Write requirements
Start with the business outcome, then add rules and examples. Let your AI Agent read the project before proposing filenames, APIs, and implementation details.
Start with one concrete interaction
“Build an order system” leaves too much open: fields, editing rules, and the definition of done. A smaller first round gives you something you can inspect.
Look back at the requirement in your first feature. It states the entry point (an order list opened from the menu), the data (order number, customer name, amount), the operations (create and edit), the limits (no negative amounts, unique order numbers), the scope (no approvals or notifications yet), and the completion criteria (data remains after refreshing). That leaves your AI Agent room to follow project conventions and gives you something concrete to check. Write new requirements along the same lines.
Clarify rules that change behavior
You do not need a long specification immediately. Include details that affect the result:
“Own orders” might mean creator, assignee, or department. Ask your AI Agent to list unresolved questions and clarify them before implementation.
Give a valid and an invalid example
If the interface has specific requirements, add a reference image or describe button and field order. Otherwise, ask your AI Agent to follow the existing application style.
Explain what an update must preserve
Describing the current state controls scope better than resending the entire original specification.
Ask for verification results
Add this to the request:
Then review the output. For larger requests, break the feature down first.

