# The first seven tests for an app you built with AI, in the order to write them

> Which journeys to test first in an app built with Lovable, Bolt, Cursor or Claude Code, why in that order, how to write each one so a tool or a coding agent can run it, how often to run them, and what to add in the first week after launch.

Published 2026-11-18, about FlowQA.

An app built with an AI tool needs about seven tests before strangers start using it: the app opens, sign up, sign in, the main action, payment if you charge, password reset, and the one failure that would embarrass you.
Each is an end-to-end test, which means a browser opens the running app, does what a person would do, and checks what the page shows.
Write them in that order: the earlier ones break the app for more people when they fail, and the later ones start from them.

The checks you do by hand before launch day are in [How to test a vibe-coded app before launch](https://testfinch.com/blog/how-to-test-a-vibe-coded-app-before-launch/); the seven tests here keep running after it.

## What is a test, if you have never written one?

A test is a list of steps and one or more checks.
The steps are what a person does: open this page, type this, press that button.
A check is a statement about what the page shows afterwards, such as "the page shows Welcome, Ada".
A program runs the steps in a browser and reports pass when every check holds, or fail with the step that went wrong.

Names you will meet for groups of tests:

- A smoke test is a short test that the app starts and its most important page works. It runs first, after every deploy.
- A regression test checks that something that worked before still works after a change. All seven tests in this post are regression tests once they exist.
- A test suite is the whole set, run together.

A journey is one trip through the app that starts where a person starts and ends at something they came for.
Each of the seven tests covers one journey.

## Which seven journeys should you test first?

| Order | Journey | Check at the end | Why it has this place |
|---|---|---|---|
| 1 | The app opens | The home page shows its main heading and the sign-in link | A bad deploy or an expired key breaks every page at once, and this is the shortest test to write |
| 2 | Sign up | A new account lands on the first screen after sign-up | Nobody can become a user without it |
| 3 | Sign in | An existing account reaches its dashboard and sees its own name or data | Every later test starts by signing in |
| 4 | The main action | The thing the person made appears in their list, and is still there after a reload | The reason the app exists |
| 5 | Payment, if you charge | After paying with the payment provider's test card, the account page shows the paid plan | A broken payment page loses money without anyone reporting a bug |
| 6 | Password reset | Asking for a reset shows "check your email" and no error | A person who cannot get back in cannot use the app |
| 7 | The one that would embarrass you | Depends on the app: see "Choosing test 7" | You named it as the worst thing a customer could see |

If the app takes no payment, test the second most used action in place of test 5.

### Choosing test 7

Finish this sentence: "It would be awful if a customer saw...".
Some answers for an AI-built app:

- another customer's record, when they open a link that is not theirs;
- a page of sample data or placeholder text the builder left in;
- an error page with a stack trace on it;
- a public share link that stops working after the owner signs out.

Pick one and write the test that would catch it.
For the first answer, the test signs in as a second account, opens the address of the first account's record, and checks that the page shows "not found".

## How do you describe a journey so a tool or an agent can run it?

Write four things for each journey: where it starts, the steps, the values to type, and the outcome to check.
A coding agent or a testing tool cannot guess any of the four.

> **Create a booking**
>
> - Start: /dashboard, signed in as the test account.
> - Steps: press New booking; fill Customer name; choose tomorrow's date; press Save.
> - Values: a customer name with the current time in it, so each run makes a new one.
> - Outcome: the bookings list shows the customer name, and still shows it after the page is reloaded.

Name buttons and fields by the words on the screen, such as "the Save button".
Keep the test account's password out of the description and refer to it by name, such as "the test account's password".
State the outcome as something that would be missing if the feature broke: "the page loads" is true of an error page too.

[How to write a test plan an AI agent can execute](https://testfinch.com/blog/writing-a-test-plan-a-model-can-execute/) goes through the four parts in detail.
If you would rather start from a filled-in sheet, the free [smoke test template](https://testfinch.com/tools/test-case-template/smoke/) and the [sign-up](https://testfinch.com/tools/test-cases/signup/) and [login](https://testfinch.com/tools/test-cases/login/) packs list rows to start from.

## How often should the tests run?

| When | What runs | What it catches |
|---|---|---|
| After every deploy | All seven | A change that broke a page you did not mention in the prompt |
| Every night | All seven | Failures with no deploy behind them: an expired API key, a payment provider's change, a full database |
| Before you push, on your own computer | Tests 3 and 4 | A broken main action, before it reaches the deployed app |

Seven tests are few enough to run all of them after every deploy.
A deploy is any time the app that people use changes, which in an AI builder is each time you publish.
A failed run has to reach you where you will see it, by email or in a chat channel, with the step that failed.

## What do you add in the week after launch?

The seven tests cover what you expected people to do, and the first week brings bugs from what they do instead.

- Turn every bug a user reports into test number eight, nine and ten.
  The steps that led to the bug are the test's steps, and the fixed behaviour is the check.
  A bug with a test cannot come back unnoticed.
- When a test fails, decide which changed: the app by accident, or the page on purpose.
  If you renamed a button on purpose, update the test.
  If you did not, the test found a bug before a user did.
- A test that fails on some runs and passes on others with no change in between is called flaky.
  Fix it or remove it in the same week, before you start ignoring its failures.
  [Flaky tests: causes and triage](https://testfinch.com/blog/flaky-tests-causes-and-triage/) lists the usual causes.
- After a month, look at which tests have failed.
  A test that has never failed in a month of changes may be checking too little, so read its check again.

## What these tests will not catch

Seven journeys walked by a browser do not look for one customer's data in another's account beyond the one case in test 7.
They do not try blank forms, very long text or a double click on Pay, and they do not judge whether a page looks right.
Do those checks by hand before launch, as [the pre-launch walk-through](https://testfinch.com/blog/how-to-test-a-vibe-coded-app-before-launch/) describes.

## Where FlowQA fits

FlowQA by TestFinch writes and runs these tests for a web app.
Give it your app's address, and its AI agent explores the site and proposes a plan of journeys.
You keep the ones on your list and cross off the rest.
If you work with Claude Code, Cursor or Codex, your coding agent can write the seven from your descriptions through the TestFinch connector; [the setup walk-through](https://testfinch.com/blog/end-to-end-tests-from-claude-code-cursor-and-codex/) has every command.

Tests run on a schedule, from CI, or after each GitHub deployment, and a failure names the step and shows the page.
When a test fails because the page changed, FlowQA proposes one small repair with its evidence, and the test changes only after you approve.

The free plan covers 1 project, up to 100 tests with 10 AI-generated a month, 1,000 cloud test runs a month, 1 schedule and a CI trigger.
Count the runs before you pick a cadence: each test in each run is one cloud test run on each browser and screen size it runs on, so seven tests every night on one browser come to about 210 in a 30-day month and every deploy adds seven more.
The [pricing page](https://testfinch.com/pricing/) has an estimator.

FlowQA tests web apps in Chromium, with phone and tablet sizes emulated.
A password reset link that arrives by email, or a payment widget from another company, is a journey you record by hand in FlowQA Desktop.
If you want someone else to plan and build the tests, [our team sets them up for you](https://testfinch.com/service/).
