CCA-F · Claude Certified Architect – Foundations
Cheat sheet
720 / 1000
Passing score (100 to 1000 scale)
4 / 6
Scenarios drawn at random out of six
1 + 3
One correct answer, three distractors, no penalty
60 · 120 min
According to prep sites, not in the official guide
15 exam reflexes
- Guarantees, money, compliance: code (hook, settings, tool, schema), never a more emphatic prompt.
- End of loop =
stop_reason. Not the text, not 'done', not a counter (the cap is a safety net). - Subagent = none of the parent's history (unless forked): everything it needs goes in the spawn prompt.
- Restricting an agent = its tools list, not the prompt, not the description.
- Wrong tool choice = rewrite name and description. Don't remove Grep, don't rely on order.
- Execution error =
isError, with a message giving the category and whether to retry. Zero results = success. - Neither silent suppression nor full stop: make the failure visible, continue, mark the gap.
- Review = an independent instance with only the diff and the criteria.
- Explicit criteria > 'be conservative'. Examples (3 to 5, varied, with reasoning) > long instructions.
- strict: true guarantees syntax, not meaning: validate in code, retry with the precise error.
- Info missing from the source: nullable field, null, or escalate. Required pushes the model to make things up.
- Blocking = synchronous. Batch = -50%, up to 24 h, join on
custom_id. - Exact facts (IDs, amounts, dates) in a structured block outside the summary; a summary loses details.
- Simplest thing that works: a fixed workflow if the steps are known in advance.
- More context doesn't fix dilution: split into passes, summary at the top.
Words in the question that point to the answer
- guarantee · cannot be bypassed · regardless
- Deterministic code: hook, check inside the tool, permissions. Never a more insistent prompt.
- first step · most efficient first diagnostic
- The cheapest fix that targets the cause: tool description, /memory.
- root cause · directly addresses
- Fix the cause (description, schema, decomposition), don't add a downstream safeguard.
- structurally eliminates
- Change the structure (consolidate, separate), don't mitigate (examples).
- before adding infrastructure
- Prompt-side answer: explicit criteria, examples.
- generalize to novel patterns
- Annotated examples that include the reasoning.
- logs show · analysis shows · 94 % of…
- The answer is in the stem's numbers: read them before the options.
- blocking · users waiting · within 30 minutes
- Synchronous API. Batch = up to 24 h, no SLA.
- same session · verify your work
- Self-review bias: independent instance, fresh context.
- all · every · always · read all files · give all tools
- Often a distractor: context or tool overload.
- classifier · vector store · separate model · async layer
- Often over-engineering compared with a simple fix.
- confidence score (self-reported) · sentiment
- Not a reliable signal without calibration; not an escalation trigger.
Pairs not to confuse
Two similar answers, decided by one detail of the question. Each number opens a quiz question that tests that detail.
Split or consolidate tools?
Description, server-side safeguard, or routing?
Examples (few-shot) or explicit criteria?
Plan mode or direct execution?
Escalate or resolve?
If: Explicit request for a human ('I want a real person NOW')
Immediate escalation, without investigating first
If: Frustration without a clear request, problem solvable right away
Acknowledge, offer to resolve now, offer the human
Resume, fork, or start over?
Error or empty result?
Retry or not?
Batch or synchronous API?
CLAUDE.md, rules, @import, command, or settings?
Sliding window or structured facts?
Fixed workflow or dynamic orchestrator?
Required or nullable field?
Restrict or pre-approve?
Tool available or tool loaded in context?
The 6 scenarios
- Customer Support
- Support agent (Agent SDK) with MCP tools
get_customer,lookup_order,process_refund,escalate_to_human. Goal: 80%+ first-contact resolution, while knowing when to escalate. - Claude Code
- A team using Claude Code (generation, refactoring, debugging, documentation): slash commands,
CLAUDE.md, plan mode vs direct execution. - Multi-Agent Research
- Coordinator + subagents (web search, document analysis, synthesis, report) that produce cited reports.
- Developer Productivity
- Agent (Agent SDK) that helps explore codebases, understand legacy code and generate boilerplate; tools Read, Write, Bash, Grep, Glob + MCP servers.
- CI/CD
- Claude Code in CI: automated reviews, test generation, PR feedback; actionable feedback, few false positives.
- Data Extraction
- Structured extraction from documents, validated with JSON Schema, robust to edge cases and integrated with downstream systems.