Which stop_reason values drive an agentic loop, and what does each one mean?

A response whose stop_reason is tool_use means Claude wants a tool run: execute it and send the result back. end_turn means Claude has finished. Branch on this field, never on whether the reply happens to contain text.

Name three anti-patterns for deciding when an agentic loop should stop

Reading natural-language cues such as the word "done", treating a fixed iteration cap as the main stopping rule, and assuming any assistant text means completion. A single response can carry prose and a tool call at the same time.

In a hub-and-spoke multi-agent system, who is allowed to talk to whom?

Every subagent reports to the coordinator and to nobody else; subagents never message each other directly. The hub decomposes the task, delegates, aggregates results and owns error handling, which keeps information flow observable and consistent.

What usually causes a multi-agent report to miss whole subtopics when every subagent succeeded?

The coordinator decomposed the topic too narrowly, so nobody was ever assigned the missing areas. Fix the decomposition, then add a refinement loop that checks the synthesis for gaps and re-delegates targeted queries until coverage is sufficient.

What must a coordinator's allowedTools contain before it can spawn subagents?

The Task tool. Subagents are spawned by Task calls, so a coordinator whose allowed tool list omits Task cannot delegate at all, however clearly its prompt describes the subagent roles it expects to use.

How do you run several subagents in parallel from one coordinator?

Have the coordinator issue every Task call it needs inside one response rather than one per turn; the subagents then run concurrently. Each prompt must still carry its own complete context, since none of them share memory.

Prompt instruction or programmatic gate: which one enforces a mandatory tool order?

A programmatic gate. Instructions such as "always verify identity first" are followed most of the time, and a non-zero failure rate is unacceptable before a financial action. Block the downstream tool in code until the prerequisite has returned.

What does a PostToolUse hook do in the Agent SDK?

It runs after a tool returns and can transform the result before the model sees it. Typical use: normalising mixed formats from different MCP tools, such as Unix timestamps, ISO 8601 dates and numeric status codes, into one shape.

Prompt chaining or adaptive decomposition: how do you choose between them?

Chain fixed steps when the work is predictable, such as reviewing each file and then running a cross-file pass. Decompose adaptively when the next subtask depends on what the last one found, as in adding tests to an unfamiliar legacy codebase.

When should you resume a session, and when start fresh with a summary?

Resume with --resume and the session name when the earlier context is still mostly valid, telling the agent which files changed so it re-reads only those. Start a new session with a structured summary when the old tool results are stale.

What does fork_session give you in the Agent SDK?

Independent branches that start from the same analysed baseline. Fork once the shared exploration is done, then try two refactoring or testing strategies side by side. Neither branch's changes leak into the other, and the shared analysis is never repeated.

What is the primary signal Claude uses to choose between similar tools?

The tool description. A one-line description such as retrieves details leaves nothing to separate near-identical tools, so selection becomes unreliable. State the inputs accepted, example requests, edge cases and when to prefer a neighbouring tool instead.

Two tools with near-identical descriptions keep getting confused. Name the two fixes.

Rename them so the names state distinct purposes, and rewrite each description with a specific scope and boundary. If one tool is genuinely doing several jobs, split it into purpose-specific tools with their own input and output contracts.

Which fields belong in a structured MCP error response?

Set isError, then give an error category (transient, validation, business, permission), a retryable boolean and a human-readable explanation. Without these the agent cannot tell a timeout worth retrying from a policy refusal it should explain to the customer.

Why does giving an agent every available tool make it less reliable?

Every added tool is another option to weigh on each turn, so selection accuracy degrades as the list grows; four or five well-scoped tools beat eighteen. Agents also tend to misuse tools outside their role, such as a synthesis agent running web searches.

Where do shared and personal MCP servers live for Claude Code?

Shared team servers go in the project's .mcp.json, committed to version control with credentials referenced as environment variables like ${GITHUB_TOKEN}. Personal or experimental servers go in the user-level ~/.claude.json. Tools from every configured server are available at once.

Grep or Glob: which built-in tool finds what?

Grep searches file contents for a pattern: function names, error strings, import statements. Glob matches file paths by name or extension pattern, such as every .test.tsx file. Use Glob to find files and Grep to find what is inside them.

Edit reports that the anchor text is not unique. What is the fallback?

Read the whole file, then Write it back with the change applied. Edit depends on matching one unique span of text, so repeated text defeats it; Read followed by Write gives a reliable modification without guessing at a new anchor.

Name the three levels of the CLAUDE.md hierarchy and who each one reaches

User level at ~/.claude/CLAUDE.md, seen only by you and never shared through git. Project level at the repository root or .claude/CLAUDE.md, shared with everyone who clones. Directory level in subdirectories for area-specific guidance. Run /memory to see which files loaded.

How do you keep a large CLAUDE.md modular instead of monolithic?

Reference external files with the @import syntax so each package pulls in only the standards it needs, or move topics into separate files under .claude/rules/, such as testing.md and deployment.md. Both keep universal instructions short and specific ones findable.

Which SKILL.md frontmatter fields restrict tools and prompt for missing arguments?

allowed-tools limits what the skill may call while it runs, for example file writes only so nothing destructive executes. argument-hint tells a developer what to pass when they invoke the skill without arguments. Skills load on demand; CLAUDE.md is loaded every session.

Where do team-shared slash commands live, and where do personal ones live?

Team commands go in .claude/commands/ inside the repository, so version control delivers them to everyone who pulls. Personal commands go in ~/.claude/commands/ and reach nobody else. Putting a shared command in the home directory is the classic mistake.

When do path-scoped rules beat a subdirectory CLAUDE.md?

When one convention applies to files scattered across many directories, such as test files beside the code they test. A rule file under .claude/rules/ with a paths glob in its YAML frontmatter loads only while matching files are edited, saving tokens.

Which tasks call for plan mode rather than direct execution?

Large-scale or multi-file changes, tasks with several valid approaches, and anything with architectural consequences, such as splitting a monolith or migrating a library across dozens of files. A one-function fix with a clear stack trace runs directly.

What beats more prose when Claude keeps misreading a transformation requirement?

Two or three concrete input and output pairs. Examples pin down the exact mapping where descriptions leave room for interpretation, and a specific failing case with its expected output is the fastest way to correct edge-case handling such as null values.

Which CLI flags run Claude Code non-interactively with machine-readable output?

-p (or --print) runs one prompt and exits instead of waiting for input, which is what stops a CI job hanging. --output-format json with --json-schema returns findings in a fixed structure that a script can post as inline comments.

Why does telling a review prompt to be conservative not reduce false positives?

Because it names no criterion. Precision improves when the prompt says which categories to report (bugs, security) and which to skip (style, local patterns), with concrete code examples for each severity level. Confidence-based filtering leaves the boundary undefined.

How many few-shot examples should you write, and what should each one show?

Two to four targeted examples, each covering an ambiguous case and showing why one action was chosen over the plausible alternative. Demonstrating the reasoning lets the model generalise to novel patterns instead of matching only the cases you listed.

What do tool_choice auto, any and a named tool each guarantee?

auto lets Claude reply in text or call a tool, so structured output is not guaranteed. any guarantees a tool call and lets Claude choose the tool. A named tool, type tool with its name, forces that exact tool this turn.

Why make an extraction field nullable rather than required?

A required field must be filled, so when the document lacks the value the model may invent one to satisfy the schema. Nullable fields let it report absence honestly. For categories, add an other value with a free-text detail field.

Which extraction failures can a retry fix, and which can it never fix?

Tool-use schemas already remove syntax errors; what remains is semantic. Retrying with the specific validation error appended fixes format and structural faults such as a total that does not sum. It cannot recover a value the source never contained.

Name the trade-offs of the Message Batches API

Roughly half the cost of synchronous calls, results arriving any time inside 24 hours, and no latency guarantee. That suits overnight reports and weekly audits, and rules out anything a person is waiting on, such as a pre-merge check.

A batch job has failures. How do you resubmit efficiently?

Match each result to its request by custom_id, resubmit only the failed ones, and change what caused the failure first, for example chunking a document that overran the context window. Refine the prompt on a sample before the next full run.

Why is a fresh Claude instance a better code reviewer than the one that wrote the code?

The generating session still holds the reasoning that produced the code, so it tends to defend its decisions rather than question them. An independent instance without that context catches subtler issues that neither a self-review instruction nor extended thinking reliably surfaces.

What is the lost in the middle effect, and how do you counter it?

Models attend reliably to the start and end of a long input and may drop findings from the middle. Put a summary of key findings first, mark each section with an explicit header, and trim tool outputs to the fields that matter.

Name three triggers that should send a support case to a human

An explicit request for a person, which you honour at once without investigating first; a policy gap or ambiguity around the request; and an inability to make meaningful progress. Negative sentiment and self-rated confidence are not reliable triggers.

What should a subagent return when it fails, so the coordinator can recover?

Structured error context: the failure type, what was attempted, any partial results, and possible alternatives. Retry transient faults locally first and propagate only what cannot be resolved. Never return an empty set marked as success, and never abort the whole workflow.

How do you fight context degradation during a long codebase exploration?

Keep a scratchpad file of key findings and re-read it rather than trusting memory, delegate verbose investigation to subagents that return summaries, inject a phase summary before spawning the next phase, and run /compact when discovery output fills the window.

Why is 97 percent overall extraction accuracy not enough to drop human review?

An aggregate can hide a document type or field where accuracy is far lower. Measure by segment first, calibrate field-level confidence scores against a labelled validation set, and keep stratified random samples of high-confidence output under review to catch new error patterns.

Two credible sources give different figures. What does the synthesis agent do?

Keep both values, each annotated with its source, and let the coordinator decide how to reconcile rather than picking one silently. Require every structured finding to carry its publication or data collection date, since a temporal gap often explains what looks like a contradiction.

No account needed

CCAR-F flashcards: 40 free cards across the five exam domains

Forty recall cards for the Claude Certified Architect - Foundations exam, one deck spread across its five official domains in proportion to their weights. Flip a card at a time, shuffle the order, or copy the whole deck as tab-separated text for Anki or a spreadsheet. Every front and back is on this page.

40 cardsfive domainsNo accountNothing stored

How the deck is built

The CCAR-F blueprint has five domains, each with its published objectives. Each card on this page was written from the official Exam Guide version 1.0 (effective July 2026) and the objective it names, then checked against the guide on the date at the foot of the page. The deck is not exam content: nothing is recalled from a sitting, and nothing is drawn from the signed-in study chapters or the gated question bank. The allocation follows the official weights, so the heaviest domain, Agentic Architecture & Orchestration, gets the most cards and the lightest, Context Management & Reliability, the fewest.

Domain	Official weight	Cards
1. Agentic Architecture & Orchestration	27%	11
2. Tool Design & MCP Integration	18%	7
3. Claude Code Configuration & Workflows	20%	8
4. Prompt Engineering & Structured Output	20%	8
5. Context Management & Reliability	15%	6

Study one card at a time

Read the front, decide your answer out loud or on paper, then flip. Shuffle once the order stops surprising you. The widget keeps your place only until you reload; nothing is saved.

Card 1 of 40

Tap the card to flip. Arrow keys move between cards.

The whole deck, by domain

All 40 cards, front beside back, grouped by official domain with the objective each one drills. Copy the deck to your clipboard as tab-separated text with a header row.

Domain 1: Agentic Architecture & Orchestration

27% of the exam, 11 cards

Which stop_reason values drive an agentic loop, and what does each one mean?: A response whose stop_reason is tool_use means Claude wants a tool run: execute it and send the result back. end_turn means Claude has finished. Branch on this field, never on whether the reply happens to contain text.; Objective O01: Design and implement agentic loops for autonomous task execution
Name three anti-patterns for deciding when an agentic loop should stop: Reading natural-language cues such as the word "done", treating a fixed iteration cap as the main stopping rule, and assuming any assistant text means completion. A single response can carry prose and a tool call at the same time.; Objective O01: Design and implement agentic loops for autonomous task execution
In a hub-and-spoke multi-agent system, who is allowed to talk to whom?: Every subagent reports to the coordinator and to nobody else; subagents never message each other directly. The hub decomposes the task, delegates, aggregates results and owns error handling, which keeps information flow observable and consistent.; Objective O02: Orchestrate multi-agent systems with coordinator-subagent patterns
What usually causes a multi-agent report to miss whole subtopics when every subagent succeeded?: The coordinator decomposed the topic too narrowly, so nobody was ever assigned the missing areas. Fix the decomposition, then add a refinement loop that checks the synthesis for gaps and re-delegates targeted queries until coverage is sufficient.; Objective O02: Orchestrate multi-agent systems with coordinator-subagent patterns
What must a coordinator's allowedTools contain before it can spawn subagents?: The Task tool. Subagents are spawned by Task calls, so a coordinator whose allowed tool list omits Task cannot delegate at all, however clearly its prompt describes the subagent roles it expects to use.; Objective O03: Configure subagent invocation, context passing, and spawning
How do you run several subagents in parallel from one coordinator?: Have the coordinator issue every Task call it needs inside one response rather than one per turn; the subagents then run concurrently. Each prompt must still carry its own complete context, since none of them share memory.; Objective O03: Configure subagent invocation, context passing, and spawning
Prompt instruction or programmatic gate: which one enforces a mandatory tool order?: A programmatic gate. Instructions such as "always verify identity first" are followed most of the time, and a non-zero failure rate is unacceptable before a financial action. Block the downstream tool in code until the prerequisite has returned.; Objective O04: Implement multi-step workflows with enforcement and handoff patterns
What does a PostToolUse hook do in the Agent SDK?: It runs after a tool returns and can transform the result before the model sees it. Typical use: normalising mixed formats from different MCP tools, such as Unix timestamps, ISO 8601 dates and numeric status codes, into one shape.; Objective O05: Apply Agent SDK hooks for tool call interception and data normalization
Prompt chaining or adaptive decomposition: how do you choose between them?: Chain fixed steps when the work is predictable, such as reviewing each file and then running a cross-file pass. Decompose adaptively when the next subtask depends on what the last one found, as in adding tests to an unfamiliar legacy codebase.; Objective O06: Design task decomposition strategies for complex workflows
When should you resume a session, and when start fresh with a summary?: Resume with --resume and the session name when the earlier context is still mostly valid, telling the agent which files changed so it re-reads only those. Start a new session with a structured summary when the old tool results are stale.; Objective O07: Manage session state, resumption, and forking
What does fork_session give you in the Agent SDK?: Independent branches that start from the same analysed baseline. Fork once the shared exploration is done, then try two refactoring or testing strategies side by side. Neither branch's changes leak into the other, and the shared analysis is never repeated.; Objective O07: Manage session state, resumption, and forking

Domain 2: Tool Design & MCP Integration

18% of the exam, 7 cards

What is the primary signal Claude uses to choose between similar tools?: The tool description. A one-line description such as retrieves details leaves nothing to separate near-identical tools, so selection becomes unreliable. State the inputs accepted, example requests, edge cases and when to prefer a neighbouring tool instead.; Objective O08: Design effective tool interfaces with clear descriptions and boundaries
Two tools with near-identical descriptions keep getting confused. Name the two fixes.: Rename them so the names state distinct purposes, and rewrite each description with a specific scope and boundary. If one tool is genuinely doing several jobs, split it into purpose-specific tools with their own input and output contracts.; Objective O08: Design effective tool interfaces with clear descriptions and boundaries
Which fields belong in a structured MCP error response?: Set isError, then give an error category (transient, validation, business, permission), a retryable boolean and a human-readable explanation. Without these the agent cannot tell a timeout worth retrying from a policy refusal it should explain to the customer.; Objective O09: Implement structured error responses for MCP tools
Why does giving an agent every available tool make it less reliable?: Every added tool is another option to weigh on each turn, so selection accuracy degrades as the list grows; four or five well-scoped tools beat eighteen. Agents also tend to misuse tools outside their role, such as a synthesis agent running web searches.; Objective O10: Distribute tools appropriately across agents and configure tool choice
Where do shared and personal MCP servers live for Claude Code?: Shared team servers go in the project's .mcp.json, committed to version control with credentials referenced as environment variables like ${GITHUB_TOKEN}. Personal or experimental servers go in the user-level ~/.claude.json. Tools from every configured server are available at once.; Objective O11: Integrate MCP servers into Claude Code and agent workflows
Grep or Glob: which built-in tool finds what?: Grep searches file contents for a pattern: function names, error strings, import statements. Glob matches file paths by name or extension pattern, such as every .test.tsx file. Use Glob to find files and Grep to find what is inside them.; Objective O12: Select and apply built-in tools (Read, Write, Edit, Bash, Grep, Glob) effectively
Edit reports that the anchor text is not unique. What is the fallback?: Read the whole file, then Write it back with the change applied. Edit depends on matching one unique span of text, so repeated text defeats it; Read followed by Write gives a reliable modification without guessing at a new anchor.; Objective O12: Select and apply built-in tools (Read, Write, Edit, Bash, Grep, Glob) effectively

Domain 3: Claude Code Configuration & Workflows

20% of the exam, 8 cards

Name the three levels of the CLAUDE.md hierarchy and who each one reaches: User level at ~/.claude/CLAUDE.md, seen only by you and never shared through git. Project level at the repository root or .claude/CLAUDE.md, shared with everyone who clones. Directory level in subdirectories for area-specific guidance. Run /memory to see which files loaded.; Objective O13: Configure CLAUDE.md files with appropriate hierarchy, scoping, and modular organization
How do you keep a large CLAUDE.md modular instead of monolithic?: Reference external files with the @import syntax so each package pulls in only the standards it needs, or move topics into separate files under .claude/rules/, such as testing.md and deployment.md. Both keep universal instructions short and specific ones findable.; Objective O13: Configure CLAUDE.md files with appropriate hierarchy, scoping, and modular organization
Which SKILL.md frontmatter fields restrict tools and prompt for missing arguments?: allowed-tools limits what the skill may call while it runs, for example file writes only so nothing destructive executes. argument-hint tells a developer what to pass when they invoke the skill without arguments. Skills load on demand; CLAUDE.md is loaded every session.; Objective O14: Create and configure custom slash commands and skills
Where do team-shared slash commands live, and where do personal ones live?: Team commands go in .claude/commands/ inside the repository, so version control delivers them to everyone who pulls. Personal commands go in ~/.claude/commands/ and reach nobody else. Putting a shared command in the home directory is the classic mistake.; Objective O14: Create and configure custom slash commands and skills
When do path-scoped rules beat a subdirectory CLAUDE.md?: When one convention applies to files scattered across many directories, such as test files beside the code they test. A rule file under .claude/rules/ with a paths glob in its YAML frontmatter loads only while matching files are edited, saving tokens.; Objective O15: Apply path-specific rules for conditional convention loading
Which tasks call for plan mode rather than direct execution?: Large-scale or multi-file changes, tasks with several valid approaches, and anything with architectural consequences, such as splitting a monolith or migrating a library across dozens of files. A one-function fix with a clear stack trace runs directly.; Objective O16: Determine when to use plan mode vs direct execution
What beats more prose when Claude keeps misreading a transformation requirement?: Two or three concrete input and output pairs. Examples pin down the exact mapping where descriptions leave room for interpretation, and a specific failing case with its expected output is the fastest way to correct edge-case handling such as null values.; Objective O17: Apply iterative refinement techniques for progressive improvement
Which CLI flags run Claude Code non-interactively with machine-readable output?: -p (or --print) runs one prompt and exits instead of waiting for input, which is what stops a CI job hanging. --output-format json with --json-schema returns findings in a fixed structure that a script can post as inline comments.; Objective O18: Integrate Claude Code into CI/CD pipelines

Domain 4: Prompt Engineering & Structured Output

20% of the exam, 8 cards

Why does telling a review prompt to be conservative not reduce false positives?: Because it names no criterion. Precision improves when the prompt says which categories to report (bugs, security) and which to skip (style, local patterns), with concrete code examples for each severity level. Confidence-based filtering leaves the boundary undefined.; Objective O19: Design prompts with explicit criteria to improve precision and reduce false positives
How many few-shot examples should you write, and what should each one show?: Two to four targeted examples, each covering an ambiguous case and showing why one action was chosen over the plausible alternative. Demonstrating the reasoning lets the model generalise to novel patterns instead of matching only the cases you listed.; Objective O20: Apply few-shot prompting to improve output consistency and quality
What do tool_choice auto, any and a named tool each guarantee?: auto lets Claude reply in text or call a tool, so structured output is not guaranteed. any guarantees a tool call and lets Claude choose the tool. A named tool, type tool with its name, forces that exact tool this turn.; Objective O21: Enforce structured output using tool use and JSON schemas
Why make an extraction field nullable rather than required?: A required field must be filled, so when the document lacks the value the model may invent one to satisfy the schema. Nullable fields let it report absence honestly. For categories, add an other value with a free-text detail field.; Objective O21: Enforce structured output using tool use and JSON schemas
Which extraction failures can a retry fix, and which can it never fix?: Tool-use schemas already remove syntax errors; what remains is semantic. Retrying with the specific validation error appended fixes format and structural faults such as a total that does not sum. It cannot recover a value the source never contained.; Objective O22: Implement validation, retry, and feedback loops for extraction quality
Name the trade-offs of the Message Batches API: Roughly half the cost of synchronous calls, results arriving any time inside 24 hours, and no latency guarantee. That suits overnight reports and weekly audits, and rules out anything a person is waiting on, such as a pre-merge check.; Objective O23: Design efficient batch processing strategies
A batch job has failures. How do you resubmit efficiently?: Match each result to its request by custom_id, resubmit only the failed ones, and change what caused the failure first, for example chunking a document that overran the context window. Refine the prompt on a sample before the next full run.; Objective O23: Design efficient batch processing strategies
Why is a fresh Claude instance a better code reviewer than the one that wrote the code?: The generating session still holds the reasoning that produced the code, so it tends to defend its decisions rather than question them. An independent instance without that context catches subtler issues that neither a self-review instruction nor extended thinking reliably surfaces.; Objective O24: Design multi-instance and multi-pass review architectures

Domain 5: Context Management & Reliability

15% of the exam, 6 cards

What is the lost in the middle effect, and how do you counter it?: Models attend reliably to the start and end of a long input and may drop findings from the middle. Put a summary of key findings first, mark each section with an explicit header, and trim tool outputs to the fields that matter.; Objective O25: Manage conversation context to preserve critical information across long interactions
Name three triggers that should send a support case to a human: An explicit request for a person, which you honour at once without investigating first; a policy gap or ambiguity around the request; and an inability to make meaningful progress. Negative sentiment and self-rated confidence are not reliable triggers.; Objective O26: Design effective escalation and ambiguity resolution patterns
What should a subagent return when it fails, so the coordinator can recover?: Structured error context: the failure type, what was attempted, any partial results, and possible alternatives. Retry transient faults locally first and propagate only what cannot be resolved. Never return an empty set marked as success, and never abort the whole workflow.; Objective O27: Implement error propagation strategies across multi-agent systems
How do you fight context degradation during a long codebase exploration?: Keep a scratchpad file of key findings and re-read it rather than trusting memory, delegate verbose investigation to subagents that return summaries, inject a phase summary before spawning the next phase, and run /compact when discovery output fills the window.; Objective O28: Manage context effectively in large codebase exploration
Why is 97 percent overall extraction accuracy not enough to drop human review?: An aggregate can hide a document type or field where accuracy is far lower. Measure by segment first, calibrate field-level confidence scores against a labelled validation set, and keep stratified random samples of high-confidence output under review to catch new error patterns.; Objective O29: Design human review workflows and confidence calibration
Two credible sources give different figures. What does the synthesis agent do?: Keep both values, each annotated with its source, and let the coordinator decide how to reconcile rather than picking one silently. Require every structured finding to carry its publication or data collection date, since a temporal gap often explains what looks like a contradiction.; Objective O30: Preserve information provenance and handle uncertainty in multi-source synthesis

How to use the cards in a study week

Run the whole deck once a day for a week, unshuffled on day one so the domains stay together, shuffled from day two so the domain tag stops giving the answer away. Say the back before you flip; a card you can only recognise, not produce, is not learned yet. Pull the cards you miss into their own list and read the matching domain section of the CCAR-F cheat sheet, which holds the decision rules the cards compress, then work through the ten CCAR-F practice questions with the reasoning written out. Then sit the CCAR-F mock, a 20-question quick set: recall is the floor, and the exam tests the judgment built on top of it. The week-by-week study plan says where the cards fit in a longer run.

What a flashcard cannot tell you

The exam reports a scaled score with a 720 cut and a percent-correct by domain. Knowing every card here says nothing about either, because the items are scenarios that ask you to choose between two defensible options on a stated constraint. Flashcard reviews are not measured answers, so they never move a readiness score on Cred Farmer; only measured answers do.

Where to go next

The deck is an independent study aid written from the published exam guide. The CCAR-F exam is 60 items in 120 minutes for a $125 fee, and these cards are not its questions. Read the originality note on why no page here carries recalled exam content.

Cards learned? Test the judgment

The CCAR-F mock is a 20-question quick set, written to the five-domain guide and timed at two minutes an item, with a raw count by domain and a reason for every option. No account, nothing stored.

Sit the CCAR-F mock, no account

Frequently asked

Quick answers to the follow-up questions.

Do I need an account to use the CCAR-F flashcards?

No. All 40 cards are on this page, every front and every back, with nothing held back and no limit on how often you come back. Everything on Cred Farmer is free. An account adds the CCAR-F study sections, drills by objective and saved progress. This deck is complete as it stands and nothing here unlocks later.

Is anything I do on this page stored?

No. Flip, shuffle and copy work in your browser without an email address or a cookie that follows you. Nothing you do on this page is sent anywhere or stored, which is also why your position in the deck resets when you reload.

Do the cards come from the live CCAR-F exam?

No. Every card was written for this page from the official CCAR-F Exam Guide version 1.0 and its published objectives. Nothing is recalled from a sitting and nothing is copied from another provider. Disclosing live exam content breaks the candidate agreement, and Cred Farmer does not use, solicit or accept it.

How many flashcards do I need for CCAR-F?

Fewer than you would think. The blueprint lists its objectives under five weighted domains, and these 40 cards give each domain a share in proportion to its weight. A deck you can finish in twenty minutes and repeat daily beats a deck of several hundred that you review once.

Can I download or print the flashcards?

Use the copy button to put the whole deck on your clipboard as tab-separated text with a header row, then paste it into Anki, Quizlet, Google Sheets or Excel. The page also prints cleanly: the study widget is hidden and the full list, front beside back, is what comes out.

Are flashcards enough to pass CCAR-F?

On their own, no. The exam asks scenario questions that test judgment between two defensible options, and recall cards only prove you know the terms. Use the deck to fix the vocabulary and decision rules, then sit the CCAR-F mock, a 20-question quick set, no account needed, to see how that knowledge holds up under the clock.

Checked against the official exam guides on 17 September 2026.

Not affiliated with, or endorsed by, Anthropic or Pearson VUE. Details are summarised from publicly published program information and can change. Always confirm against the official exam guide before booking.