Reduce "Malformed Responses" by a minor change in the prompt

[omri123]

This change has positive effect on ratio of malformed response.
I discovered it while testing another change, in which I add line numbers to code.

I run the Exercism python benchmarks, consisting of 135 tests, using gpt-4-turbo.

• The number of chats with malformed response went down from 14 to 9
• The total number of malformed responses went down from 21 to 15

Does this difference justifies a change?
Do you want additional benchmarking?

About performance, I saw minor improvment.
community -> community+change:

• 1st try 50.04% -> 51.1%
• 2nd try 61.5% -> 63%

I know it doesn't reproduce your results, I guess difference comes from chatgpt randomness.
I counted malformed responses from the chat history, using excel and this script:
https://gist.github.com/omri123/6c0f4a07e3c25ac059e04ddeeaf7f62c

[omri123]

Edit: I counted the number of chats incorrectly.
The number of chats with malformed response went down from 12 to 9, and not from 14 to 9.

[paul-gauthier]

Thanks for putting together this PR!

The indents were intentional, to try and encourage GPT to keep leading whitespace intact in S/R blocks.

Also, the changes you are reporting in benchmarking result seem within the range of random varition. The benchmark suite tries very hard to be deterministic, but OpenAI's API is not.

[omri123]

Cool, if we will continue with the line numbers I will re-add the spaces. (this change appears in both pull requests).