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

The recurring question in this domain is where does this belong — which file, which scope, which layer. Get the precedence tables into memory and most of these questions answer themselves.

3.1 CLAUDE.md and the memory hierarchy

How loading actually works — this is exam material, not trivia:

• Files from the working directory and every directory above it load at launch. They are concatenated, not overridden, ordered filesystem-root-down, so the file closest to where you launched is read last.
• Within a directory, CLAUDE.local.md is appended after CLAUDE.md.
• Files in subdirectories do not load at launch. They load on demand, when Claude reads a file in that subdirectory. This is the monorepo answer: a root CLAUDE.md for shared conventions plus per-package files for stack-specific rules, each loading only when relevant.
• Imports use @path/to/file, resolve relative to the importing file, recurse to a maximum depth of four hops, and are skipped inside backticks or code fences. Imported content still loads at launch, so imports organise but do not save context.
• Target under 200 lines per file. Adherence degrades as files grow.
• claudeMdExcludes (glob patterns against absolute paths) skips other teams' ancestor files in a monorepo. Managed policy files are exempt.

The distinction the exam turns on — CLAUDE.md is context, not configuration. It is delivered as a user message after the system prompt; Claude reads it and tries to follow it. If the requirement is "must always" or "must never" — a compliance rule, a forbidden path, a mandatory pre-commit step — the answer is a hook or managed settings, never a line in CLAUDE.md.

Rules in .claude/rules/*.md break large instruction sets into topic files. With paths: frontmatter they become path-scoped and load only when Claude touches a matching file:

---
paths:
  - "src/api/**/*.ts"
---
# API rules
- Every endpoint validates input before use.
- Use the standard error envelope.

Rules without paths load unconditionally at the same priority as .claude/CLAUDE.md. User rules (~/.claude/rules/) load before project rules, so project rules win.

3.2 Settings precedence

Highest wins. Memorise the order:

1. Managed settings — managed-settings.json / MDM / console (your organisation)
2. Command line — claude --settings …
3. Project local — .claude/settings.local.json
4. Shared project — .claude/settings.json
5. User — ~/.claude/settings.json

List keys such as permissions.allow merge across scopes rather than replacing one another. deny and ask rules apply immediately; allow rules and additionalDirectories wait until the folder is trusted.

3.3 Permission modes

Set the starting mode with permissions.defaultMode or --permission-mode. Organisations can block the dangerous ones with permissions.disableBypassPermissionsMode and permissions.disableAutoMode.

Permission rule syntax

Evaluation: deny beats ask beats allow. Deny rules cannot be overridden by a lower-precedence allow.

3.4 Hooks — the enforcement layer

Hooks run your code at fixed lifecycle events regardless of what Claude decides. When a question says "guarantee", "always", "block", or "must not", look for the hook.

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{ "type": "command", "command": "$CLAUDE_PROJECT_DIR/.claude/guard.sh", "timeout": 30 }]
    }]
  }
}

Exit codes: 0 = success, JSON on stdout is parsed for a decision; 2 = block the action, stderr becomes the reason Claude sees; anything else = non-blocking error, the action proceeds. A JSON decision looks like:

{ "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "deny",
    "permissionDecisionReason": "Refund tool blocked: verification not recorded for this session."
} }

Handler types go beyond shell: command, http, mcp_tool, prompt, and agent. Matchers accept exact names, Edit|Write alternation, and regex such as mcp__.*.

3.5 Slash commands and skills

Custom commands live in .claude/commands/ (project, committed) or ~/.claude/commands/ (personal); subdirectories namespace them. Skills live in .claude/skills/<name>/SKILL.md and carry frontmatter that controls invocation:

Arguments arrive as $ARGUMENTS or positionally as $1, $2. Dynamic context is injected by executing a command with !`git diff HEAD`, and files are referenced with @path. Built-ins worth knowing: /init, /memory, /context, /compact, /clear, /permissions, /mcp, /agents, /plan, /review, /rewind, /resume, /doctor, /add-dir.

3.6 Headless mode and CI/CD

claude -p "Review this diff for security issues" \
  --output-format json \
  --permission-mode acceptEdits \
  --allowedTools "Read,Grep,Bash(npm test *)" \
  --append-system-prompt "You are a security reviewer. Report only exploitable findings." \
  --bare

Exit code 0 on success, non-zero on failure; SIGTERM exits 143 and leaves the turn resumable. Without --bare, a -p session runs the repository's hooks and connects its .mcp.json servers in a folder you have never trusted — a supply-chain consideration the exam has been known to probe.

The CI false-positive scenario — When an automated reviewer floods PRs with noise, the correct fix is project-specific exceptions plus few-shot examples of good findings — narrow the rule, show the standard. Removing the security check entirely, or raising a generic "be less strict" instruction, is the wrong answer every time.

← Previous: Domain 2 — Tool Design & MCP Integration  |  Next: Domain 4 — Prompt Engineering & Structured Output →