Post 1 of 2. This post walks through the discovery ceiling in Anthropic's Agent Skills standard, the obvious first attempt at fixing it (a hook that didn't work), and the architecture that did. A follow-up post will cover the hook story in detail — why Claude Code's UserPromptSubmit hook can't reliably inject knowledge, what's filed about it, and what would actually fix it. A small puzzle to start Last month I installed Google's new Workspace CLI, gws . It's a command-line tool that wraps every Workspace API — Gmail, Calendar, Drive, Sheets — and ships with 95 ready-made "skills" that explain how to use each one. The pitch is that any AI coding agent that supports the Agent Skills standard can read these skills and operate Workspace on your behalf. I asked Claude Code to fetch my most recent email. Simple task; the skill for it is sitting right there in ~/.claude/skills/gws-gmail/SKILL.md , with the exact commands and recommended flags.…