← Learn Claude

CCA-F Exam Domain Traps

Use this as a drill sheet for Exam Domains (CCA-F). For each trap, write your own one-sentence rule before moving on.

D1 — Agentic Architecture & Orchestration (27%, ~16 questions)

Topics: Multi-agent design, task decomposition, supervisor/worker patterns, sequential vs parallel execution, failure handling (retry, fallback, abort).

Exam trap: Two plausible architectures appear in the same question. The correct answer minimises context propagation and handles errors at the right abstraction layer.

Drill rule: Prefer the design that passes less context between agents and fails/retry at the layer that owns the failure — not by flooding the supervisor with full worker logs.

Self-check:
- [ ] Can explain when parallel beats sequential (independent subtasks)
- [ ] Can describe what happens when worker B fails mid-pipeline
- [ ] Can identify over-sharing context vs minimum necessary delegation

YouTube: Ep 02–04 · slides lesson 2.pdf / lesson 3.pdf · Study map

D2 — Claude Code Configuration & Workflows (20%, ~12 questions)

Topics: CLAUDE.md project vs global scope, custom skills, slash commands, PreToolUse/PostToolUse hooks with exit codes, permission modes, subagent spawning.

Exam trap: Exam matches docs.anthropic.com exactly — do not rely on summaries. Questions get surprisingly specific on CLAUDE and hook configuration.

Drill rule: When in doubt, cite official doc behaviour: hook exit codes, file scope hierarchy, permission matchers.

Self-check:
- [ ] Know project vs user vs global CLAUDE.md precedence
- [ ] Know hook exit code meanings (0 = allow/continue, 2 = block on PreToolUse)
- [ ] Built P4 project: ../p4-claude-code-workflow

YouTube: Ep 05, 08–10 · built-in tools Ep 09 · CCA-STUDY-MAP

D3 — Prompt Engineering & Structured Output (20%, ~12 questions)

Topics: Production structured outputs, system prompt architecture for multi-user apps, output failure detection, Constitutional AI, minimum constraint principle.

Exam trap: Two prompts both seem right — pick the more constrained, more specific one. Production prompts are contracts, not conversations.

Drill rule: More explicit criteria, schema, and validation beats friendly open-ended instructions.

Self-check:
- [ ] Prefer tool_use + JSON schema over prose-only output when reliability matters
- [ ] Can spot vague system prompts that leak behaviour across tenants/users

YouTube: Ep 01 (roles), 06 (descriptions), 10; later Ep 15+ few_shot_examples.py · CCA-STUDY-MAP

D4 — Tool Design & MCP Integration (18%, ~11 questions)

Topics: Tool schema and descriptions, API key vs OAuth auth, MCP transport (stdio local vs HTTP/SSE remote), partial failure handling, enterprise security.

Exam trap: Most close-but-wrong distractors live here. Transport types and authentication patterns change the answer — know them cold.

Drill rule: Local dev → stdio; remote/shared → HTTP/SSE. Auth choice depends on deployment model, not convenience.

Self-check:
- [ ] Built P5 project: ../p5-custom-mcp-server
- [ ] Can explain when stdio vs HTTP/SSE is correct for a given deployment
- [ ] Tool descriptions include when not to use the tool

YouTube: Ep 06–08 · .mcp.json sample · CCA-STUDY-MAP

D5 — Context Management & Reliability (15%, ~9 questions)

Topics: Summarise vs trim vs archive, minimum necessary context, model selection (Haiku/Sonnet/Opus), retry/backoff, production evaluation metrics.

Exam trap: Correct answer always gives Claude minimum context needed — not maximum. Look for the most token-efficient option that preserves accuracy.

Drill rule: Smallest context window that still contains decision-critical facts wins.

Self-check:
- [ ] Progressive summarisation risks understood (lost nuance, stale facts)
- [ ] Can justify Haiku vs Sonnet vs Opus for a given subtask

YouTube: Ep 03, 07, 09 (incremental exploration) · CCA-STUDY-MAP

Speed strategy (exam day)

Mastery checklist