---
title: "How I passed the Claude Certified Architect exam as a non-developer"
description: "What the Claude Certified Architect exam covers, why it matters, and how to get a seat if your company is not a Claude partner."
author: Agentmatik (https://agentmatik.ai)
source: https://agentmatik.ai/guides/how-i-passed-the-claude-certified-architect-exam-as-a-non-developer
date: "20 October 2026"
tools: "claude"
---
> [!TLDR]
> The Claude Certified Architect exam is Anthropic's own proctored test of whether you can make sound decisions when you build with Claude. I passed it in Los Angeles in September without an engineering background. Only members of the Claude Partner Network can sit it, so you need a partner company email, and there are ways to get one if your company is not a partner.

I wrote this for founders, operators and consultants who build with AI and have never heard of the exam. With Agentmatik AI I help a few selected companies and executives put AI to work.

## What the exam is

Claude Certified Architect - Foundations, exam code CCAR-F, is Anthropic's proctored exam for the people who design and build AI agent systems with Claude. It is one of [four credentials](https://claude.com/blog/four-role-based-claude-certifications): Associate for everyday users such as consultants and project leads, Developer for engineers who build on the Claude API, and two Architect levels for the people who design the systems, Foundations and Professional (governance and evaluation at enterprise scale). Foundations covers AI agent design, the Claude API, MCP, Claude Code setup, structured output and context management, and [Anthropic's exam guide](https://anthropic.skilljar.com/claude-certified-architect-foundations-access-request) says it checks whether you can make informed decisions about tradeoffs when you build real solutions.

Claude Code is Anthropic's AI coding tool, the Claude Agent SDK is its toolkit for building AI agents, and the Model Context Protocol (MCP) is an open standard for connecting tools and data to Claude. The questions come from realistic customer situations, such as a support AI agent, a research system made of several AI agents, Claude Code inside a code review pipeline or data pulled out of messy documents.

## Why it matters

Small and mid-sized companies and operators are swamped with offers for AI readiness audits and AI transformation sprints. I believe a proctored exam from the lab that builds the model cuts through that noise. Anthropic says its own [Applied AI team helped design the exams](https://claude.com/blog/four-role-based-claude-certifications) around real field work. You sit them at a Pearson test centre or online with a proctor, your identity is checked, and you get no notes and no AI assistant.

Anthropic is putting real money behind the people who implement AI. In March it launched the [Claude Partner Network](https://www.anthropic.com/news/claude-partner-network) with a $100 million investment, alongside this exam. By July it reported [more than 36,000 certified consultants across more than 1,300 organisations](https://claude.com/blog/four-role-based-claude-certifications), and Accenture, Deloitte and PwC have each [committed to certify](https://claude.com/blog/four-role-based-claude-certifications) between 15,000 and 50,000 people. [The Financial Times](https://www.ft.com/content/df70bf20-e809-48d0-8485-f8f514a7fbc4) reported on 23 October that Anthropic and the private equity firm Hg are [setting up an AI consultancy for Europe](https://thenextweb.com/news/anthropic-hg-european-ai-consultancy), expected by the end of 2026.

The questions test decisions. Often two answers look right, and you pick the one that fits the constraint: the cheaper option or the faster one, a rule that is enforced or one that is only requested. The idea the exam keeps returning to is that a rule which must hold every time belongs in the system, not in the prompt. A prompt can be ignored on any run, while a hook, a permission rule or a schema is enforced, so the AI cannot skip it.

I sat it in Los Angeles, on a trip to the US for Meta Connect. It is a hard exam and worth the investment, especially if you do not come from an engineering background. It certifies one vendor's tools, not AI in general.

## What the exam looks like

You get 60 questions and 120 minutes, at a Pearson VUE test centre or online with a proctor, and it costs $125 per attempt. The pass mark is 720 out of 1000 on a scaled score, so there is no fixed number of correct answers to hit. The questions come in four scenarios of 15, picked at random from a bank of six. The scenario text sits in a panel on the left and stays there for all 15 questions, so you read it once. Each question is multiple choice or multiple response and tells you how many answers to pick.

![The exam at a glance: 60 questions, 120 minutes, 4 scenarios of 15, 720 to pass, and the five domain weights](/guides/claude-certified-architect/exam-at-a-glance-v2.png)

Five areas share the points. Agentic architecture carries the most at 27%. Claude Code configuration and prompt engineering with structured output are 20% each, tool design and MCP is 18%, and context management and reliability is 15%. It surprised everyone I studied with how many questions are about an engineering team's process with Claude Code: CI pipelines, code review, permissions, hooks, built-in tools and structured output that feeds another system. Read each question as the CTO who wants that process to be predictable.

Anthropic's guide pictures the ideal candidate as a solution architect with six months or more of hands-on building experience. I am not that person. I build with Claude Code, and what I needed was to understand how a development team works with it. That is a different skill from writing the code, and a non-developer can learn it. I did.

A pass is valid for 12 months. Renewing before it lapses is a free, non-proctored assessment, and if it lapses you sit the full exam again at full price. You get up to four attempts per exam in any twelve months, with a waiting period after each fail. To register you need an email address at a company in the Claude Partner Network, and [Getting a seat](#registration) explains how to get one.

The next guide like this one can come by email. The box below is optional.


## Where to study: one site

One free website taught me everything the exam needed, and I would call it the only necessary resource. It has thirty lessons, one for each exam task. Each lesson has a "What you need to know" section and a practice scenario that explains why every option wins or loses, plus Exam Trap boxes: a wrong answer that looks right, and the one fact that makes it wrong. The lessons are free with no account. The mock exam became a one-time $9.95 purchase in October, and the people I studied with called it the closest thing to the real exam. The site's name and the order I read it in are in the [subscriber bonus](#registration).

The other thing to keep open is the official exam guide. Download it from the [certification page on the Anthropic Partner Academy](https://anthropic-partners.skilljar.com/claude-certified-architect-foundations-certification), or from the public [access-request page](https://anthropic.skilljar.com/claude-certified-architect-foundations-access-request) if you are not in the academy yet.

> [!TIP] The exam guide is the answer key
> The exam is scored against that document. Where today's Claude Code documentation and the guide's wording differ, the guide wins, so study the guide's wording.

What I skipped: the paid prep courses, such as [Learning Tree](https://www.learningtree.com/courses/claude-architect-foundations-prep/)'s three-day instructor-led course with a full mock at $2,785, Anthropic's extra Skilljar material beyond its free prep courses, and most third-party question banks. A course buys you structure and a cohort.

## How I prepared

I had about two and a half days, and they went into the site's theory. I answered each practice question before reading the lesson, and for lessons I had no time to read in full, I read only the Exam Trap boxes. I also kept one notes page. After every mistake I wrote one line in my own words, in a Notion page I could open on my phone: when the question says X, the answer is Y, not Z. I wrote down only mistakes, and I went through the page on the morning of the exam.

![Three of my one-line rules, in the format I used](/guides/claude-certified-architect/rules-card.png)

The last piece was Claude as a tutor. I pasted in the exam guide and the lesson text and asked it to explain every miss in plain words, with an analogy from my own work. Without that context, Claude answers from today's documentation, which is not always what the exam wants.

> [!TIP] The same setup as a free Claude Code plugin
> I packaged this method as an open-source Claude Code plugin, Cert Coach. It writes fresh practice questions from the public exam guide, explains every miss in plain words, and saves your one-line rules to a file on your own machine. It contains no exam questions. In Claude Code, run these one at a time:
>
> ```
> /plugin marketplace add mattzimak/claude-cert-coach
> /plugin install cert-coach@mattzimak-cert
> /cert-coach:foundations
> ```
>
> The code is on [GitHub](https://github.com/mattzimak/claude-cert-coach). No Claude Code? The tutor step above works in any Claude chat with the exam guide attached.

I did not take a single mock exam in the end. Time was short, and the lessons were worth more per hour. If you have a week, take one two days before the exam and learn from the misses, but never on the morning of the exam. The two and a half days went like this. On Monday and most of Tuesday I read prompt engineering and structured output, six lessons in full, and on Tuesday evening the six Claude Code configuration lessons. On Wednesday from 9:00 to 10:30 I read the five lessons on the topics I knew least, then went through the Exam Trap boxes and my rules page and started nothing new. I sat the exam that day, 30 September, in Los Angeles, at 13:45, and scored 851 out of 1000 against a pass mark of 720.

[![The certificate, issued 30 September 2026](/guides/claude-certified-architect/certificate.jpg)](https://www.credly.com/badges/9f11a83d-fbb3-430b-a46e-ccd0aa886abc/public_url)

Ten conventions have to be known cold, because two answer options will differ by exactly one of them. These are the ones I kept on one page. Open any to read it.


- **The agent loop.** It continues while the model asks for a tool and stops when the model says it is done. Read that signal from the response, never from the wording of the text.
- **Tool choice.** Three settings matter: let the model decide (it may answer in prose), force it to use some tool, or force one named tool.
- **The Batch API.** It is about half price and may take up to 24 hours. Use it for overnight work, never for a check that blocks a merge.
- **Hooks.** A pre-tool hook can stop an action before it happens. A post-tool hook runs after it, which is where formatting, linting and tests belong, with the failure fed back so the model fixes it.
- **CLAUDE.md files.** Files at different levels add up; they do not override each other. A file in your home folder is yours alone and never reaches the team.
- **MCP configuration.** The shared configuration is the `.mcp.json` file in the repository. The one in your home folder is personal.
- **Skills.** They are folders with a `SKILL.md` and load on demand. The guide treats their `allowed-tools` line as a restriction on what the skill may use.
- **Pipeline flags.** In a pipeline, `claude -p` runs without a human, `--output-format json` makes the result machine-readable, `--append-system-prompt` adds rules, and `--system-prompt` replaces them.
- **Built-in tools.** Glob finds files by name, Grep searches inside files, Read opens one. Discover step by step; do not read everything first.
- **Subagents.** A subagent starts with an empty memory. The coordinator hands over what it knows, and subagents never talk to each other.

## How I read each question

Two answers always look almost identical and differ by one technical detail. These seven habits are how I read every question.

- **1. Read the last line first.** What is being asked, and how many answers? Then read the scenario.
- **2. Read each scenario once.** It stays on the left for all 15 questions. Re-reading it per question is where the time goes.
- **3. Find the one deciding fact.** Every question has one: a number, a "must", "team-wide", "blocking", a log line. If you can answer without it, you are pattern-matching.
- **4. Must, always, guarantee: enforce it.** When a rule must hold on every run, the answer is something the system enforces, such as a permission rule, a hook or a schema, not a better-worded prompt. CLAUDE.md is advice. Settings and hooks are rules.
- **5. Pick the team-wide answer.** Shared config in the repo, a check in the pipeline, a separate review session. One person's workaround is usually the distractor.
- **6. Distrust the bigger option.** Reading everything "to be safe", a bigger model, a bigger context window and more AI agents are standing distractors. The narrower, cheaper option is right about a third of the time.
- **7. Spend two minutes a question.** That is 30 minutes per scenario. Flag anything over three minutes and move on. Never leave a blank: it scores the same as a wrong answer.

The questions that cost me the most time were the developer-heavy ones: GitHub workflows, and choosing between the built-in tools (Grep, Glob, Read, Bash). If you do not write code daily, budget a few extra minutes for those and move faster through the questions about prompts and orchestration.

![The seven reading habits on one card](/guides/claude-certified-architect/read-a-question.png)

Here is a worked example in the exam's style. It is my own question, not one from the exam. Pick an answer before you read on.

```quiz
A support team runs an AI agent that drafts refund decisions. Compliance requires that no refund above $500 is ever issued without a human sign-off, on every run, with no exceptions. Which change meets the requirement?
- [ ] A) State the rule in capitals at the top of the system prompt.
- [x] B) Add a pre-tool hook on the refund tool that blocks any amount above $500 and routes the case to a human.
- [ ] C) Instruct the agent to ask the customer to confirm amounts above $500.
- [ ] D) Move to a larger model with better instruction following.
> Last line first: one answer, and the question is about meeting a requirement. The deciding fact is "on every run, with no exceptions": the rule must be enforced, so it cannot live in a prompt. A and C ask the model to comply; D is the standing "bigger model" distractor. B is the only option where the client enforces the rule. Now flip one word: if the stem had said the team "wants the agent to be careful with large refunds", with no hard requirement, the prompt change would be the proportionate answer and the hook would be over-engineering. Same scenario, one word, different answer, and most questions on the paper work like that.
```

Five more like this one, each with its reasoning, are in the [subscriber bonus](#registration).

## Getting a seat

Registration needs an email address on the domain of a company in the Claude Partner Network. Check this before you pay for anything.

> [!IMPORTANT] You need a partner email
> A personal Gmail or Outlook address is refused with "You aren't signed in with a partner company employee email". If your employer is a partner but its domain is not recognised yet, partner-support@anthropic.com can link it. That takes 7 to 10 days, so ask early.

If your employer is not a partner, there are two ways in, and one of them is free. Both are in the subscriber bonus below, a free extra that opens with your email address: the exact steps and the 15-day timing that catches people, every registration screen in order, the route into the Professional exam through a project, the name of the study site and the order I read it in, five practice questions and a five-page cheat sheet. The rest of this guide stays open.


## Exam-day rules that cost people their attempt

> [!WARNING] The 48-hour rule
> Rescheduling and cancelling close 48 hours before the appointment. Inside that window you sit it or lose the fee. The July exam guide said 24 hours, but Anthropic's certification FAQ and Pearson both say 48 now, so plan for 48. After a reschedule, wait for Pearson's confirmation email, or the change may not have gone through.

- **Cancelling in Pearson is not a refund.** It only frees the slot. To get your money back, email certifications-support@anthropic.com.
- **Plan about 135 minutes in the seat** for the 120-minute exam: check-in, instructions and a short survey. Arrive after the late-arrival window (Pearson centres treat 15 minutes as that window) and the fee is forfeited.
- **Online proctoring** needs Pearson's system test passed on the machine you will use, and some corporate networks and managed laptops block it. If yours does, book a centre. IDs from Belarus, Cuba, North Korea, Russia, Syria or restricted regions of Ukraine require a test centre.
- **It is closed book.** No notes, no AI assistants, no browser translation tools. At a centre, your phone, watch and bag go in a locker.
- **Retake waits** are 14 days after a first fail, 30 after a second and 90 after a third, with at most four attempts per exam in a rolling twelve months. Each retake is the full fee, with your partner discount.
- **Accommodations** for documented needs exist, but Pearson VUE must approve them before you schedule, not after.

You accept a non-disclosure agreement before the first question. Nothing in this guide is exam content, and the worked example above is mine.

## What other people who passed say

I collected 127 LinkedIn posts from people who sat the Claude exams between March and October 2026 and read the ones with real advice in them. The same lessons came up again and again, and most match what happened to me. Open any for the details and the post:


- **Practice scores mislead.** Andrey Arykov scored 88 to 95% on eight practice exams, then 690 on the real one, and 898 on the second attempt with a different preparation method. Arykov's explanation: practice exams test whether you remember the answer, while the real exam asks which option is best under the constraint in the question. [Post](https://www.linkedin.com/posts/andrey-arykov_ccarf-claudecertifiedarchitect-anthropiccertification-activity-7515022145178206208-iu4b)
- **Two answers look right.** Chidiebube Onah, an engineering team lead who passed on the first attempt after two focused weeks, found the hard questions were the ones where two answers were both technically correct and the job was to pick the one that fits the architecture principles in Anthropic's material. [Post](https://www.linkedin.com/posts/chidiebube-onah_i-passed-the-claude-certified-architect-foundations-activity-7493939477053272064-INzV)
- **The official courses alone are not enough.** Vitaliy Kalinichenko compared the three technical exams and rated Architect - Foundations the hardest, with the broadest scope and the most plausible wrong answers. His estimate is that the official courses cover 20 to 30% of what the exam asks. [Post](https://www.linkedin.com/posts/vitalii-kalinichenko_anthropic-ccarf-activity-7515705171335622657-Pa__)
- **Non-engineers pass.** Santanu Majumdar, a technical programme manager who does not ship code, scored 904 after a month of preparation. Almost half the score sits in two domains, and the wrong answers are written to sound like good engineering decisions. [Post](https://www.linkedin.com/posts/santanumajumdar_claudecertifiedarchitect-aicertification-activity-7455882458308448257-rnYk)
- **Start with the official exam guide.** Adhithiyan Sathasivam, who passed all four Claude exams in two months, read each exam guide and prep course first and used other resources only to fill gaps. [Post](https://www.linkedin.com/posts/adhithiyan-sathasivam-b96223b3_over-the-past-two-months-i-completed-anthropics-activity-7519315181215318016-93FB)
- **Non-native speakers, watch the vocabulary.** Qin Li, who passed the Professional exam, found a handful of uncommon English words tricky. [The ten words below](#vocabulary) are the ones that tripped me. [Post](https://www.linkedin.com/posts/qinli11_claude-ccarp-claudecertified-activity-7518899022401658880-JyVP)
- **Leave feedback on bad questions.** Anton Rogozin points out that after you finish you can comment on individual questions without it costing exam time. If a question looks wrong to you, say so there. [Post](https://www.linkedin.com/posts/anton-rogozin_hi-linkedin-im-a-little-surprised-by-the-activity-7519707588293439488-79xB)
- **There is no official mock.** As Michael F. notes, Anthropic publishes a handful of sample questions in the exam guide and no practice exam, which is why so many community mocks exist. [Post](https://www.linkedin.com/posts/michael-f-f_claude-anthropic-aiengineering-activity-7518217822880235522-tk68)

## Ten words the exam assumes you know

The exam is written for people who ship software. These were the words I had to look up, in the sense the exam uses them. The same words are underlined where they first come up in this guide: tap or hover one for its definition, or open any row below.


- **CI/CD, pipeline.** The automated assembly line a code change travels down: submit it, and a pipeline builds it, tests it and ships it, with nobody watching. Think of an automated workflow whose trigger is "someone submitted code".
- **Pull request (PR).** A proposed code change waiting for review before it joins the main codebase.
- **Repository, repo root.** The project folder with all the code and its history. The root is its top level, where shared configuration files live.
- **Non-interactive, headless.** Running without a human present to answer prompts. In Claude Code that is the `-p` flag.
- **Session.** One continuous conversation with its own memory. "A separate session" means starting fresh with nothing carried over, which is why a review should not happen in the session that wrote the code.
- **Hook.** A small piece of code that fires at a fixed moment, such as before or after a tool call. Hooks are enforced by the client; prompts are not.
- **Schema.** A declared shape for data: which fields exist, their types, which are required. A schema is how you guarantee the format of an answer.
- **Glob.** A filename pattern with wildcards. `**/*.test.ts` means every test file in any folder.
- **Stateless.** The server remembers nothing between requests, so every request carries the whole conversation. That is why context size is a design problem.
- **Frontmatter.** The small settings block at the top of a markdown file, between two `---` lines, where things like a skill's allowed tools are declared.

## What it cost, and the questions people ask me

The exam is $125 per attempt, less for partner tiers (the last question below has them). The route I used added a CHF 100 membership of the association that gave me my partner email. The study site was free when I used it, and today its mock exam is $9.95. Nothing else.


**What is the Claude Certified Architect exam?** Anthropic's own proctored exam on building systems with Claude. The Foundations exam has 60 scenario questions in 120 minutes, covers Claude Code, the Claude Agent SDK, the Claude API and MCP, and needs 720 out of 1000 to pass.

**Is the Claude Certified Architect exam worth it?** Yes, if you build with Claude or sell Claude work to clients. It is a hard, proctored exam that Anthropic's Applied AI team helped design, and it certifies one vendor's tools, not AI in general. It is worth the investment, especially if you do not come from an engineering background.

**Is the exam hard?** Hard on reading, not on knowledge. Two answers usually differ by one detail. If you read carefully and know the conventions above, a non-developer can pass. I did.

**How long should I prepare?** Plan a focused week with the site. I had about two and a half days, which is tight for someone who does not write code.

**What if I fail?** The score report lists every objective with a percentage and stays in your Pearson account under View score reports. Read the lessons behind your weakest objectives in depth, write your one-line rules, and book the retake for the earliest date you can prepare for properly. The waits after a fail are in the [exam-day rules](#exam-day).

**Do I need to be a developer?** No. You need to understand how a development team works with Claude Code: pipelines, reviews, permissions, hooks. [The vocabulary list](#vocabulary) covers the words, and the site covers the rest.

**Do I need to be in the Partner Network?** Yes. Registration needs a partner email on a recognised company domain ([the seat section](#registration) has the details). If your employer is not a partner, there are two ways in. Anthropic says a firm can join the network directly for free, and the route I took, through a membership association, is in the subscriber bonus with the steps.

**Online or test centre?** Both are offered. A centre removes the room-scan and connection risks; online saves the trip. Either way the name on the booking must match your ID.

**Does the badge show the score?** No. Credly shows pass and dates only. Mine was 851.

**How long is it valid?** Twelve months, then a free non-proctored renewal if you do it on time.

**Is it free for partners?** For some. Global Premier partners pay nothing until the end of 2026, and Select and Preferred partners get half off. Registered-tier partners and their members pay the full price.

Made it through? Tag [@mattzimak.ai](https://www.instagram.com/mattzimak.ai) and tell me your score.