Testing as Spec With Claude Code

Context

Claude Code can generate 50 tests in seconds. The problem isn't quantity: it's whether those tests validate real behavior and read like a specification, not mechanical code that always passes. Using Claude Code for testing means steering the generation with judgment, not delegating the judgment.

Setup

1. Have the test runner configured in the project (Vitest, Jest, Pytest, whatever you use) before you start.
2. Give Claude Code context: which function, what behavior you expect, which runner. The test command and conventions ideally live in CLAUDE.md.
3. Golden rule (from Claude Code best practices): give it a way to verify its work. Have it run the tests and read the result, not just make them "look" right.

State as of July 2026. Claude Code evolves; commands and flags may change.

Examples

A prompt that produces tests-as-spec:

Generate tests for canUserCheckout(user, cart) with Vitest.
Make each test read like a business sentence, not "test 1, test 2".
Cover: verified / not verified, with / without pending orders,
empty cart, user with no credit limit.
Then run the suite and show me the output.

Expected result — the names are the spec:

describe('canUserCheckout', () => {
  it('allows checkout when verified and no pending orders', ...)
  it('blocks checkout when the user is not verified', ...)
  it('blocks checkout when there is an unpaid pending order', ...)
})

If Claude Code hands you "100 passing tests" but none break when you deliberately break the code, they're worthless. Ask it to make the tests fail when a bug is introduced, as proof they validate something.

Particularities

• Combines with edge-cases-with-ai: first you ask it to discover the cases, then to turn them into tests. Discovering and covering are two distinct steps.
• Real verification: require it to run the suite and show the output. "Looks done" isn't done — it's the central principle of Claude Code best practices.
• Coverage ≠ quality: a high coverage number with tests that assert nothing is smoke. Prioritize critical behavior over percentage.
• Limitation: Claude Code writes correct tests, but you decide WHAT is worth testing. That judgment isn't delegated.