Skip to main content

📚 The 5 Exam Domains

Per-domain breakdown: key concepts, anti-patterns, and official Anthropic resources. Weights (%) are on the Exam Overview.

Domain 1 — Agentic Architecture & Orchestration (~25%)​

Design and implement agentic systems using Claude's Agent SDK.

Key Concepts:

  • Agentic loop lifecycle: stop_reason (tool_use vs end_turn)
  • Hub-and-spoke multi-agent architecture — coordinator role, subagent context isolation
  • Task tool for subagent spawning (allowedTools must include 'Task')
  • Parallel subagent execution with fork_session for branched exploration
  • PostToolUse hooks for deterministic enforcement (beats prompt-based guidance)
  • Session management: --resume, fork_session, named sessions, stale context handling
  • Task decomposition: prompt chaining vs dynamic adaptive decomposition

Anti-Patterns to Know:

  • Parsing natural language for loop termination (use stop_reason instead)
  • Arbitrary iteration caps without conditions
  • Flat multi-agent vs hub-and-spoke (flat = no coordinator = chaos at scale)
  • Overly narrow task decomposition → coverage gaps

📚 Official resources (Anthropic):

Domain 2 — Tool Design & MCP Integration (~20%)​

Design effective tools and integrate with Model Context Protocol servers.

Key Concepts:

  • Tool description best practices: input formats, examples, edge cases in description
  • Structured error responses: isError, errorCategory, isRetryable
  • Tool distribution: 4–5 tools per agent max, scoped tool access per subagent
  • MCP server config: .mcp.json (project-level) vs ~/.claude.json (user-level)
  • Built-in tools: Read, Write, Edit, Bash, Grep, Glob — when to use each

Anti-Patterns to Know:

  • Generic error messages (no isRetryable → agent retries indefinitely)
  • Too many tools per agent → decision paralysis
  • User-level MCP for project-specific servers (should be project-level)

📚 Official resources (Anthropic):

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

Configure Claude Code for team development workflows.

Key Concepts:

  • CLAUDE.md hierarchy: user → project → directory levels
  • @import syntax, .claude/rules/ for topic-specific rules
  • Custom slash commands (.claude/commands/) vs skills (.claude/skills/)
  • SKILL.md frontmatter: context: fork, allowed-tools, argument-hint
  • Path-specific rules: YAML frontmatter with paths glob patterns
  • Plan mode vs direct execution decision criteria
  • CI/CD: -p flag, --output-format json, --json-schema
  • Session context isolation in CI (generator vs reviewer roles)
  • Message Batches API: 50% cost savings, 24-hour processing window

Anti-Patterns to Know:

  • Single monolithic CLAUDE.md vs hierarchical imports
  • Using direct execution mode for complex multi-file refactors
  • Same session for generator + reviewer (loses independent perspective)

📚 Official resources (Anthropic):

Domain 4 — Prompt Engineering & Structured Output (~20%)​

Master prompt engineering for production-scale systems.

Key Concepts:

  • Explicit criteria over vague instructions (false positive impact on precision)
  • Few-shot prompting: 2–4 examples for ambiguous classification cases
  • tool_use with JSON schemas: guaranteed schema compliance vs semantic errors
  • tool_choice: 'auto' vs 'any' vs forced specific tool
  • Schema design: required vs optional, enums with 'other' + detail field
  • Validation-retry loops: append specific errors to prompt (not generic "try again")
  • detected_pattern fields for tracking dismissal pattern escalation
  • Multi-pass review: per-file local analysis + cross-file integration pass

Anti-Patterns to Know:

  • Vague instructions → inconsistent outputs
  • Self-review in same session (retains reasoning context, defeats the purpose)
  • Generic retry prompts without specific error context

📚 Official resources (Anthropic):

Domain 5 — Context Management & Reliability (~15%)​

Manage context effectively in long-running production systems.

Key Concepts:

  • Progressive summarization risks, 'lost in the middle' effect
  • case facts blocks, trimming verbose tool outputs, position-aware context ordering
  • Escalation patterns: customer demands, policy gaps, sentiment ≠ complexity
  • Error propagation: structured context vs generic error messages
  • Local recovery before coordinator escalation, partial results reporting
  • Context degradation in extended sessions → scratchpad files
  • /compact, subagent delegation, crash recovery manifests
  • Human review: stratified sampling, field-level confidence scoring
  • Information provenance: claim-source mappings, temporal data freshness

Anti-Patterns to Know:

  • Appending all context chronologically (no position-awareness)
  • Escalating on sentiment rather than actual complexity/policy gap
  • No crash recovery manifest → lost progress on failure

📚 Official resources (Anthropic):

➡️ Next: 🎓 Courses + Mock Exam — put each domain into practice with the free official courses.