What to get right before you build an AI agent
Work through this before anyone writes a prompt. The model is the part of this project least likely to be your problem.
- 01
Name the one job it does.
An agent that "helps customers" cannot be built or tested. An agent that "answers where-is-my-order and reschedules deliveries" can be. Scope is the single biggest predictor of whether an agent reaches production.
- 02
Decide what it is allowed to do, not just what it can say.
Reading an order is one risk profile. Issuing a refund is another. Write the list of actions before anything is built, because that list is the security boundary and it is much harder to narrow later.
- 03
Find the source of truth for every fact it will state.
in the PDF
- 04
Write twenty real examples before you write a prompt.
in the PDF
- 05
Decide what a wrong answer costs.
in the PDF
- 06
Define the handoff before the happy path.
in the PDF
- 07
Decide what it must never do, in writing.
in the PDF
- 08
Plan how you will know it is getting worse.
in the PDF
- 09
Work out what it costs at ten times the volume.
in the PDF
7 more checks in the PDF
The complete checklist — every check with the reasoning under it, the closing note, and the other 12 documents — comes as one PDF.