Claude Certified Architect (CCA-F) Exam Guide

What the exam rewards

• ✓Determine loop completion using stop_reason — never parse natural language signals or apply arbitrary iteration caps
• ✓For financial or security-critical operations, enforce tool ordering programmatically with hooks; prompt instructions alone have a non-zero failure rate
• ✓In hub-and-spoke orchestration, all communication routes through the coordinator — subagents never communicate directly with each other
• ✓Emit multiple Task tool calls in a single response to spawn subagents in parallel and reduce round-trip latency
• ✓Trace failures to their root cause: a research report missing entire topic areas is a coordinator decomposition failure, not a subagent failure

Anti-patterns to reject

• ✕Parsing natural language to detect when the agentic loop should stop
• ✕Treating iteration caps as the primary stopping mechanism
• ✕Assuming subagents automatically inherit the coordinator's context
• ✕Using prompt instructions to enforce critical business rules

What the exam rewards

• ✓A strong tool description specifies: what it does, expected input formats, example queries it handles well, and explicit boundaries versus similar-looking tools
• ✓When misrouting occurs, improve descriptions first — before adding few-shot examples, routing classifiers, or consolidating tools
• ✓Scope each subagent to 4–5 tools relevant to its role; giving one agent 18 tools degrades selection reliability
• ✓Know tool_choice: 'auto' (model may return text), 'any' (must call a tool, picks which), forced name (must call a specific named tool)
• ✓Evaluate existing community MCP servers before building custom ones; only build custom servers for workflows community servers cannot handle

Anti-patterns to reject

• ✕Near-identical tool descriptions causing constant misrouting between similar tools
• ✕Giving all agents access to all available tools
• ✕Building a custom MCP server before checking community servers
• ✕Using Glob to search file contents, or Grep to match file names

What the exam rewards

• ✓Three config levels: user (~/.claude/CLAUDE.md, personal, not shared), project (.claude/CLAUDE.md, version-controlled, shared via git), directory (subdirectory files, applies to that path only)
• ✓Path-specific rules in .claude/rules/ with YAML glob frontmatter apply across the entire codebase — unlike directory-level CLAUDE.md which is location-bound
• ✓Use plan mode for multi-file migrations, architectural decisions, or any task where the correct approach needs exploring before committing; use direct execution for clear, single-file tasks
• ✓The -p flag is mandatory for non-interactive CI/CD pipelines — without it, the pipeline job hangs waiting for input
• ✓Skills with context: fork run in isolated sub-agent context, keeping verbose output out of the main conversation window

Anti-patterns to reject

• ✕Putting team standards in user-level CLAUDE.md — new team members won't receive them
• ✕Using directory-level CLAUDE.md when test conventions need to span many directories
• ✕Running Claude Code in CI without the -p flag
• ✕Putting task-specific procedures in CLAUDE.md, or universal standards in skills

What the exam rewards

• ✓Replace vague confidence instructions with specific categorical criteria: 'flag only when claimed behaviour contradicts actual code behaviour' outperforms 'be conservative'
• ✓Few-shot examples (2–4 targeted cases showing reasoning for ambiguous calls) are more effective than detailed instructions for consistency
• ✓tool_use with JSON schemas eliminates syntax errors, but not semantic errors — schemas won't prevent wrong field values, mismatched sums, or fabricated data
• ✓Make fields nullable when source documents may not contain the information — this prevents the model from fabricating required fields
• ✓Batch API is for latency-tolerant workflows (overnight reports, nightly audits). Blocking pre-merge checks that developers wait on must stay synchronous

Anti-patterns to reject

• ✕Using confidence thresholds ('only high-confidence findings') instead of categorical criteria
• ✕Adding more instructions when 2–3 concrete examples would resolve the inconsistency
• ✕Routing blocking pre-merge checks through the batch API
• ✕Assuming JSON schemas prevent all extraction errors

What the exam rewards

• ✓Extract transactional specifics (amounts, dates, IDs) into a persistent case facts block and include it verbatim in every prompt — never let it be summarised
• ✓Three valid escalation triggers: customer explicitly requests a human, a policy gap exists, unable to make meaningful progress. Sentiment and self-reported confidence are not reliable triggers.
• ✓The 'lost in the middle' effect is real — place key summaries and findings at the beginning of long inputs, not buried in the middle
• ✓When propagating errors in multi-agent systems, include: failure type, what was attempted, partial results obtained, and alternative approaches. Never silently suppress failures or terminate the full pipeline on one error.

Anti-patterns to reject

• ✕Using progressive summarisation on conversations containing transactional data
• ✕Escalating based on customer frustration or detected sentiment
• ✕Escalating based on the model's self-reported confidence score
• ✕Terminating an entire pipeline workflow because a single tool failed