what does it do? it does some god-awful special-case hacks to make it look like they made LLMs do a good job at code when in fact it just keeps trying and trying and trying until it makes something that compiles. haaahahahahahahahaha

@pikhq When I run models on my own I usually do it in a harness that uses BPF grammar to make it impossible for the model to generate ungrammatical tokens, you can use this trick to load up a JSON schema and force the model to obey. It's kinda neat and is the basis of how tool calling works everywhere but with Anthropic lol