Specs and tests as the contract

Write the requirement first, turn it into a failing test, and let the agent drive it to green - including the edge cases it would otherwise skip.

The requirement comes first

A test written before the code is a specification the agent cannot negotiate with. A test written after is a description of whatever the agent happened to build, which is a very different document and much less valuable.

Requirement: a discount code applies once per customer.

Acceptance criteria:
- Applying a used code returns 409 with code ALREADY_REDEEMED.
- Applying an expired code returns 409 with code EXPIRED.
- Applying an unknown code returns 404.
- The discount is applied to the subtotal, before tax.
- Codes are case-insensitive and trimmed.

1. Write the criteria before the prompt. Half of them will be decisions you had not actually made.
2. Turn each criterion into a failing test - or ask the agent to, and then read the tests before any implementation exists.
3. Only then ask for the implementation, with the failing test as the target.
4. Run the suite and require real output, not a claim.
5. Delete any test the agent added that does not correspond to a criterion.

Red to green

import { test, expect } from "vitest";
import { applyDiscount } from "./discount";

test("rejects a code that has already been redeemed", () => {
  const result = applyDiscount({ code: "SAVE10", redeemedBy: ["user_1"] }, "user_1");
  expect(result).toEqual({ ok: false, error: "ALREADY_REDEEMED" });
});

test("trims and lowercases the code before lookup", () => {
  const result = applyDiscount({ code: "  save10  " }, "user_2");
  expect(result.ok).toBe(true);
});

npm test -- discount.test.ts
# FAIL  src/discount.test.ts
# Error: Cannot find module './discount'

# now, and only now:
> "Implement src/discount.ts so these tests pass. Do not modify the tests."

• The red step proves the test can fail. A test that passes before the code exists is testing nothing.
• 'Do not modify the tests' is worth saying out loud; without it the shortest path to green may run through the assertions.

Edge cases agents skip

"Now add tests for the cases you would expect to break:
two simultaneous redemptions of the same code, a code with leading
whitespace, and a user who has never redeemed anything.
Report any case where you had to guess the correct behaviour."

That last instruction is the useful one. Cases where the agent had to guess are exactly the cases where the requirement was underspecified, and they are cheaper to settle now than after the code is deployed.

# a mutation check: do the tests actually bite?
npx stryker run --mutate src/discount.ts

# or simply break the code by hand and confirm a test fails
# if nothing fails, the test is decoration

FAQ

Related

The generate, run, correct loop Types, linters and static checks as the first filter

Last refreshed 2026-09-18.