Cursor Not Following Instructions? How to Verify What It Changed

You gave Cursor clear instructions. It still changed something else.

AI coding agents can sometimes reinterpret instructions, make adjacent changes, or complete only part of a task. The chat can look finished while the diff still includes an edit you did not ask for, or leaves a constraint you stated unfinished.

Common signs the agent did not follow your instructions

• A requirement was skipped
• A file changed that you did not ask to touch
• A public API changed despite being out of scope
• The implementation drifted from the original plan
• The agent reported completion even though part of the task is missing

Why rules and prompts are not the same as verification

Rules tell the agent what should happen. Verification checks what actually happened.

Rules, prompts, AGENTS.md, and hooks can guide an AI coding agent, but they do not guarantee that every final change matches the original task.

Why asking the same AI to review its own work may not be enough

The same agent may review the result using the same assumptions it used while implementing the change. That second pass can confirm the interpretation it already followed.

An independent review that starts from the original task and compares it against the actual Git diff is a different type of check.

A better way to verify the result

Example

How ChangeVerify helps

ChangeVerify compares:

It helps identify:

• Missing requirements
• Unexpected changes
• Scope creep
• Possible plan deviations
• Human review priorities

ChangeVerify does not prove runtime correctness, security, or production safety.