Write an AI change brief your WordPress teammate can approve
“Fix the site” is a poor change request, whether you give it to a developer or an AI assistant. The person approving the work cannot tell which files might change, what counts as success or when the operator should stop.
A useful brief gives the reviewer something smaller to approve. Start with a visible problem and an exact target.
A brief for a real review
Imagine a client asks to change the enquiry button on one landing page. This is a hypothetical task, not a customer result. A usable brief could read:
“Inspect the enquiry button on /services/. Explain where its label and link are configured. Do not change anything yet. Propose a change from ‘Get started’ to ‘Request a quote’, keeping the existing destination, styling and tracking attributes. Do not edit other buttons, shared templates or form settings.”
That first step is inspection. After reading the proposal, the reviewer can authorize a bounded edit or explain why a shared component needs a different approach. A request to inspect should not quietly become permission to deploy a replacement template.
Put verification in the request
For this example, completion means the new label appears at the intended URL on desktop and mobile, the link destination is unchanged, and unrelated buttons retain their original text. If the button belongs to a shared template, the operator should stop and report that before editing it.
Those checks are more useful than “make sure it looks good.” They also help a teammate review the result without replaying the entire conversation.
When the task involves code, describe the behavior rather than prescribing a guessed implementation. Include the observed error and a way to reproduce it. Remove passwords and personal customer data from any examples you attach.
Approval does not replace recovery
Before a significant mutation, identify an appropriate recovery point and what it actually covers. A file rollback cannot reconstruct a deleted database record. A full database restore may overwrite orders placed after the backup. The reviewer needs that distinction before saying yes.
IrmaWP’s team approval workflow and operation history keep the request and its outcome connected. Available actions still depend on the site’s connection, permissions and safety policy. An AI-generated patch is a proposed implementation, not evidence that the site works.
After execution, record the check you performed and any uncertainty. “Label changed; destination checked; mobile layout verified” is useful. “Done” leaves the next operator guessing.
Have a correction or a workflow we should cover? Tell the IrmaWP team.