Prompting

Executes tasks based on user input using natural language processing. This command is invoked when the user input does not start with a / command. The system interprets the input and attempts to carry out the task specified by the user.

Example Usage

> click sign up 
    
    thinking...

    To accomplish the goal of clicking "Sign Up," we need to
    focus on the Google Chrome application and then click on
    the "Sign Up" button.

    Here are the steps:

    1. Focus the Google Chrome application.
    2. Click on the "Sign Up" button.

    commands:
      - command: focus-application
        name: Google Chrome
      - command: hover-text
        text: Sign Up
        description: button in the header
        action: click

command='focus-application' name='Google Chrome'
command='hover-text' text='Sign Up' description='button in the header' action='click'

Prompting Tips

The agent is selecting the wrong thing!

This is the most common issues encountered with our agent, here are some possible reasons.

You're asking the AI to interact with elements it can not see

TestDriver uses the context from your prompt and the computer screen to make a decision of what commands to run. You should only prompt the AI to interact with elements it can currently see.

Incorrect Prompt

A common example of this is interacting with a dropdown. We often see users prompt the agent to interact with a dropdown and choose a state.

> Click on 'options' and select 'edit'

Recommended prompts

Instead, simply treat these as two separate prompts. This allows the UI to render and gives the AI the opportunity to parse the new screen data.

> click on options

> select edit

You're asking the AI to click on elements it does not understand

The relies on visual understanding, not functional. Like any user, the AI does not understand what the function of a button will be. It can only guess.

Incorrect Prompt

> click on the "new task icon"

Correct Prompts

> click on the "plus icon"

Describing Images Properly

If you're uncertain of how to describe an icon, simply ask ChatGPT-4o what it would call it, and use that as your input.

The AI can not find small images

Small, isolated images smaller than 15x15px appear like "noise" to the AI and may not be clickable. However, you can use the match-imagecommand to select these using manually made screenshots.

No matter what I do, TestDriver will not select my element.

The AI has trouble selecting some specific elements, like empty gray boxes, some substrings, or conditions where there is a lot of similar text close together.

If that's the case, you can always fall back to match-image. We've seen typically a test suite of 10 tests could require a single screenshot.