Don't use OpenAI's outdated tools. Also, don't rely on prompt engineering to force the output to conform. Instead, use a local LLM and something like jsonformer or parserllm which can provably output well-formed/parseable text.