Claude Certified Architect – Foundations (CCAR-F) lesson one: what the exam measures, how it asks, how to be ready
If you are moving toward Claude solution architecture and wondering what to study first, lesson one of Alloy's Claude Certified Architect – Foundations (CCAR-F) course answers it: what the exam measures, how it asks, and how to be ready. The lesson is free, and so is its podcast, "How to Take This Exam". It has four parts.
Part 1: how the assessment works
CCAR-F is a real published certification, and its pass mark is 720 on a scaled score of 100–1,000 — a scaled score, not a percentage of items answered correctly, and it does not convert to one. Scoring is on the total scaled score only; the per-domain percentages on your result are informational, not separate bars.
The lesson draws the practical consequence straight away: do not set your preparation target by guessing how many questions you can afford to miss, because the scoring model does not support that conversion. Doing it anyway produces false confidence in one of two directions — stopping too early, or over-investing in the wrong area because the number was read as something it is not.
Part 2: how to take the exam
Knowing the material is half a pass; the other half is what you do in the seat. A certification question's wrong answers are written to compete with the right one, so the losses come from options that look right rather than from material you never learned. The misconception underneath every wrong-answer shape is that the most impressive answer must be the best. The test is never could this work? — plenty of wrong options would — but does this best satisfy this exact requirement?
The lesson names five shapes, each with its tell and its counter:
| Shape | The tell | The counter |
|---|---|---|
| The absolutist | always / never / all / only / every | Name one counter-example. You only need to show it is not universally true. |
| True, but not the question | A correct statement that does not answer what was asked | Re-read the stem's final sentence: does it want the best design, the first action, the main benefit, the most appropriate control, or the explanation for a specific failure? |
| Right action, wrong layer | The right remedy in the wrong place — the prompt instead of retrieval, the application instead of the model | Name the layer the problem lives in before reading the options. |
| The over-engineered one | More architecture than the scenario needs — an agent where one call works | Prefer the simplest option that fully satisfies the requirement. |
| Answers a different question | A strong answer to an adjacent question | Restate the question in your own words first, then match. |
Changing an answer. Change it when you can name what you missed on the first read, and not on a feeling. A valid change has a specific trigger — an overlooked qualifier, a missed requirement, a misread outcome, a fact that changes the comparison. Anxiety is not new evidence, and the ticking clock is the worst reason of all. The rule prevents two opposite failures: keeping an answer after discovering a genuine miss, and abandoning a defensible one because a distractor has become psychologically louder.
Part 3: what the Architect exam measures
The exam is built to defeat a vocabulary strategy. It measures whether you can make an architect's design decisions on real systems: should this be a fixed pipeline or an adaptive agent loop, how does control flow terminate, where does grounded context come from and how do you know it has gone stale, what output can the next system parse, and which tools does the model get and how are they described? None has a one-word answer, so most questions are scenarios — a realistic design and "what do you do?", or a flawed approach and "what fails, and what is the correct alternative?" Recall questions exist but are the minority.
Five domains organise that judgement:
- Agentic Architecture & Orchestration (27%) — an agent whose control flow is driven by
stop_reason, subagents coordinated through a clear router, and the choice between a fixed pipeline and a dynamic workflow. The largest domain. - Claude Code Configuration & Workflows (20%) — plan, edit, verify, and keeping an autonomous run honest.
- Prompt Engineering & Structured Output (20%) — machine-parseable output you can act on, reliably.
- Tool Design & MCP Integration (18%) — tools and MCP servers whose descriptions and schemas make the model use them correctly.
- Context Management & Reliability (15%) — grounding in retrieved evidence, citation, and spotting a fluent answer built on thin, stale or unverified context.
Part 4: the five domains as a study plan
Weight your effort to the exam, not your comfort. The lesson's first pass through the path:
- Agentic architecture — study first, study most. Agentic loops, coordinator and subagent orchestration, context passing and the pipeline-versus-workflow trade-off, paired with the lab where you build a real loop.
- Claude Code workflows. The plan → edit → verify loop and the verification discipline that stops an autonomous run confidently shipping something broken. Practise questions about what a run should do next.
- Context reliability. Retrieval, grounding and citation, and the anti-patterns learners most often miss — over-weight them.
- Prompting and structured outputs. The difference between asking nicely for JSON and constraining output so it parses every time.
- Tool design and MCP. Descriptions and schemas that make the model use tools correctly — the smallest domain by concept count and a fast source of points.
The discipline that works: read just enough to understand why a mechanism exists, then practise immediately — commit, rate your confidence, review where your reasoning broke — and interleave domains rather than cramming one.
How to use Alloy to get ready
The lesson ends with the workspace itself, in order:
- Take the diagnostic first, before a single lesson, in one sitting — 24 questions, about 20 minutes, from the same question bank and weighted like the exam, with a breakdown by domain on the summary screen.
- Rate your confidence honestly on every answer. Certain-and-wrong is what fails people; Alloy can hand back exactly those questions.
- Do flashcards on the day they come up, rating recall Again, Hard, Good or Easy — "Again" is the most useful answer.
- Prove it rather than tick it. Reading is the weakest evidence; being right when confident is stronger; a scenario or lab stronger still. Each concept carries a state — studied, practised, built, passed, demonstrated — and a passing lab, especially one done with no help, is the strongest proof you can produce.
- Open Recommended next when you do not know what to do; it weighs what you are closest to forgetting, where you are weakest, and what carries the most weight that you have not started.
- Watch readiness — not ready, almost, ready, strongly ready — and sit a practice exam when it says to. When it reaches ready, the course's final exam unlocks.
- Come back, because the curriculum is watched and kept current by people.
- Use labs and scenarios where they appear, and the Curriculum as the map everything is indexed against.
The exam-strategy lesson is worth a 10–20% score bump, because it stops you giving away questions you already knew.
Next step
Take the free CCAR-F diagnostic, then start lesson one and its podcast. The second lesson, on agentic loops and stop_reason control flow, is free too.
Alloy is exam preparation built from the official exam guides — only what the exam asks for.