LLMs are bad at returning code in JSON
「 Benchmarks show that models struggle with syntax errors in the code they write, related to quoting and escaping it into JSON. The benchmark results also imply a decreased capacity for solving coding problems due to the burden of JSON formatting 」
