From user story to Cypress test in 60 seconds

The average QA engineer writes a Cypress test in 2–4 hours. With a well-built AI workflow it takes 60–90 seconds— not because AI is faster than a tester, but because it removes the repetitive boilerplate work.

This article is a hands-on guide. We use Claude Code, but the principle works with Cursor or Copilot Chat too.

Prerequisites

• Existing Cypress + TypeScript project
• Page Object pattern (Claude follows it)
• CLAUDE.md with project conventions (see the previous article)
• A clearly written user story

Step 1: Prepare the user story

A bad user story = a bad test. This is the format that works:

AKO registrovaný používateľ
CHCEM pridať produkt do obľúbených
ABY som ho vedel nájsť pri ďalšej návšteve.

Acceptance:
- Klik na srdiečko v produkt detail pridá produkt do obľúbených
- Návrat na /favorites zobrazí pridaný produkt
- Klik znovu na srdiečko produkt odstráni
- Neprihlásený používateľ dostane modal "prihláste sa"

When AI has concrete acceptance criteria it can write 4 test cases — not 1 generic one.

Step 2: Prompt

> Napíš Cypress test pre túto user story.

[paste user story sem]

Context:
- Test go do cypress/e2e/favorites/
- Použij existujúci productPage a favoritesPage (ak existujú, inak
  ich vygeneruj v cypress/support/pages/)
- Testuj aj scenar neprihláseného používateľa (očakávaj modal s
  data-qa="auth-required-modal")
- Pre login použij custom command cy.loginAs('standard')

Step 3: Validate the output (60 seconds)

AI should be fast, not limitlessly trusting. A junior reviewer can do these 3 checks in a minute:

1. Run it — npx cypress run --spec 'cypress/e2e/favorites/*.spec.ts'. If it fails on a selector, AI hallucinated.
2. Acceptance coverage — open the spec and count: are there 4 it() blocks for 4 acceptance criteria? If not, ask again.
3. Negative cases — does it test the logged-out scenario? Does it test double-click (removal)?

Common AI mistakes when generating tests

• Invented selectors — AI sees data-qa="login-btn" in a previous test and assumes the product page has the same pattern. Validation = actually open the page and look at the DOM.
• Missing beforeEach() — AI sometimes omits cleanup, which leads to flaky tests under parallel runs.
• Hardcoded cy.wait(2000) — a classic. Even with a good CLAUDE.md it sometimes slips in. Stop & fix.
• Tested happy path only — if acceptance includes a negative branch, AI sometimes omits it. Ask for it explicitly.

A meta-prompt that rules out 80% of issues

Paste this at the start of every generation prompt:

Pravidlá:
- Pred napísaním testu otvor relevantné page objects a over selektory
- Žiadne cy.wait(ms). Používaj cy.intercept() alebo .should()
- Pre každé acceptance kritérium 1 it() blok
- Zahŕň negatívne scenáre, ak sa logicky ponúkajú
- Použi existujúce factories a fixtures namiesto inline dát
- Ak niečo nevieš z kódu, pýtaj sa — nehalucinuj

The difference in output quality is noticeable.

When AI isn’t enough

If the user story describes a flow with critical business logic (payments, regulation, GDPR), don't use AI as the author of the test — use it as a helper. Write the skeleton yourself, AI fills in Page Object calls and assertions.

---

Want the same approach at your company? Get in touch — dohodneme 30-minute discovery call.