Working effectively with AI-assisted delivery

Provide context an implementation can use

State the intended outcome, the users, the permitted inputs and the constraints. Supply approved examples and explain which sources are authoritative. If information is missing, ask for the uncertainty to be surfaced rather than filled with an assumption.

An example of a useful request

“Add a request review step for coordinators. A coordinator can approve or return a request for correction. A return requires a reason. A submitter can see the reason but cannot approve their own request. Demonstrate approval, return, resubmission and a denied self-approval attempt using synthetic records.”

This request gives the implementation and the reviewer the same target. It also names the important exception and a way to verify it.

Keep work small enough to review

• Ask for one coherent workflow or change at a time.
• Separate a proposed approach from permission to change production.
• Review the plan when the request affects architecture, access or external systems.
• Use test results and working behavior to assess completion.
• Escalate repeated uncertainty to the owner of the missing information.

Treat generated material as reviewable work

Generated code, explanations, tests and documents can contain errors or omit important cases. Review them against authoritative requirements and actual behavior. If a result is wrong, record the discrepancy, correct the relevant input or implementation, and rerun the affected checks. A confident explanation is not a substitute for a passing test.