I love that all of AI’s prompting/agents/skills/etc. infrastructure is 100% text based. It makes it so easy to monitor and update exactly what is being fed to the LLM. It would, of course, have been impossible to have such human-auditable configurations before software could “interpret” settings.