I find myself wondering if there is a job in optimising LLM agent skills by converting them to run deterministically with real programming/scripting languages.
I see so much “if…then”, decision trees, and even literal “for each” mechanics in instructions to LLMs.
This week I reduced a part of a skill that took 14s to run with an LLM to a PHP script that took 2 seconds (mostly querying an API) and used zero tokens.
With companies asking people to reduce token use, this seems like a no-brainer.