Wow, seems pretty obvious what this was written for (even if they didn’t know it): building commands/prompts for LLM ensembles. It’s just too perfect —- the LLMs could easily implement their own combinators. Would be a huge step towards self-improving systems…