AgentReceipt

Write prompts that lead to reviewable AI changes

By the AgentReceipt team · Updated 13 November 2026

How you ask decides how hard the result is to review. Narrow requests give small diffs.

Say what is in scope, and what is not

Name the files or the area, and say what to leave alone. “Change the label in the checkout button; do not touch other components.”

One task per session

Mixing a fix, a refactor and a feature makes one large change with three reasons. Do them separately and commit between them.

State what must not change

Tell the agent not to modify tests to make them pass, and not to add dependencies without asking. Then check the receipt to see whether it kept to that.