Automated test case generation with Claude Code

The setup

You need Claude Code with the E2Easy MCP connector enabled. See the setup guide if you haven't connected yet — it's a one-time 5-minute step.

From a Jira ticket

Paste the ticket into Claude Code:

// PROJ-1284: New tenant onboarding wizard
// As a new admin, I should be able to:
//   1. Sign up with company email
//   2. Verify the email
//   3. Pick a workspace name
//   4. Invite at least one teammate
//   5. Land on the dashboard with welcome banner

Generate a test for this against staging.acme.com.

Claude Code reads the ticket, calls inspect_page on each onboarding screen to gather selectors, composes the steps and asks you to confirm before create_test saves it. Total: 17 steps, 4 assertions, ~$0.07 in LLM tokens.

From a Figma flow

If your design team ships flows in Figma, paste the link plus the staging URL:

Here's the new pricing-page flow in Figma: [link]
Add E2E coverage that matches it on staging.acme.com.
Don't break anything else.

Claude reads the design, lists existing tests with list_tests to avoid duplication, drafts the new test and confirms before saving.

From a bug report

Bug reports often describe the regression in plain English ("clicking the X breaks the form"). Use them as test seeds:

Bug #418: Submitting the contact form with non-ASCII
characters returns 500 instead of validating client-side.

Add a regression test so this doesn't ship again.

Claude builds a targeted test that fills the form with cyrillic + emoji, asserts the validation message client-side and that no 500 leaks through. create_test tags it with the bug ID for later traceability.

What Claude Code does NOT do well

Honest limits:

• Implicit business rules — if "valid" means something only your team knows, write it down in the ticket. Claude won't infer.
• Multi-tenant data isolation — tell Claude which tenant ID / fixture to use; otherwise it'll pick one and stick with it.
• Visual regression — generation works for behavioral assertions; pair with a visual tool for pixel-perfect checks.

Cost vs writing the test by hand

For a typical 17-step generated test:

• Generation cost: ~$0.07 in LLM tokens, takes ~90 seconds end-to-end.
• Equivalent hand-written test: 1-3 engineer hours at $80/hr = $80-240.

Quality is comparable for behavioral coverage. The big win is cycle time — tests get added in the same PR that ships the feature, instead of in a follow-up sprint.