Not sure if the author saw in the OpenAI docs, but their current "system" message has a really weak effect on the output. OpenAI explicitly recommends using a simple "system" message and then a separate instructions "user" message that encodes your prompt instructions.
So if you're struggling to get the chat completions API to follow instructions, don't rely on the system message alone.
(I'm the author of the `chatgpt` npm package and run a community of 10k+ ChatGPT hackers, so we've run into a lot of these kinks and found that this method works much better than using the "system" message exclusively. It's even mentioned in their official chat completions guide as something that they will be improving in future versions)
So if you're struggling to get the chat completions API to follow instructions, don't rely on the system message alone.
(I'm the author of the `chatgpt` npm package and run a community of 10k+ ChatGPT hackers, so we've run into a lot of these kinks and found that this method works much better than using the "system" message exclusively. It's even mentioned in their official chat completions guide as something that they will be improving in future versions)