Skip to content

Privacy and control

How to review an AI action before it changes an app

Check the destination, exact effect, and permission scope before an AI companion clicks or types. Understand why uncertain actions should not repeat.

··3 min read

Explaining a button and pressing it are different tasks. When a companion can interact with an app, the useful question is not only whether it understands your request, but whether the exact action still matches the current screen.

Observe → review → checkIllustrative diagram
From intent to a checked actionThis illustrates a selected-window action. Confirmation approves the reviewed effect; a fresh observation is still needed to inspect the result.Selected windowDraftObserveReviewAction + destinationAllow onceCheck✓Fresh observationUncertainty is not a reason to repeat
  1. 01
    Inspect the actual window

    The action should refer to the selected app and current visible state.

  2. 02
    Review the effect

    Check what will change and how broadly you want to authorize it.

  3. 03
    Verify the result

    After one action, inspect the app again instead of assuming the request succeeded.

This illustrates a selected-window action. Confirmation approves the reviewed effect; a fresh observation is still needed to inspect the result.

Read the effect, not just the button label

An action review should make the consequence understandable. “Type this text into the draft” is different from “send the draft.” “Select these files” is different from “delete them.” Before approving, check the destination and the exact content when either matters. A correct sentence sent to the wrong place is still the wrong action.

Rayito’s app-control workflow alternates observation, planning, explanation, an action, and another observation. That sequence is useful because a plan can become stale while you read it. A document can change, focus can move, or a different conversation can become selected. The review belongs to the current situation rather than an abstract command that remains valid forever.

Choose a scope that matches your task

Approving one action is a useful starting point for an unfamiliar workflow. A task-level permission can reduce repeated prompts while completing the request you are already discussing. App-level and broader saved permissions cover more situations, so they deserve a more deliberate choice. The point of these options is to express your preference, not to make every person choose the broadest access.

A wider permission does not remove the need to inspect the target or verify the result. It also does not make hidden content trustworthy. If a webpage or document tells the companion to perform another action, that content is not a new instruction from you. Keep the task you actually requested as the reference when deciding whether a proposed change belongs.

Separate a refusal from an uncertain outcome

Sometimes an action is refused before input is sent. Sometimes input may have reached the app but the result cannot be confirmed. These situations need different explanations. A clear no-input refusal can be addressed by fixing the stated precondition. An uncertain outcome requires checking the app before considering another attempt.

Do not ask a companion to repeat a send or purchase simply because it did not report success. A missing confirmation is not proof that nothing happened. Look for the actual result in the selected app. Rayito is designed not to automatically replay an uncertain action; this protects the boundary between a lost response and a duplicate real-world effect.

A question to try

“Before doing that, show me exactly what will change and where.”

Keep a simple way to take over

You should be able to stop the active task and continue manually. If you change the selected app or the relevant content, describe the new state before resuming. Avoid layering a second request over a half-finished consequential action. A compact contextual card can keep the current question or approval visible while leaving most of the app unobstructed, which makes review easier than following a long running transcript.