Skip to content
Data AI Starter

TUTORIALS

CCA-F Part 7: Exam Scenarios, Tips from People Who Passed, and All Resources

The six exam scenarios, eight tips from candidates who scored above 720, and a comprehensive list of free and paid prep resources.

By Mohamed AL-Kaisi 11 min read 3 views

CCA-F Part 7: Exam Scenarios, Tips from People Who Passed, and All Resources

Part 7 of 7 of Claude Certified Architect — Foundations: a complete self-study tutorial. See all parts.

The six exam scenarios

Four of these six appear on every exam, chosen randomly.

1. Customer support resolution agent

Multi-agent dispatch (returns / billing / accounts), confidence-gated escalation, structured tool errors, session state for repeat callers.

2. Code generation with Claude Code

Team-wide CLAUDE.md plus per-language .claude/rules/ with paths:. Plan mode for refactors, direct mode for typo fixes. Skills with context: fork to keep one-off jobs out of the main session.

3. Multi-agent research

Coordinator delegates to search / analysis / synthesis / reporting. On partial failure of the analysis step, proceed with what synthesis can do and flag the confidence — do not abandon.

4. Developer productivity tools

Choose the right built-in tool (Grep for content; Glob for paths; Read for files; Bash for execution). Add custom MCP servers for internal systems (Jira, internal docs).

5. Claude Code in CI/CD

claude -p "..." --output-format json for headless review. Two-pass review pattern: first pass flags candidate issues, second pass validates each against the actual file. Stateless per PR.

6. Structured data extraction

tool_use with a JSON-schema tool. Nullable fields with "output null if not present verbatim" rule. Few-shot examples of the null case. Message Batches API for large jobs.

<a id="reference"></a>

Reference: Claude models, pricing, prompt caching

Latest models (June 2026)

Model Model ID Context Output Input price Output price Strengths
Claude Fable 5 claude-fable-5 1M 128k $10 / MTok $50 / MTok Most capable widely released
Claude Opus 4.8 claude-opus-4-8 1M 128k $5 / MTok $25 / MTok Top Opus tier; complex reasoning, agentic coding
Claude Sonnet 4.6 claude-sonnet-4-6 1M 64k $3 / MTok $15 / MTok Best speed-to-intelligence ratio
Claude Haiku 4.5 claude-haiku-4-5-20251001 200k 64k $1 / MTok $5 / MTok Fastest, near-frontier

Pick Sonnet 4.6 as the default for production agents; it is the right balance for most loops. Opus 4.8 for the hardest reasoning steps; Haiku 4.5 for cheap, parallelisable classification.

Prompt caching

Use prompt caching for any prompt over a few KB that you reuse. The economics:

  • 5-minute TTL (default): cache writes cost 1.25× base input price; cache reads cost 0.1×.
  • 1-hour TTL: cache writes cost 2× base input price; reads still 0.1×.
  • Up to 4 cache breakpoints per request.
  • Lookback window of 20 blocks from the breakpoint.

Minimum cacheable prompt sizes vary by model — 1,024 tokens on Opus 4.8, Sonnet 4.6, Sonnet 4.5; 4,096 tokens on Haiku 4.5 and the older Opus 4.6.

Cache hierarchy: tools → system → messages. Changing the tools invalidates everything below; changing the system invalidates messages but not tools; changing messages invalidates nothing above.

Mark cache breakpoints on the last block that stays identical across requests — typically the end of the system prompt or the end of a long static document the user is iterating on.

Batch processing

Use the Message Batches API when results within 24 hours are fine. 50% discount on input and output tokens. Up to 100,000 requests per batch, up to 256 MB.

<a id="study-plan"></a>

Study plan and practice exams

A 3-week plan

Week 1 — Map the territory. Read this tutorial top to bottom. Take the diagnostic (the practice exam attached to this article — 30 questions across the five domains). Note which domains you scored under 70% on; those are your focus for Week 2.

Week 2 — Build the patterns. Pick three of the six scenarios. For each, build a working prototype in Claude Code or against the Claude API:

  • Two agents that hand off (Scenario 1 or 3).
  • A CLAUDE.md + .claude/rules/ + a custom skill setup (Scenario 2 or 4).
  • A tool_use extraction pipeline with validation-retry and a small Batches API job (Scenario 6).

Week 3 — Mock exam loop. Re-take the practice exam in this article under timed conditions (60 minutes for 30 questions ≈ exam pace). Review every wrong answer the same day. Book the real exam when you score above 80% twice in a row.

Practice exam

The next part of this post is a 30-question practice exam matching the format and weighting of the real CCA-F. Each question has a single correct answer; the explanation tells you why the right answer is right and why each wrong answer is wrong.

The exam is also available as an interactive runner via your account — click Take this exam on the page.

<a id="tips-from-passers"></a>

Tips from people who have already passed

Synthesised from public pass write-ups by candidates who scored above 720 (one reported a 911). Eight tips that recur in the post-exam reflections:

1. Domain 1 is underestimated. Spend more time there. At 27% of the exam, Agentic Architecture & Orchestration is bigger than any other single domain — yet candidates routinely report giving it the same study time as Domains 4 and 5 combined. Plan to spend roughly one-third of your study hours here.

2. "If it must always happen, use code. If it should usually happen, use a prompt." This sentence comes up in nearly every pass guide. It is the single best mental model for choosing between two seemingly valid answers. A categorical rule (refund > £500 → human) should be a code-level gate, not a prompt instruction.

3. Mental models beat memorisation. You will not get a multiple-choice fact question like "what is the context window of Sonnet 4.6?" You will get four options that all look plausible, and the right one is the one that obeys good architectural principles. Internalise the patterns; you can look up the numbers later.

4. Eight recurring distractor patterns. Pass write-ups consistently flag the same wrong-answer shapes:

  • Putting business logic into a system prompt instead of a code-level gate.
  • Over-engineering before trying the simpler fix.
  • Choosing the heavier model when the brief or schema was the real issue.
  • Adding more tools when fewer, better-named tools would do.
  • Unbounded retry loops instead of bounded retries plus escalation.
  • "Make Claude more confident" answers (raise temperature, add prose hedging) instead of structural fixes.
  • Silent failure / silent retry instead of structured errors with categories.
  • Pushing problems into the model instead of into the architecture.

If a tempting answer fits one of these shapes, it is almost certainly wrong.

5. Build something before you study. "Reading alone does not bridge that gap. Practice does." Spend a weekend building one end-to-end agent before you start grinding the lessons. Pick the customer-support scenario (it is the most common in the wild) and build a coordinator with two specialists. Everything you read afterwards will land deeper because you have a real system to map it onto.

6. The "three-Claude project" hands-on setup. A recurring tip from successful candidates: open a Claude Project, upload the official CCA-F exam guide PDF as the knowledge base, and create three coaches in it:

  • Learning coach — quizzes you on the concepts and explains your gaps.
  • Practice coach — generates scenario-style multiple-choice on demand.
  • Mock test simulator — runs a timed 30-question mock against the exam-style format.

Talk to each daily. The structured tutoring loop is reported as the single biggest study lever after hands-on building.

7. Read every option before choosing. Scenario questions are designed so the first option you read often looks right. Train yourself to read all four, eliminate two, then compare the remaining two against the four architectural principles in tip 4. The exam rewards the discipline.

8. Time per question. 60 questions in 120 minutes = 2 minutes per question, but scenario questions take longer than fact questions. Practical rhythm: aim for 90 seconds on the first pass, flag anything you are unsure of, return at the end. A passer scored 911 by leaving 12 minutes for review at the end.

Reality check from passers: "With six months of hands-on Claude work, I prepared in five focused days and scored 911. Without prior hands-on, plan five weeks, not five days, and build three small projects rather than reading three books."

<a id="resources"></a>

All resources

A curated reading and practice list, grouped by purpose. The free resources alone are enough to pass.

Anthropic official

Anthropic Academy (free courses — required reading)

  • Claude 101
  • AI Fluency: Framework and Foundations
  • Building with the Claude API
  • Introduction to Model Context Protocol
  • Model Context Protocol: Advanced Topics
  • Introduction to Subagents
  • Introduction to Agent Skills
  • Claude Code 101
  • Claude Code in Action
  • Introduction to Claude Cowork

(Hosted on Anthropic Skilljar; signed-in URLs vary. Search by title.)

Free, community-built study guides

Free practice question banks

Pass-experience write-ups (read at least two)

Paid courses (only if you want video and structured pacing)

Reference reading on the techniques the exam tests

This is an independent study guide. It is not affiliated with, endorsed by, or sponsored by Anthropic. Synthesised from Anthropic's official documentation, the Model Context Protocol architecture spec, and publicly shared pass experiences and community study guides linked above (June 2026).

← Part 6 of 7: Context Management and Reliability (15%) · All parts

Tags LLMs Prompt Engineering
M

Written by

Mohamed AL-Kaisi

Editor-in-chief of the Data & AI Hub.

Continue reading

More from Tutorials