages you'll defend); and the launch-blocking risks from your error map's expensive quadrant, each with its mitigation named.
4. 4Pencil the unit economics: cost per use from your prototype's actual token counts, value per success with a range, break-even rate — one honest paragraph. Then assemble the scope doc (2 pages max) and run the adversarial read: 'you're the engineering lead: what did the PM get wrong or skip?' Fix what's fixable; append what's arguable.