[Lab 07 Released] Context Harness — A harness that reads only the code needed for each question
I’ve uploaded the practice materials for Section 2 (Lessons 4–8). Run git pull, and labs/07-context-harness/ will be created.
▶ What do we do?
Only read the code and documentation needed for each type of request, expanding one level at a time only when the evidence is insufficient.
You can build the context harness yourself or work through the hands-on exercises.
- CLAUDE.md decomposition (259 lines → 28 lines + path-specific rules)
- Task Router (regex, injected via the UserPromptSubmit hook)
- Search first (search_context.sh)
- 4 Skills + references (progressive disclosure)
- Progressive disclosure (only one level deeper when evidence requires it)