Debugging with AI — Reading Errors, Minimal Reproductions and Asking the Right Questions

"It's not working, fix it" + 400 lines of code is the most common way people ask AI for debugging help — and the least effective. Debugging is a process. AI makes each step faster, but it can't replace the process.

1. The debugging loop

1. Reproduce reliably (exact steps, input, environment).
2. Read the error message and stack trace.
3. Shrink to a minimal reproduction.
4. Form one hypothesis, test it.
5. Fix, then add a test for it.

2. Read the stack trace yourself first

import traceback

def load_marks(rows):
    return {r["roll"]: int(r["marks"]) for r in rows}

def class_average(rows):
    marks = load_marks(rows)
    return sum(marks.values()) / len(marks)

rows = [{"roll": 1, "marks": "78"}, {"roll": 2, "marks": "AB"}, {"roll": 3, "marks": "91"}]
try:
    class_average(rows)
except Exception as e:
    print("Traceback (most recent call last):")
    for frame in traceback.extract_tb(e.__traceback__):
        print(f'  in {frame.name}: {frame.line}')
    print(f"{type(e).__name__}: {e}")

Traceback (most recent call last):
  in <module>: class_average(rows)
  in class_average: marks = load_marks(rows)
  in load_marks: return {r["roll"]: int(r["marks"]) for r in rows}
ValueError: invalid literal for int() with base 10: 'AB'

(That's the core of the traceback Python prints — each frame with the function name and the line of code.) How to read it:

• Bottom line first: ValueError: invalid literal for int() with base 10: 'AB' — the what.
• Walk up the frames: the error happened in load_marks, called from class_average. The last frame in your code is usually where to look.
• Form a hypothesis: one row has "AB" (absent) instead of a number.

You often don't need AI at all once you read the last line carefully. When you do, you'll ask a much better question.

3. Minimal reproduction

Shrink the problem until removing anything else makes the bug disappear. In the example: one row {"roll": 2, "marks": "AB"} and one line int("AB"). A minimal reproduction:

• often reveals the bug to you while making it;
• fits easily in a prompt, a Stack Overflow question or a GitHub issue;
• removes private data and unrelated code.

4. Ask AI the right way

A strong debugging prompt has five parts:

1. What I'm doing: computing class averages from a CSV export of marks (Python 3.12).
2. Minimal code: <the 6 lines>
3. Exact error / wrong output: ValueError: invalid literal for int() with base 10: 'AB'
4. Expected behaviour: absent students ("AB") should be skipped, not crash the report.
5. What I tried: printing rows showed 'AB' for absent students.
Question: what's the cleanest way to skip non-numeric marks and report how many were skipped?

Compare with "my average code crashes help". The first gets a correct, targeted fix; the second gets a guess.

The fix that follows from it:

def class_average(rows):
    marks, skipped = [], 0
    for r in rows:
        if r["marks"].strip().isdigit():
            marks.append(int(r["marks"]))
        else:
            skipped += 1
    avg = sum(marks) / len(marks) if marks else 0.0
    return round(avg, 2), skipped

rows = [{"roll": 1, "marks": "78"}, {"roll": 2, "marks": "AB"}, {"roll": 3, "marks": "91"}]
print(class_average(rows))
print(class_average([{"roll": 9, "marks": "AB"}]))

(84.5, 1)
(0.0, 1)

5. One hypothesis at a time

When AI suggests five possible causes, don't apply five changes at once — you won't know which one mattered (or which new bug you added). Test them one by one, cheapest first:

Print/log debugging is fine. So is a real debugger (VS Code breakpoints, pdb, browser DevTools) — AI can explain how to set them up for your stack.

6. When AI goes in circles

Signs: each "fix" introduces a new error, the code keeps growing, or it re-suggests something you already tried. Then:

• Stop and reset context. Start a fresh conversation with a clean summary: minimal code, error, what's been ruled out.
• Give it the missing facts. Versions (pip show, npm ls), config, the actual data.
• Check the official docs / changelog — the API may have changed after the model's training cut-off.
• Revert to the last working commit (git stash / git diff) and bisect what changed (git bisect finds the commit that introduced a bug).

7. Never paste these into a prompt

• API keys, passwords, tokens, .env contents (replace with <REDACTED>)
• Real customer/student personal data (use fake rows with the same shape)
• Proprietary code your company hasn't approved for the tool (lesson 7)

8. Case study: "works locally, fails on the server"

A student's Flask app returned 500 errors only in production. AI suggestions about CORS and routes went nowhere for an hour. Then they followed the loop: read the server log (not the browser), found KeyError: 'DATABASE_URL', made a two-line reproduction, and the prompt became precise: "env var available locally via .env, missing under systemd". Answer in one reply: systemd doesn't read .env — add EnvironmentFile= to the unit (Cloud & DevOps track, lesson 2). Total fix time after reading the log: five minutes.

Common mistakes

• Pasting a whole project instead of a minimal reproduction.
• Not including the exact error text and expected behaviour.
• Applying many suggested changes at once.
• Debugging the symptom in the browser instead of reading server logs.
• Pasting secrets or real personal data.

Quick recap

• Loop: reproduce → read → shrink → hypothesise → test → fix → add a test.
• Read stack traces bottom-up; the last frame in your code is the usual suspect.
• Ask AI with: context, minimal code, exact error, expected behaviour, what you tried.
• Test one hypothesis at a time; reset when going in circles; check docs and versions.
• Never paste secrets or real personal data.