Create a test from a prompt

Tests made from a prompt

Describe a workflow to your AI assistant, Claude, ChatGPT or a local AI model we provide or you bring, and it records the test for you: it opens the app or the web page on your computer, performs the steps as real clicks and keys, verifies as it goes and saves an ordinary GUI Test Studio test.Open it in GUI Test Studio, run it in GUI Test Runner-Recorder, add it to your pipeline with gtr-cli: nothing about the test says a model made it.

It is the second way to create a test, beside recording it yourself: the result is the same test either way.The assistant also edits an existing test in place, creates one from a Figma design, runs a test and shows you its Confidence Report.

It works in Claude Desktop, in Claude Code (the terminal, and its VS Code and JetBrains extensions), in ChatGPT with Codex (the ChatGPT desktop app, the Codex command line and Codex's VS Code extension) and in VS Code's Copilot Chat.With no assistant at all, a local AI model runs on this computer, one we provide in GUI Test Studio's built-in engine or one you bring, and creates the test with the very same tools.The connector is installed with GUI Test Studio; it needs your account signed in on this computer, on the free trial or a paid plan, and nothing else: no key, no setting.This page says Claude; ChatGPT and a local model do the same.

The Create a Test From a Prompt window

In GUI Test Runner-Recorder, click Create Test From Prompt: it stands beside Record on the home page, in the Create a new test card, and in the toolbar.Either app also has it under Help, Use with Claude or ChatGPT....The window has four parts, from top to bottom.

Enlarge

1. What should the test do?

Write what the test should do and verify, the way you would brief a tester.The box is all the assistant gets, so name the app or the address, the steps and the outcome.

2. Or start from an example

Six dashed tiles fill the box with a ready-made request: Whole app, One workflow, Browsers & devices, Run a test, Change a test and From a Figma design.For the first two, the For row says what you test: a desktop app, a web app or an app on your phone.Every example opens with "With GUI Test Studio, ..." and marks what to replace between « »; the first part is selected, so you only type over it.

3. Run this prompt with your AI assistant

One card per assistant: Claude Desktop, Claude Code, ChatGPT & Codex and VS Code, each with its state and one button that does what comes next.

Claude Code is one card for the terminal and for its VS Code and JetBrains extensions: they run the same program and share the connection.ChatGPT & Codex is one card too: the ChatGPT desktop app, the Codex command line and Codex's VS Code extension read one setting, so one Connect serves the three; ChatGPT on the web and on a phone cannot reach a program on your computer.The VS Code card is Copilot Chat's agent mode.

Any other assistant that takes a connector works too: Copy the Connector's Path, under the cards, copies the one command the assistant needs; paste it where it asks for a connector.On another computer, or in a terminal of your own, this one command does the Claude Code setup (the window's "?" shows it):

claude mcp add --transport stdio --scope user gui-test-studio -- "C:\Program Files\GUI Test Studio\Runner\gts-mcp.exe"

And this one the ChatGPT and Codex setup (the ChatGPT app reads it when it starts):

codex mcp add gui-test-studio -- "C:\Program Files\GUI Test Studio\Runner\gts-mcp.exe"

For one project only, a .mcp.json file at the project's root does the same for Claude Code, and a .vscode/mcp.json file for VS Code.

4. Or create it on this computer

The This computer card, under the assistants, creates the test with a local AI model, one we provide or one you bring: Settings... sets it up once, Create Here starts.See the next section.

On this computer, with a local AI model

No AI assistant needed: the This computer card creates the test with a local AI model, one we provide in GUI Test Studio's built-in engine or one you bring through LM Studio or any OpenAI-compatible local server.Your request and your screen never leave the computer, and there is no account with an AI company and no cost per test.The model gets the very same tools Claude and ChatGPT get, so the test is the same kind of test.

Set it up once

Click Settings... on the This computer card.The window says what to expect, shows this computer's graphics card and memory, and recommends the model that runs best on it.Click Set up for me, or pick another model and click Download: it is downloaded once, from its publisher, and kept for every test.

Enlarge

These choices are separate from AI Smart Verification's: one model can create your tests here while another one, or a cloud provider, explains failed checks.

Create the test

Write the prompt, then click Create Here.A card in the corner of the screen names each step as the model takes it.Keep your hands off the mouse and the keyboard meanwhile: the card steps aside while the model clicks and types, and it never shows in the test's pictures.At the end it says where the test was saved and whether it passed, with Open in GUI Test Studio and Show in Folder.

Enlarge

What to ask

Say what the test should do and what it should verify, the way you would brief a tester:

Open the calculator, compute 5 + 5 and verify the result shows 10.

Create a test for my app C:\Apps\Shop.exe: add a client named Atlas Textiles and verify it is listed.

Test the sign-up form at https://shop.example/signup in Chrome, Edge and Firefox and on an iPhone 15.

Run the Checkout test and tell me what failed.

With GUI Test Studio, test the whole of C:\Apps\Shop.exe: list its main workflows and create one test per workflow, each verifying its outcome.

Show me the report of the Checkout test.

In the Checkout test, rename the check "Visual Assert #3" to "Pay button" and let it wait 10 s.

Create a test from my Figma design C:\Designs\Login.gtsdesign.

Open with "With GUI Test Studio, ..." when the app you ask could do it another way: the common tasks of the window word every request so.For an app on an Android phone plugged in by USB (USB debugging on), say "on my phone": the test shows the phone in a window on this screen and Claude works through that window.

Every task ends the same way: Claude tells you where the test was saved and the verdict of its run, and asks what comes next, whether to run it again, edit it in GUI Test Studio or open it in GUI Test Runner-Recorder to read its Confidence Report.A run that fails is explained from the report, never answered by recording the test again, unless you ask for that.

Say "edit the test in GUI Test Studio" and Claude opens it there; "open it in GUI Test Runner-Recorder" opens the window where you run it yourself; "show me the report" opens the test there with its newest Confidence Report on screen.The prompts menu of your Claude app (the "+" in Claude Desktop) lists the common asks - create a test, test a web page in several browsers, test an app fully, run, edit, open in GTR - each taking a short line; details go in the chat.

Claude does exactly the workflow you asked for.It explores an app on its own only when you ask it to ("test all the basic features of this app"), and then creates one test per workflow it finds, each with its own Scenarios and checks.

What the assistant does on your screen

Keep your hands off the mouse and the keyboard while Claude works: it uses the real screen.It starts the app or opens the page, looks at the screen, clicks and types, and after each step checks the result the way the recorder would.The time Claude spends thinking between two steps is left out of the test; a pause Claude asks for on purpose becomes a Wait step.

The test lands in Documents\GUI Test Studio unless you name another folder, in a folder of its own, with its results beside it.

Edit an existing test by prompt

Claude can change a test you already have without recording it again: rename a step, change the text a check expects, give a check a longer timeout, decide what happens if it fails, remove or ignore a step, add a wait, change the test's name or description.Name the test and say what to change:

In the Checkout test, let the "Order confirmed" check wait 10 s and continue the test if it fails.

In the Sign in test, ignore the step that clicks "Remember me", and add a 2 s wait before the Sign in click.

Rename the Checkout test to "Checkout with a coupon" and describe it as the coupon scenario.

It can also change where a check looks, how strict a picture comparison is and what the AI is told about a check, move a step, group steps into a Scenario, and add a Test Flow step or a Marker.A new check is tried first on the test's own recording at that moment, and refused if it could not pass there.

In the Checkout test, also verify that the total is shown after step 12.

In the Checkout test, group steps 3 to 9 as a Scenario called Payment.

Before it writes anything, Claude shows what it is about to change and asks you to confirm, every time: an edit overwrites the test file.Recording the whole test again happens only when you ask for it.Afterwards, open the test in GUI Test Studio to see the change on the timeline: an edit by prompt is an ordinary edit of the test, and the next run uses it.

After a run: why it failed, and what to do about it

Ask what failed and Claude reads the run's own results, without running anything: what each check expected and what the run found, with the pictures and every difference outlined, and what AI Smart Verification said.It also reads the test's history: since when it fails, and whether it goes back and forth.

Why did the Checkout test fail? Show me the difference.

When did the Checkout test start failing? Is it flaky?

The new button is correct: accept what the last run found.

File this failure as a Jira bug in project SHOP.

Run the Checkout test from step 12 on, and retry it once if it fails.

Put the Checkout test in TeamCity.

Claude decides nothing for you.Accepting a result makes what the run found correct from then on, and a Jira issue is read by other people: both happen only after you confirm what Claude shows you.A PDF or an Excel report is written at once.

A test from a Figma design

Give Claude the design package your designer exported with the suite's Figma plugin, a .gtsdesign file, and it creates a test from it through the same import GUI Test Studio offers: every screen becomes a reference picture, every element a check, and a prototype flow becomes the click path.A login page design gives a login scenario test, ready before the feature is built.

Create a test from C:\Designs\Login.gtsdesign, following its prototype flow.

The design is exported from Figma as Import a design from Figma explains; what the test can do before the feature exists is in Test a feature before it exists.

Web apps: browsers, devices, conditions

For a web page, ask for the other browsers, devices and conditions in the same sentence ("in Edge and on an iPhone 15, in French and at 150 % zoom"): Claude records the page once, generates the test for every combination once it is saved, and can run each one.The recording runs with a browser window, as it was recorded; the generated browsers and devices run without one.

A saved web test can ask for more browsers and devices later, and generate them again after the page changed.Generating one again replaces what it holds, so Claude asks first.

Every browser and device from one recording

What the assistant can and cannot do

If something does not work

The connector's own log is %LOCALAPPDATA%\GUI Test Studio\logs\claude-connector.log.