Post 2 of 2. A walkthrough of the wrapper that breaks UserPromptSubmit hook output, the issues that document the problem (with current states and a canonical open fix request to monitor), two empirical tests against current Claude Code that confirm the wrapper-authority story, and the structural lesson about agent design that this all points to. Where this picks up In the previous post , I wrote about a discovery ceiling in Anthropic's Agent Skills standard: at around 32 skills the description budget is exhausted, the agent's index of available skills starts truncating, and even the survivors get skipped 56% of the time according to Vercel's evaluation. I built a small open-source daemon called PreBrief that does per-prompt semantic search over the full skill content as a way around it. The daemon worked. But the first thing I tried for getting that output to the agent — a Claude Code hook — didn't work.…