My biggest takeaway from this video is actually not just that Fable is smarter. It is that a smart model only becomes insanely useful when the context around it is structured properly. I think a lot of agents are still created with instructions that are way too thin. Something like: โYou are a sales agent, help with sales.โ That is not really an agent. That is just a prompt. A real agent needs a proper operating manual: role, scope, examples, source rules, stop rules, output format, tone of voice, escalation rules, and a way to verify its own work. And the folder/data structure matters just as much. If your files, skills, customer context, projects, and logs are messy, the model has to read too much, spend more tokens, and guess more often. So for me the real shift is: not just using better models, but structuring your business so the LLM can find the right information quickly, cheaply, and reliably. Claude.md then becomes more than an instruction file. It becomes the router for your whole AI operating system.