How to Prompt an AI for Code (That Actually Works)

In short: Goal + Language + Input/Output + Rules + Example. Ask for small pieces. Ask for tests too.

Why prompts matter

The AI cannot read your mind. A vague prompt gives vague code. Then you spend more time and more tokens fixing it. A clear prompt saves both.

What a good code prompt has

• Goal: one sentence: what the code must do.
• Language and tools: Python 3.12, Node 20, Java 21, “no extra libraries”.
• Input and output: the types and shape, e.g. “takes a price (float) and coupon (string) and returns a float”.
• Rules / edge cases: what to do on empty input, negative numbers, bad values.
• Example: one input and the exact expected output.
• Format: “only the code” or “code + a short explanation”.

How a prompt turns into working code

Bad prompt vs good prompt

The code the AI wrote (tested here)

def final_price(price: float, coupon: str | None = None) -> float:
    """Return price after coupon. SAVE10 = 10% off, FLAT50 = 50 off. Never below 0."""
    if price < 0:
        raise ValueError("price cannot be negative")
    if coupon == "SAVE10":
        price = price * 0.9
    elif coupon == "FLAT50":
        price = price - 50
    elif coupon is not None:
        raise ValueError(f"unknown coupon: {coupon}")
    return round(max(price, 0), 2)

import pytest
from discount import final_price

def test_no_coupon():
    assert final_price(200) == 200
def test_save10():
    assert final_price(200, "SAVE10") == 180
def test_flat50_never_negative():
    assert final_price(30, "FLAT50") == 0
def test_bad_coupon():
    with pytest.raises(ValueError):
        final_price(100, "FREE")

pip install pytest
pytest -q
# -> 4 passed

How to write prompts, step by step

1. Start with the goal in one sentence.
2. Name the language, version and libraries.
3. Give the function name and inputs/outputs.
4. List the edge cases (empty, zero, negative, wrong type).
5. Add one example with the exact output.
6. Say the format: “only the code”.
7. Ask for tests in a follow-up, then run them.
8. Fix with a short follow-up: paste only the failing test and error lines.

Same pattern in other stacks

• JavaScript: “Write a Node 20 function slugify(title), no npm packages, return lowercase words joined by dashes. Example: “Hello, World!” → “hello-world”.”
• Java: “Write a Java 21 static method isStrong(String pw): at least 8 chars, one upper, one lower, one digit. Null returns false. Add JUnit 5 tests.”
• SQL: “PostgreSQL. Table orders(id, customer_id, amount, created_at). Write a query for total amount per customer in 2026, highest first.”

Settings and config: save your rules once

Most tools read a project file with your standing rules. Then you do not repeat them in every prompt.

1. GitHub Copilot: create .github/copilot-instructions.md in your repo. Write rules like “Use Python 3.12, pytest, type hints”.
2. Claude Code: create CLAUDE.md in the project root (or run /init to make one). Put the same rules there.
3. Keep it short. These files are sent with requests, so a long file costs tokens every time.

Quick quiz

Save tokens and money

Ask for one function at a time. Add “only the code”. Paste only the lines that matter, not the whole file. Put standing rules in a short instructions file instead of retyping them.

Save these cards