Claude Certification Blog
A CCAR-F cheat sheet: 25 decision rules for a five-domain scenario exam
This CCAR-F cheat sheet reduces the five-domain, scenario-framed architect exam to decision rules you can hold under time: the official weights, five rules per domain with the trap each catches, the terms the items lean on, and a final-week checklist.
A CCAR-F cheat sheet has to start from the flattest blueprint in the programme: Agentic Architecture and Orchestration 27%, Claude Code Configuration and Workflows 20%, Prompt Engineering and Structured Output 20%, Tool Design and MCP Integration 18%, Context Management and Reliability 15%. No domain can be skipped. Below are 25 CCAR-F decision rules, five per domain, each with the trap it catches, then a traps table, a quick reference of terms and a seven-day checklist. Every rule is written from the official exam guide’s task statements, and none is an answer.
How the CCAR-F cheat sheet is weighted
The CCAR-F guide publishes five domain weights and 30 task statements, and unlike the developer paper it publishes no skill-level weights, so the domain is the finest grain available. The item column below is the nearest whole-number split of 60 and is arithmetic rather than a published count. The objective column is the guide’s own list, and it matters: the largest domain also has the most objectives, so its extra items are spread thin rather than piled onto one topic.
| Domain | Official weight | Of 60 items | Task statements | Rules below |
|---|---|---|---|---|
| 1. Agentic Architecture and Orchestration | 27% | 16 | 7 | 5 |
| 2. Tool Design and MCP Integration | 18% | 11 | 5 | 5 |
| 3. Claude Code Configuration and Workflows | 20% | 12 | 6 | 5 |
| 4. Prompt Engineering and Structured Output | 20% | 12 | 6 | 5 |
| 5. Context Management and Reliability | 15% | 9 | 6 | 5 |
The other structural fact is the scenario frame: four scenarios are presented from a published bank of six, and each scenario names its primary domains. Three of the six lean on Agentic Architecture, three on Tool Design and three on Claude Code Configuration; four name Context Management, more than any other domain, despite its being the lightest. The post on the six scenarios works through that map. The twelve rules in the Claude exam cheat sheet are assumed here rather than repeated, and the two-week CCAR-F study plan turns the same weights into a near-even fortnight.
Agentic Architecture and Orchestration: five rules for 27%
Seven task statements and about sixteen items, so roughly two items per objective and no room to master three and skip four. The domain covers the loop itself, coordinator and subagent topologies, enforcement and handoff, hooks, decomposition and session state. Two of the guide’s twelve sample items sit here, and both turn on the same distinction: whether an option changes what the system can do or only what it is asked to do.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| The loop stops on stop_reason, never on the text | Continue while the model asks for a tool; stop when it ends its turn. Every tool result goes back into the conversation before the next call. | Loop control that parses the reply for a finished phrase, or an iteration cap used as the primary stopping mechanism. |
| A subagent knows only what its prompt contains | Workers run in isolated context and inherit nothing, so the coordinator passes complete findings, structured with source metadata, and its own tool list must permit spawning them. | A synthesis agent expected to know what the search agent found without being handed it. |
| When compliance must be guaranteed, enforce it in code | A prerequisite gate or a hook that blocks the call is deterministic; a prompt instruction has a non-zero failure rate. Financial and identity steps get the gate. | Strengthen the system prompt so that verification is described as mandatory before a refund. |
| Decompose to the shape of the work | Predictable multi-aspect reviews chain fixed steps; open-ended investigation adapts its subtasks to what it finds. A coordinator that cuts the topic too narrowly is the root cause of missing coverage. | Blaming the search or synthesis agent when the coordinator assigned only part of the topic. |
| Resume when context is still valid, restart when it is stale | A named session resumes an investigation whose files have not changed; a fresh session with an injected summary beats resuming over stale tool results; a fork compares approaches from one baseline. | Resuming a session after edits without telling it which files changed. |
The third rule is the one the guide illustrates first. When an agent skips customer verification in about one case in eight and refunds the wrong account, the answer is a programmatic prerequisite that blocks the downstream call, and the two options that make the prompt firmer are rejected because they rely on probabilistic compliance where the consequence is financial. That reasoning generalises: every time a stem says must, look for the option that enforces rather than requests.
Tool Design and MCP Integration: five rules for 18%
Five task statements and about eleven items. This domain is easy to mistake for a subset of the first, but it turns on contracts rather than orchestration: what a description promises, what an error tells the caller, who holds which tool, where a server is configured, and which built-in tool searches what. The guide’s sample item on two tools with one-line descriptions is the archetype.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| The description is the routing mechanism | Input formats, example queries, edge cases and when to use this tool rather than its neighbour are what the model selects on. Overlapping descriptions misroute, and a generic tool is split into purpose-specific ones. | Five to eight few-shot examples, or a keyword routing layer, offered as the first fix for two tools with one-line descriptions. |
| Errors carry a category and a retry flag | Transient, validation, permission and business failures need different responses, so the error says which it is, whether a retry can help, and what a person should be told. An empty result is a success, not an error. | A uniform operation-failed message that leaves the agent guessing whether to try again. |
| Scope the tool set, and control the call when it must happen | Four or five tools per role select reliably; eighteen do not. A worker gets only its role tools plus a narrow cross-role tool for the frequent case, and tool choice is forced when a specific call must come first or any call must happen. | Giving the synthesis agent the full web-search toolkit to save a round trip. |
| MCP scope follows audience, and secrets are expanded, not written | Shared team servers live in the project configuration with environment variables for tokens; personal or experimental servers live in the user configuration. Resources expose catalogues so the agent stops exploring blind. | A token committed in project config, or a custom server built for an integration a community server already covers. |
| Pick the built-in tool by what it searches | Content search is one tool and path search is another; targeted edits need unique anchor text and fall back to a full read and write when the anchor is not unique. Understanding is built incrementally from entry points, not by reading everything first. | Reading every file in the repository into context before tracing a single call path. |
Notice how the first rule is argued in the guide: expanding the descriptions is a low-effort, high-leverage first step that addresses the root cause, while few-shot examples add token overhead without fixing it, a routing layer over-engineers around the model, and consolidating the tools is a legitimate design that costs more than a first step warrants. Four distractors, four named reasons, and the word first in the stem doing most of the work.
Claude Code Configuration and Workflows: five rules for 20%
Six task statements and about twelve items. Three scenarios, code generation, developer productivity and continuous integration, name this domain as primary, and the items are the most concrete on the paper: which directory a command lives in, which flag runs a pipeline, which file scope reaches a teammate. Concrete does not mean trivial, because every option names a real-looking location and only one is right.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Scope instructions by who must receive them | Project-level files travel with the repository and reach every teammate; user-level files reach one person. Imports and a rules directory keep the project file modular, and a memory command shows what actually loaded. | A new team member missing standards that live in a colleague’s home directory. |
| Path-scoped rules for conventions that cross directories | A rule file with path globs loads only when a matching file is edited, which is the answer for test files scattered through a codebase. A subdirectory instruction file covers one subtree; a skill covers an on-demand procedure; the main file covers always-on standards. | Every convention consolidated in one root file with headers, relying on inference to pick the right section. |
| Plan mode for architectural change, direct execution for a scoped fix | Many files, several valid approaches or an infrastructure decision means plan first. A single-file fix with a clear stack trace does not. Verbose discovery goes to an exploration subagent so the main context survives the investigation. | Starting in direct execution and promising to switch to planning if it gets complicated, when the stem already said it is. |
| Iterate with examples and tests, not with adjectives | Two or three input and output pairs settle a transformation that prose describes inconsistently; a test suite written first turns iteration into sharing failures; the interview pattern surfaces considerations before code exists. Interacting fixes go in one message, independent ones in sequence. | Rewording the same instruction a fourth time because the output is still inconsistent. |
| A pipeline runs headless, reviews independently and remembers its findings | Non-interactive mode with a structured output format and a schema is how automation consumes Claude Code. Prior findings and existing tests go into context so a rerun reports only what is new, and the instance that reviews is not the instance that generated the change. | The same session that wrote the code asked to review it, or an interactive invocation that hangs in CI. |
Four of the guide’s twelve sample items come from this domain. Three are about scope and mode: a shared review command belongs in the project’s commands directory so it travels with the repository; a monolith-to-microservices restructuring enters plan mode before touching anything; and test conventions spread across a codebase are applied by path-scoped rule files rather than by one root file, a skill per file type or a subdirectory file in every folder. The fourth is a pipeline that hangs waiting for input, fixed by the non-interactive flag. Each rejected option is rejected for a stated reason, which is the habit to copy.
Structure beats instruction on every CCAR-F domain
A hook beats a mandatory sentence, a schema beats a request for valid JSON, a scoped tool set beats a warning to choose carefully, a facts block beats a longer summary, and an independent reviewer beats a self-review instruction. When two options survive your first read, the one that changes what is possible is almost always the key.
Prompt Engineering and Context Management: rules for the last 35%
Prompt Engineering and Structured Output is 20%, about twelve items over six task statements, and it is where the extraction and continuous integration scenarios live: explicit criteria, few-shot examples, schemas through tool use, retry loops, batch strategy and multi-pass review. The connecting question is whether a second attempt knows anything the first did not.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Criteria, not confidence | Define what to report and what to skip by category, with an example per severity level. An instruction to be conservative or to report only high-confidence findings does not move precision; a category with a high false-positive rate is switched off while it is fixed. | Asking the model to be more careful in place of stating which comment patterns count as a finding. |
| Two to four examples of the ambiguous case | Few-shot examples that show the reasoning for choosing one action over a plausible alternative teach judgment the model can generalise. They also fix the shape of the output and reduce empty extractions from unusual document layouts. | More instructions added when the instructions were never the problem. |
| A schema through tool use removes syntax errors, not semantic ones | Forcing the tool call guarantees the shape. Values that do not add up, or land in the wrong field, still need a check. Fields the source may lack are optional so the model is not pushed to invent them, and an enum carries an other and an unclear value. | A required field for a value that some documents never contain. |
| Retry with the error attached, and only when a retry can learn something | A follow-up that includes the document, the failed output and the specific validation failure fixes format and structure errors. It cannot fix information that is absent from the source; that case is routed elsewhere. Paired fields such as a stated and a calculated total flag disagreements. | A second attempt at extracting a figure that exists only in a document the model was never given. |
| Batch when nobody is blocked, and split reviews that dilute attention | The batch API halves cost inside a window of up to a day with no latency guarantee and no mid-request tool calls, so overnight reports use it and pre-merge checks do not; failures are resubmitted by correlation id. A fourteen-file review runs per-file passes plus one integration pass. | Switching a blocking check to batch for the saving, or buying a larger context window to fix inconsistent review depth. |
Context Management and Reliability is the lightest domain at 15%, about nine items, and the one named as primary by the most scenarios. Every failure in it is silent: a summary that dropped an amount, an error reported as an empty success, a claim whose source was lost, an aggregate accuracy that hides a failing segment. Nothing raises an exception, which is why the rules here are about what must be preserved.
| Rule | Why it holds | The trap it catches |
|---|---|---|
| Exact facts leave the narrative | Amounts, dates, order numbers and statuses go into a persistent facts block outside the summarised history; tool output is trimmed to the fields that matter before it accumulates; key findings lead the aggregate, with headers for the middle. | A progressive summary that keeps the topic and loses the refund amount. |
| Escalate on triggers, not on sentiment or self-reported confidence | When the customer asks outright for a person, hand over at once; a policy gap or silence escalates; no progress escalates. Frustration alone does not, and a confidence score the model assigns itself is not calibrated. Several matching customers means asking for another identifier. | A confidence threshold or a sentiment detector routing cases to people. |
| A failure travels with its context | Failure type, what was attempted, partial results and possible alternatives let the coordinator decide. An access failure and a valid empty result are reported differently. Workers recover from transient faults locally and propagate only what they cannot resolve; the final report annotates coverage gaps. | A timeout swallowed as an empty success, or a whole workflow terminated on one subagent’s failure. |
| Explore through delegation and scratchpads | Specific questions go to subagents so the main context holds the coordination; findings are written to a scratchpad and read back; a phase is summarised before the next begins; agents export state to a manifest for crash recovery; compaction reclaims a filling session. | A long session that starts describing typical patterns instead of the classes it found an hour ago. |
| Provenance survives synthesis, and accuracy is checked by segment | Claim-to-source mappings, dates and methodology travel through every summarisation step; conflicting credible figures are both kept and annotated. A high aggregate accuracy is checked by document type and field, with stratified sampling of high-confidence outputs, before review is reduced. | A ninety-seven percent overall figure used to automate a segment that fails one time in four. |
The guide’s escalation item is the clearest test of the second rule. An agent resolving 55% of contacts against an 80% target, escalating the simple cases and attempting the hard ones, needs explicit criteria with examples of each side of the line. Self-reported confidence fails because the agent is already confidently wrong on hard cases, a trained classifier is infrastructure before prompting has been tried, and sentiment does not track complexity at all.
CCAR-F exam tips: the traps table and the terms
The rules are organised by domain; the traps are organised by shape, because the same distractor pattern recurs across domains and is easier to spot once it has a name. The guide’s own sample rationales supply most of these, since every rejected option there is rejected for a reason that repeats.
| Trap | What it looks like | The check |
|---|---|---|
| The firmer instruction | A stronger system-prompt sentence where the stem describes a rule that must hold every time. | Ask what happens on the run where the model ignores the sentence. If nothing stops it, it is not a control. |
| The downstream blame | A fix aimed at the search or synthesis agent when the coordinator’s logs show the decomposition was too narrow. | Find which agent decided the scope. The others executed it correctly. |
| The over-provisioned worker | Every tool handed to one agent so it never needs a round trip. | Least privilege: role tools, plus one scoped tool for the frequent case. |
| The bigger window | A higher tier or larger context proposed for inconsistent review depth or a drifting session. | Attention dilutes; split the passes or delegate the discovery instead. |
| The generic error | Retries exhausted, then a one-word failure status returned upward. | The coordinator can only recover from what it is told. Category, attempt, partial results. |
| The batch everywhere | Both a blocking check and an overnight report moved to batch for the saving. | Who is waiting? Batch has no latency guarantee. |
| The retry that cannot learn | A second extraction attempt for information the source never contained. | Retry fixes format and structure. Absence is routed to a person or a null. |
| The select-count slip | Three options marked on a select-two item, or one dropped on review. | Every multiple-response stem states the count. Match it, then check it again. |
The quick reference below is the vocabulary those rules depend on: working definitions in plain terms, each there because an item can turn on it. Configuration file names and flags are described by what they do rather than spelled out, because the items test the choice, not the syntax.
| Term | Working definition |
|---|---|
| stop_reason | The field that says why the model stopped: it wants a tool run, or it has finished its turn. The only correct loop control. |
| Coordinator and subagent | A hub agent that decomposes, delegates, aggregates and routes every message; workers in isolated context that inherit nothing. |
| Task tool and allowedTools | The mechanism that spawns a subagent, and the list that must include it before a coordinator can delegate at all. |
| Hook | Code that intercepts a tool call before it runs or a result before the model sees it. The deterministic alternative to an instruction. |
| Fork and resume | Branching one analysis into independent explorations, and continuing a named session whose context is still valid. |
| isError and isRetryable | Structured error metadata on a tool result: that it failed, what kind of failure it was, and whether trying again can help. |
| tool_choice | Auto lets the model answer in text; any forces some tool call; a named tool forces that specific call first. |
| Project and user MCP scope | Shared servers configured with the repository and environment-variable tokens; personal servers configured for one user. |
| Path-scoped rule | A rules file whose frontmatter lists path globs, loaded only when a matching file is being edited. |
| Forked skill context | A skill that runs in its own sub-context so verbose or exploratory output never reaches the main conversation. |
| Plan mode | Investigate and design before changing anything. For architectural, multi-file or many-valid-approaches work. |
| Headless output with a schema | Non-interactive Claude Code emitting machine-readable results against a schema, so a pipeline can post them. |
| Message Batches API | Asynchronous processing at half cost inside a window of up to a day, no latency guarantee, correlated by custom id. |
| Lost in the middle | The tendency to process the start and end of a long input reliably and drop findings from the middle. |
| Stratified sampling | Measuring error rates inside high-confidence output by segment, so an aggregate cannot hide a failing document type. |
The final-week checklist
A cheat sheet belongs to the last seven days, after the Claude certification study guide routine has been followed and before the sitting. The sequence below turns the rules into a diagnostic on your own misses and folds in the administrative checks that catch people at the door.
Seven days out
Read the 30 task statements in the official guide once more and write each as a one-line rule in your own words. Compare your list with the 25 rules here and mark every gap.
Six days out
Sit the free 20-item CCAR-F mock, timed. It is written to the five-domain guide. Score it as a raw count only, then map every miss to a rule; a miss that maps to nothing means a rule is missing.
Five and four days out
Repair the largest domain first. Agentic Architecture is seven objectives and about sixteen items, so a gap in any one objective costs about two marks. Then the two twenty-percent domains.
Three days out
Reread the six scenario descriptions in the guide and the primary domains listed under each. Then the administrative checks: registration name matches your ID exactly; your work email is on a recognised Claude Partner Network company domain.
Two days out
Confirm the appointment. The exam guide allows changes up to 24 hours before; pearsonvue.com and the partner FAQ say 48. Treat 48 as the deadline and neither rule can catch you.
The day before
Nothing new. Workspace cleared, ID ready, and a pacing plan: two minutes an item leaves nothing for review, so decide now to flag and move on rather than linger.
Exam day
Read the scenario once, then each stem for its controlling sentence: what must hold, who is waiting, what cannot be undone. Count on every multiple-response item before you move on.
The policy facts behind those checks: CCAR-F moved to Pearson VUE on 30 June 2026 and its fee rose from $99 to $125 per attempt. Select, Preferred and Global Premier partners pay 50 percent, and Global Premier partners pay nothing through 31 December 2026. Registration needs a work email on a recognised Claude Partner Network company domain. A failed attempt can be retaken after 14 days, then 30, then 90, with four attempts per exam in any rolling twelve months, each at the full fee with your tier discount applied. The credential lasts twelve months and renews on time through a free, non-proctored assessment. The rules are identical for CCDV-F, whose cheat sheet follows the same layout across eight domains instead of five. Checked against the official sources on 17 September 2026: the CCAR-F exam guide v1.0, the Anthropic Certification Exam Policy dated 25 June 2026 and the partner certification FAQ.
On practice: the free mock at /practice/ccar-f is 20 items, timed, with no account, written to the five-domain guide and with a stated reason for every option. It reports a raw count only. Cred Farmer’s older CCAR-F timed forms were built on a retired seven-domain outline; they are supplemental study context, not a picture of this exam, and no result on them is a reason to book. No practice percentage anywhere converts to the scaled score. Our guide to Claude certification practice exams covers what else exists and how to read it.
Key takeaways
- Flattest blueprint, so no domain is skippable. 27% down to 15% is a 1.8 to 1 spread; the lightest domain is still about nine items.
- Structure beats instruction. Hooks, gates, schemas, scoped tool sets and facts blocks beat firmer sentences on every domain.
- Find who decided the scope. When coverage is missing, the coordinator’s decomposition is the root cause, not the workers.
- Ask what a retry can learn. Format and structure errors retry well; information absent from the source never does.
- Batch only when nobody is blocked. Half the cost, up to a day, no latency guarantee.
- Practice results are raw counts. The free mock is written to the current guide; the older timed forms are not, and neither converts to 720.
Test the 25 rules on a free CCAR-F mock written to the five-domain guide
Twenty items, timed, no account, with a reason for every option after you submit. The result is a raw count and a per-domain breakdown, which is what the final-week checklist asks for; it is not a scaled score, a pass prediction or a booking signal. From there, the study sections are the next step.
Start the free CCAR-F mockQuestions
Frequently asked
The follow-up questions people search next.
Is there an official CCAR-F cheat sheet?
No. Anthropic publishes the exam guide, which names five domains, six scenarios and 30 task statements, and nothing shorter. This page compresses the task statements into 25 decision rules with the trap each one catches. It is derived from the guide and not endorsed by it, and it lists no answers.
What are the CCAR-F domain weights?
Agentic Architecture and Orchestration 27%, Claude Code Configuration and Workflows 20%, Prompt Engineering and Structured Output 20%, Tool Design and MCP Integration 18% and Context Management and Reliability 15%. On 60 items that is about 16, 12, 12, 11 and 9. It is the flattest blueprint of the four.
How is the CCAR-F exam structured?
Sixty items in 120 minutes, framed by four scenarios drawn at random from a published bank of six. Items are multiple-choice and multiple-response, and every multiple-response stem states how many options to select. The pass mark is a scaled 720 on a 100 to 1,000 scale; the fee is $125.
What changed for CCAR-F on 30 June 2026?
Delivery moved to Pearson VUE and the fee rose from $99 to $125 per attempt. Partner-tier discounts still apply: Select, Preferred and Global Premier partners pay half, and Global Premier partners pay nothing through 31 December 2026. Confirm the current figure at checkout before booking.
Is a CCAR-F quick reference enough on its own?
No. It compresses judgment you already have, so it works as a diagnostic on practice misses in the final week rather than as first reading. The five domains are close in weight, which means the useful check is coverage: can you state at least three rules for every domain without looking?
Do the older Cred Farmer CCAR-F timed forms match this exam?
No. They were built on a retired seven-domain outline and are supplemental study context only; they are not a picture of the five-domain exam and no result on them is a reason to book. The free 20-item mock is written to the current guide, and even its result is a raw count, not a scaled score.
Keep reading
Related posts
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.