I read this earlier on Anthropic's blog, and their wording was a breath of fresh air

They definitively said it didn't work and the fix was more prompting, which at some point is just pseudo-programming in a LLM