An enquiries skill and a search-reviewer subagent in Claude Code

The configuration is three files: a skill that holds the Protocol rules and the enquiry format, a subagent that reads exactly one search, and a settings file that keeps both read-only. The rest of the project is there to answer the exercise's questions with evidence instead of impressions. claude -p runs on your subscription and writes a stream-json transcript; src/stream.ts reads it for Skill and Agent calls, and src/enquiries.ts checks every quote in the final list against the file it cites. The decision that matters most is writing the skill's description as the task that should trigger it, because the description is the only part of the skill that's in context until the model chooses to load it.