Domain 3 — Claude Code Configuration & Workflows (20%)

How to set up Claude Code for teams, CI, and complex workflows.

Topics

3.1 CLAUDE.md hierarchy

• ~/.claude/CLAUDE.md — personal global.
• <repo>/CLAUDE.md or <repo>/.claude/CLAUDE.md — project-shared (committed).
• <repo>/.claude/CLAUDE.local.md — project-local, gitignored.
• Subdir CLAUDE.md — scoped to that subtree.

Memorize precedence and what belongs where.

3.2 Settings / permissions

• Files: .claude/settings.json (shared), .claude/settings.local.json (local), ~/.claude/settings.json (user).
• Precedence: CLI flags > local > shared > user > enterprise.
• deny evaluated first, always wins.
• allow skips the confirmation prompt.
• Patterns are Tool(arg-glob) style: Bash(npm test:*), Read(./**).

3.3 Slash commands

• Live in .claude/commands/<name>.md (shared) or ~/.claude/commands/<name>.md (user).
• Body is the prompt. $ARGUMENTS interpolates user input. Backticks run shell.
• Frontmatter: description, allowed-tools, argument-hint.
• Cheap, parameterized, no context overhead.

3.4 Skills

• .claude/skills/<name>/SKILL.md — required.
• Frontmatter description triggers selection; the body is loaded on invocation.
• May include scripts/resources alongside SKILL.md.
• Lazy-loaded — many skills are fine.

3.5 Subagents

• .claude/agents/<name>.md — frontmatter description, tools, model, isolation.
• Run in isolated context. Receive only the parent's prompt; return only a final summary.
• Use for: bounded research, parallel exploration, isolated risky writes.

3.6 Hooks

• Events: PreToolUse, PostToolUse, UserPromptSubmit, Stop, SubagentStop, Notification, SessionStart.
• Hook = shell command, runs synchronously, gets event JSON on stdin.
• Exit code 2 blocks the action; stdout becomes the reason for the model.
• Use for: lint/format on edit, test on stop, audit, secret scanning.

3.7 Plan mode

• --permission-mode plan or Shift+Tab cycle.
• Read-only; agent proposes plan, user approves, agent executes.
• Best for multi-file changes, refactors, ambiguous tasks.

3.8 Non-interactive / CI use

• claude --print "task" for one-shot.
• --output-format json for parseable output.
• --output-format stream-json for streaming events.
• Combine with narrow allow list + hooks for safety.

3.9 Sessions

• /clear to reset; /resume to continue.
• Sessions persist working directory and history.
• Worktree isolation (subagent isolation: worktree) keeps filesystem changes scoped.

Common question shapes

• "Where should a team-wide convention live?" → <repo>/.claude/CLAUDE.md, committed.
• "Where should personal style preferences live?" → ~/.claude/CLAUDE.md.
• "How to enforce that pre-commit lint always runs?" → PostToolUse hook on Edit|Write, or Stop hook.
• "Run Claude in CI without blocking on confirmations?" → --print + permissive allow for read-only tools + denies for destructive.
• "Refactor across 12 files safely?" → Plan mode → user approval → execute.

Cheatsheet

• Shared > Local > Personal for team consistency.
• Deny always wins.
• Hooks > prompt rules for non-negotiable behavior.
• Plan mode for ambiguity.
• CI uses --print + narrow allow + hooks.

See also

• claude-code-overview
• claude-code-md
• claude-code-settings
• claude-code-hooks
• claude-code-plan-mode
• agent-skills
• subagents
• claude-code-slash-commands