On 10/15/26 15:37, Theodore Tso wrote: > On Wed, Oct 14, 2026 at 09:50:47PM -0500, Kees Cook wrote: >> Exactly. And the rate of frontier model improvement has me convinced >> that there is no need to spend any time writing documentation to be >> consumed _only_ by the agents. If an agent gets it wrong after reading >> our existing documentation, ask it what would need to change for it to >> understand it, and then see if that makes it better for a human too, and >> send those patches. :) > > We might not want to assume that everyone has access to frontier > models. At the very least some people might want to use things like > Gemini Flash instead of Gemini Pro, and there will also be people who > wants to use local LLM's. Even if we don't care about usage of more > modest LLM's, there is still an economic (and environmental!) reason > why we might want to make sure that we promote token-efficient usage. > We certainly don't want to assume that LLM's need to read *all* of > Documentation for every single prompt. Agreed, but maintaining the same information twice is work that as of now humans need to do -- and having information for one task in two different places targeting different audiences (humans and LLMs) most likely will lead to small or huge inconsistencies that cause confusion or trouble.
Can't we prevent all of that somehow -- maybe while also improving the docs for humans? Is one of those frontier models maybe now or soon capable of generating and maintaining dedicated skill files for less capable LLMs from documentation written for humans? Or at least pointing out inconsistencies when someone updates one side but not the other? Ciao, Thorsten
