Working code vs Correct code

Working code is code that runs successfully in one observed case—often the example you gave the AI. Correct code is code that meets the full requirements and works across all valid inputs, not only the one you tested. The difference matters when your data or situation changes.

AI models often produce code that matches the example you provide. If your example is too narrow, the code may only work for that specific case. Without more details, the AI cannot anticipate every valid input or requirement.

1. Test the code with several different valid inputs, not only the original example.
2. Check the requirements or rules the code is supposed to follow.
3. See if the code handles edge cases—unusual but valid situations.
4. Ask a colleague to try the code with their data.

• Fails when you change the input slightly.
• Breaks when a colleague tries it on their data.
• Misses important checks or rules from your policy.
• Only passes the original test, not others.

Yes. Even if the code runs, you need to check it against your requirements and test it with different inputs. AI does not know your full context unless you spell it out.

No. A single successful run only shows the code worked for that input. Correct code needs to handle all valid cases, including those you have not tested yet.

• Give clear, detailed requirements.
• Provide examples of different valid inputs.
• Describe edge cases or tricky situations.
• Ask the AI to explain how it handles errors.

You receive a script from an AI assistant that processes your sales report perfectly. When your colleague tries it on her report, it crashes because her file has an extra column. The code worked for you, but was not correct for all cases.

On the surface, both look like they do the job. If you only see one successful run, you might assume the code is reliable. Only when you try new data or requirements do the gaps show up.

Requirements describe what the code must do and what situations it must handle. If the code only works for one case, it is not meeting the requirements. Correct code covers all the rules and cases you set.

Yes, if you extend it to handle all valid cases and meet all requirements. Often, you start with working code and improve it until it is correct.

• Unexpected failures with new data.
• Missed deadlines when code breaks.
• Rework and lost time fixing errors.
• Embarrassment if a client sees the failure.

1. List the requirements and valid inputs.
2. Try the code with different data sets.
3. Check if it handles errors gracefully.
4. See if it gives the right results every time.

No. Correct code meets the stated requirements and handles all valid cases, but it might not be optimised for speed or style. Perfection is not the goal—reliability is.

1. Compare your data with your colleague's to spot differences.
2. Check if the code assumes a specific format or value.
3. Update the code to handle both cases.
4. Test again with both data sets.

No. If the requirements are vague or missing, the AI can only guess what you want. Clear requirements are needed for correct code.

An edge case is an unusual but valid situation—like a blank value or a very large number. Correct code handles these cases, while working code often misses them.

Testing with different inputs and scenarios shows whether the code only works for one case or for all the cases you care about. It is the fastest way to spot gaps.

No. A single success does not guarantee reliability. Production code needs to be correct—able to handle all valid cases and meet your requirements every time.

• What requirements does this code need to meet?
• Have I tested it with different valid inputs?
• Does it handle errors or edge cases?
• Would it work for my colleague's files too?