How to Debug Code with AI

In short: reproduce → collect error lines → ask for one hypothesis → apply a small fix → re-run tests. Never paste secrets or .env files into the chat.

Why AI helps debugging

Reading a long stack trace is slow when you are tired. An AI is good at pattern-matching common mistakes (wrong type, off-by-one, float money math). You stay in charge: it guesses, you verify.

What to give the AI

• Failing test name: so it knows the expected behaviour.
• Error / stack lines: the exception type and the 5–15 lines that point at your file.
• The small function: not the entire repository.
• What you already tried: one sentence, to avoid looping the same bad fix.

How debugging flows

Example: GST total was wrong

# Common AI mistake: float money math
def gst_total(price, rate_percent):
    return round(price + price * rate_percent / 100, 2)
# 99.99 * 0.18 in float can surprise you — use Decimal for money.

"""Indian GST on a price. Rate is percent, e.g. 18 for 18%."""
from decimal import Decimal, ROUND_HALF_UP


def gst_total(price: str | Decimal, rate_percent: str | Decimal) -> Decimal:
    p = Decimal(str(price))
    r = Decimal(str(rate_percent))
    if p < 0 or r < 0:
        raise ValueError("price and rate must be >= 0")
    tax = (p * r / Decimal("100")).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
    return (p + tax).quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)

from decimal import Decimal
import pytest
from gst import gst_total

def test_18_percent():
    assert gst_total("100", "18") == Decimal("118.00")

def test_rounds_half_up():
    assert gst_total("99.99", "18") == Decimal("117.99")

def test_zero_rate():
    assert gst_total("50", "0") == Decimal("50.00")

def test_negative_raises():
    with pytest.raises(ValueError):
        gst_total("-1", "18")

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

How to debug step by step

1. Make it fail the same way twice with a test or a 5-line script.
2. Copy only the useful error lines (file, line, exception). Drop huge HTML dumps.
3. Ask for one hypothesis, not five random rewrites.
4. Apply the smallest patch. Reject drive-by refactors.
5. Re-run the same test. If it still fails, paste the new error only.
6. Start a new chat when the thread is full of wrong guesses.

Where to debug in real tools

1. VS Code + Copilot / Claude Code / Cursor: select the function, open chat or inline chat, paste the error.
2. Agent mode: allow it to run the test command; keep approval on for shell.
3. Terminal CLI (Claude Code): run from the project root so it can see the failing test file.
4. Settings: put the test command in project rules so you do not retype it.

Quick quiz

Save tokens and money

Send the error and the function, not megabytes of logs. One hypothesis per turn. Prefer inline chat for a single function; save agents for multi-file bugs.

Save these cards