How to write a ChatGPT prompt for code

What to put in the prompt

1. 1.

   Paste the real code

   The actual function, query or error, in a code block. Not a description of it. The model can't debug code it can't see, and paraphrased code tends to lose the bug.

2. 2.

   Say what it should do

   One sentence on the intended behaviour, like “this should return the user's last 10 orders, newest first”. That's what the model checks its answer against.

3. 3.

   Say what happens instead

   The exact error message, the wrong output, or how slow it is. “It takes 4 seconds on 200k rows” is something the model can work with. “It's broken” isn't.

4. 4.

   Give the setup

   Language, version, framework, database. “Python 3.12, Django 5, Postgres” takes five seconds to type and rules out a whole class of wrong answers.

5. 5.

   Say what must not change

   The function signature, the output format, a library you can't swap out. Without this you'll often get a rewrite of the parts that were fine.

6. 6.

   Ask for a specific answer

   Just the fixed code, the fix plus why it broke, or a diff. Say which. If you want to understand it, ask for the reason first and the code second.

Common mistakes

• Just writing “fix it”

  Without saying what's wrong, the model guesses. Sometimes it guesses right. Often it tidies something else and the real bug stays.

• Pasting the whole file

  Paste the function that fails and what it calls, not 500 lines. If it's long, say which line errors.

• Leaving out the error

  Copy the full error and stack trace. The line number in it is often most of the answer.

• Not saying the version

  APIs change between versions. An answer written for an older release can look right and still not run.

• Starting a new chat when it fails

  Paste the new error back into the same chat. It keeps what it already knows about your code.

A template you can copy

Goal:
[What the code should do, in one sentence.]

Context:
[Language, version, framework.]
Here's the code:
```
[paste the code]
```
What happens instead: [the exact error or the wrong output]

Requirements:
- Keep [the function signature / the output format / this library] the same.
- [Anything else it must or must not do.]

Output format:
[The fixed code only / the reason it fails, then the fix / a diff.]

Fill in the brackets and delete the lines you don't need. A short prompt with the code and the error beats a long one without them.

A real before and after

From our test runs. A one-line question, and the prompt Polido rewrote it into. The code in backticks comes through exactly as typed.

Questions

Should I tell ChatGPT to act as a senior developer?

It rarely hurts, but it isn't what makes the answer good. The code, the error and what you want back do far more than a persona line.

How much code should I paste?

Enough to reproduce the problem: the function that fails, the code it calls, and the error. For a big file, paste the relevant part and say what's around it.

Does this work for Claude and Gemini too?

Yes. The same structure works in any AI chat, because the problem it solves, the model guessing what you meant, is the same everywhere.

Why does ChatGPT keep changing code I didn't ask it to touch?

Because nothing in the prompt said to leave it alone. Add a line like “keep the function signature and don't change anything outside this function”.