How to Write Tests with AI

In short: describe behaviour → ask for tests first or right after the function → run pytest / node --test / JUnit → keep only tests you understand.

Why AI-written tests help

Writing every edge case by hand is slow. AI is good at listing empty, zero, negative and “unknown input” cases. Bad AI tests just copy the bug. You must read assertions and run them.

What a good test prompt has

• Behaviour: what “correct” means in one sentence.
• Framework: pytest, node --test, JUnit 5, xUnit, cargo test…
• Cases: happy path + at least two edges + one error path.
• Rules: “no network”, “no real clock”, “use temp dirs” when needed.

How writing tests flows

Example: discount helper + pytest

def final_price(price: float, coupon: str | None = None) -> float:
    if price < 0:
        raise ValueError("price must be >= 0")
    if coupon is None:
        return round(price, 2)
    if coupon == "SAVE10":
        return round(price * 0.9, 2)
    if coupon == "FLAT50":
        return round(max(price - 50, 0), 2)
    raise ValueError(f"unknown coupon: {coupon}")

import pytest
from discount import final_price

def test_no_coupon():
    assert final_price(200) == 200.0

def test_save10():
    assert final_price(200, "SAVE10") == 180.0

def test_flat50_floor():
    assert final_price(30, "FLAT50") == 0.0

def test_unknown():
    with pytest.raises(ValueError):
        final_price(100, "BOGUS")

cd code/tests
python3 -m pytest -q
# -> 4 passed

How to get useful tests

1. Prefer tests that would fail if the code were wrong — not tests that only assert True.
2. Name tests after behaviour (test_flat50_floor), not test1.
3. Ask for table-driven tests in Go, or parametrize in pytest, when you have many similar cases.
4. For legacy code: ask for characterisation tests that lock today’s outputs before you refactor.
5. Run the suite in your IDE or CI — do not trust a chat that says “all passed” without a log.

Settings tip

1. Put the exact test command in project rules (pytest -q, npm test, mvn test).
2. In Copilot / Claude Code / Cursor Agent: “add tests and run them; stop if red.”
3. Keep fixtures small. Do not paste production databases into the chat.

Quick quiz

Save tokens and money

Ask for the test file once with a clear case list. Do not regenerate the whole suite after every tiny code tweak — update the one failing case.

Save these cards