Aim: 7/8+.
1. A new teammate lacks the team's shared instructions. Likely cause — they live in:
- User-level
~/.claude/CLAUDE.md, personal and not shared
- The project root CLAUDE.md committed to the shared repository
- An imported standards file tracked in version control already
- A project
.claude/rules/ file with broad glob patterns
Option 1. User scope isn't version-controlled. (L1)
2. One convention must apply to test files scattered across the whole tree. Use:
- A
.claude/rules/ file with a glob path for test files
- A single root CLAUDE.md relying on Claude to infer the section
- A CLAUDE.md in each directory that happens to contain tests
- A manually invoked skill loaded before editing any test file
Option 1. Glob rules apply by file type across directories. (L1)
3. A /review command must reach every developer on clone. Put it in:
.claude/commands/ in the project repository~/.claude/commands/ in each home directory- the project root CLAUDE.md as a command section
- a
.claude/config.json commands array entry
Option 1. Project-scoped commands ship with the repo; config.json array doesn't exist. (L2)
4. A verbose codebase-analysis skill clutters the main chat. Isolate it with:
allowed-tools narrowing it to read-only operationscontext: fork running it in an isolated sub-agentargument-hint prompting for the analysis target first- a directory CLAUDE.md placed next to the skill definition
Option 2. context: fork isolates output. (L2)
5. Monolith→microservices, dozens of files, boundary decisions. Approach:
- Enter plan mode to explore and design before changing code
- Start direct and let service boundaries emerge from the code
- Direct execution with detailed upfront per-service instructions
- Begin direct, switch to plan only if complexity shows up later
Option 1. Stated architectural complexity → plan mode now. (L3)
6. Multi-phase task exhausts context from reading many files. Use:
- Plan mode, which avoids reading any files during the task
- Automatic
/compact after every single file is read
- The Explore subagent to isolate discovery and return a summary
- A larger token limit so all files remain in the main context
Option 3. Explore subagent isolates discovery. (L3)
7. A CI job hangs running claude "..." awaiting input. Fix:
- Export a
CLAUDE_HEADLESS=true variable before the call
- Add the
-p / --print non-interactive flag
- Redirect stdin from
/dev/null as a shell workaround
- Append a
--batch flag to bypass the session prompt
Option 2. -p is the documented non-interactive mode. (L4)
8. Most objective review of just-generated code comes from:
- The same session re-reading and critiquing its own output
- Turning on extended thinking in the generating session
- An independent instance without the generator's reasoning
- Expanding the window so the generator reviews more at once
Option 3. Independent instance beats self-review. (L4)
Scoring8/8 → Domain 4. Misses → re-read the flagged lesson.
Tell your teacher your score.