Claude Certified Associate – Foundations

Exam code CCAO-F. This single page holds two things: a fast high-level overview to orient you, and a thorough study guide covering every exam domain, worked examples, decision tables, and study materials. This exam is written for business and productivity users — no coding, no API calls, no SDKs. If you have never written a line of code, you are exactly who this certification is for. Overview items link straight into their deep-dive sections.

60 questions 120 minutes 720/1000 to pass $99 USD Valid 12 months
⚡ Open the 1-page Cram Jam sheet

Part 1 — High-Level Overview

What & Why

The Claude Certified Associate – Foundations credential validates that you can apply Claude to real business and productivity tasks with minimal guidance: using built-in platform features to streamline work, spotting where Claude could improve a process, weighing quality against speed and cost, and knowing when a task is too technical for you and belongs with a Claude Architect or Claude Developer instead. It is Anthropic's own foundational, non-engineering certification, sitting at the base of the Claude Certification Program alongside the Developer and Architect tracks.

You are not expected to write code, call an API, or configure an SDK. You are expected to write a clear prompt, judge whether the answer is trustworthy, pick the right Claude surface and model for a job, set up a shared Project sensibly, and recognize a data-handling or governance problem before it happens.

Who it's for Anthropic aims this exam at operations, marketing, project management, education, communications, and general knowledge-work roles that use Claude daily and build Claude Projects, not at software developers building against the API. Full detail in §1 Exam Details.

Exam at a Glance

AttributeValue
Exam codeCCAO-F
Questions60, standalone items (no multi-question scenarios)
Question typesMultiple choice & multiple response — each item states how many responses to select
Time120 minutes
Passing score720 out of 100–1000 (scaled, criterion-referenced)
Cost99 USD
DeliveryPearson VUE — online proctored or a Pearson test center, closed book
Validity12 months, free on-time renewal assessment

→ Deep dive: §1 Exam Details

The 7 Domains

Anthropic's own Exam Guide PDF publishes seven weighted content domains. Output Evaluation is the single biggest chunk — this is fundamentally a "can you judge Claude's work" exam, not a "can you write a clever prompt" exam.

14%
1 — Prompting & Task Execution
21%
2 — Output Evaluation & Validation
12%
3 — Product & Model Selection
16%
4 — Workflow Integration & Solution Design
12%
5 — Configuration & Knowledge Management
15%
6 — Governance, Risk & Responsible Use
10%
7 — Troubleshooting & Optimization
DomainWhat it testsDeep dive
1 · Prompting 14%Writing effective prompts, decomposing complex requests, iterating, adapting strategy to task type.→ §3
2 · Output Evaluation 21%Judging accuracy, catching hallucinations and bias, fact-checking, knowing when a human must review, choosing an output format.→ §4
3 · Product & Model Selection 12%Picking the right Claude feature and model for the job, managing context and memory.→ §5
4 · Workflow Integration 16%Using Claude for requirements analysis, research, process redesign, and explaining Claude's value and limits to others.→ §6
5 · Configuration 12%Setting up Projects with instructions and knowledge sources, connectors, keeping configuration current.→ §7
6 · Governance & Responsible Use 15%Appropriate use cases, data sensitivity and privacy, organizational policy, ethics.→ §8
7 · Troubleshooting 10%Diagnosing a weak prompt or poor output, adjusting approach, optimizing a workflow.→ §10

Core Skill Snapshot

If you master this shortlist, you cover the majority of questions. Each links to its detailed treatment.

Prompting

Clear instructions, examples, role framing, step-by-step structure, iteration. → details

Judging output

Hallucination spotting, fact-checking, knowing when to escalate to a human. → details

Claude surfaces

Chat, Projects, Artifacts, research mode, and matching Haiku/Sonnet/Opus to the task. → details

Configuration

Project instructions, uploaded knowledge, connectors, keeping it current. → details

Responsible use

What never to paste in, organizational policy, bias awareness. → details

Team adoption

Shared Projects, consistent prompts, admin basics. → details

→ Full "which one" decision tables (§11)

Recommended Materials

This is a brand-new certification (the exam guide is dated July 2026), so the practice-question market is thin and full of low-quality guesswork. The only material worth trusting fully is Anthropic's own.

The honest stack: read the official Exam Guide PDF in full → work through Anthropic's own help center and documentation for Projects, Artifacts, Memory, and Skills → use Claude on real work daily for two weeks and keep a short prompt journal. Treat any third-party "exam dump" or "guaranteed pass" practice set for this specific exam with real skepticism until it has a track record.

→ Full breakdown with sources (§12)

Study Plan (Quick Version)

Built on learning science — active recall + spaced repetition — but the real engine here is daily hands-on use. You cannot cram judgment about AI output quality; you build it by using Claude on real tasks and checking your own work.

  1. Days 1–3: Read the exam guide's blueprint, skim Anthropic's help center for Projects/Artifacts/Memory/Skills, and start a daily prompt journal.
  2. Days 4–9: Use Claude on real work every day. After each session, write down one prompt that worked and one that didn't, and redo the weak one.
  3. Days 10–12: Study Domains 2 and 6 hard (Output Evaluation and Governance are the two most conceptually dense, together over a third of the exam). Teach the shared-responsibility-style "what Claude can and can't do" idea out loud.
  4. Days 13–14: Work every worked example in this guide, review the Cram Jam, sleep well — don't cram new material.

→ Detailed retention-based plan (§13)  ·  ⚡ Cram Jam one-pager  ·  → Exam-day tips & traps (§14)

Part 2 — Thorough Study Guide

1 · Exam Details

Confirmed vs. reported Anthropic's public certification page confirms the $99 price and that candidates must review the Exam Guide, Certification Terms and Conditions, and Certification Exam Policy before the exam session. This guide was written after directly retrieving Anthropic's official CCAO-F Exam Guide PDF (version 1.0, effective July 2026) from the Anthropic Partner Academy, so the facts below — question count, timing, domain weights, scoring, delivery, and policies — come from that primary document, not third-party guesswork. Even so, the guide itself states it "is subject to change without notice," so re-check the current PDF on anthropic-partners.skilljar.com before you book.

The certification validates that a candidate can apply Claude to complete business and productivity tasks with minimal guidance: using built-in features and tools to streamline workflows, identifying opportunities to improve a process with Claude, selecting an approach that balances quality, efficiency, and cost, and recognizing limitations that should be escalated to a Claude Architect or Claude Developer.

Who it's for

Professionals who use Claude as a productivity tool and build Claude Projects, across operations, marketing, project management, education, communications, and general knowledge work — both internal staff and external consultants. The guide is explicit that this is not for software developers who build against APIs or design agentic systems, and candidates are not expected to design enterprise-scale AI architectures; that scope belongs to the Architect and Developer credentials. There are no mandatory prerequisites, and no software-development or API experience is needed — the credential is awarded on exam performance alone.

Scoring mechanics

  • 60 items, standalone. Each is a self-contained multiple-choice or multiple-response question — there are no shared-scenario item sets. Multiple-response items tell you how many answers to select.
  • Scaled score 100–1000, pass at 720. It is criterion-referenced: you are measured against a fixed performance standard set by a formal standard-setting study, not ranked against other candidates.
  • Your score report shows pass/fail plus percent-correct by domain — useful for retake prep, but domain scores do not themselves determine pass/fail; only the total scaled score does.

Logistics

ItemDetail
Price99 USD (partner-tier discounts may apply at checkout).
DeliveryPearson VUE, online proctored or at a Pearson VUE test center. Closed book: no Claude, no documentation, no browser tabs, no external tools, no notes, phones, or secondary monitors.
RegistrationThrough the Anthropic Partner Academy; you create a Pearson VUE account to schedule. You can cancel or reschedule up to 24 hours ahead without penalty.
RetakesWaiting periods of 14 / 30 / 90 days after the 1st / 2nd / 3rd failed attempt; up to 4 attempts per exam within a rolling 12 months. Full fee applies each attempt.
Validity & renewal12 months from the award date. Renew on time with a free, non-proctored assessment; a lapsed credential requires retaking the full paid exam.
Before you sitYou must accept a confidentiality/non-disclosure agreement before the exam starts, and you must review the Exam Guide, Certification Terms and Conditions, and Certification Exam Policy beforehand.
Closed-book, no exceptions This is a proctored exam like any other professional certification. You cannot have Claude open, cannot browse documentation, and cannot use a second monitor or phone. Prohibited items include phones, smart watches, headphones, and any recording device. Treat it exactly like an in-person exam even if you sit it from home.

↑ back to top

2 · What Claude Actually Is

A surprising number of exam-worthy mistakes come from a fuzzy mental model of what a large language model actually does. Get this part solid first — it underlies almost every other domain.

2.1 · What an LLM Is (and Isn't)

Claude is a large language model (LLM): a system trained on enormous amounts of text to predict, one piece at a time, what text should come next given everything before it. From that simple mechanism, useful behavior emerges — Claude can summarize, draft, explain, translate, brainstorm, and reason through multi-step problems in plain language.

  • Claude is not a search engine. Left to its training data alone, it does not "look things up" — it generates plausible, fluent text. This is why it can state a false fact with complete confidence: fluency and accuracy are not the same thing.
  • Claude is not a database. It does not query a system of record for exact figures, and it does not know today's date, your company's internal numbers, or anything after its training cutoff unless you provide it or it uses a connected tool like web search.
  • Claude does not "remember" by default the way a person does — see §5.3 for what Memory and context actually do.
  • Claude is a tool you direct, not an autonomous coworker with judgment equal to a human's. It has no stake in the outcome, no accountability, and no ability to independently verify a claim against reality unless you give it a way to check.
The one-sentence version Claude is extremely good at manipulating and generating language and reasoning through structured problems, and it is not a reliable, verified source of fact on its own — that gap is exactly what Domain 2 (Output Evaluation) is built around.

2.2 · Tokens & the Context Window

Claude reads and writes in tokens, roughly chunks of a word (a short word is often one token; a long or unusual word may be several). You don't need to count tokens by hand, but two consequences matter for daily use:

  • The context window is the total amount of text — your prompt, any uploaded files, the conversation history, and Claude's reply — that fits into one exchange. It is large, but it is finite. A very long conversation or a huge pasted document can fill it.
  • When context fills up or a conversation gets too long and wandering, quality degrades: Claude may lose track of an instruction from early in the chat, or conflate details. The fix is not "add more instructions" — it's to start a fresh conversation, summarize what matters so far, or move stable, reusable context into a Project (see §7.1) instead of re-pasting it every time.
  • Uploaded files count against the same context budget. A very large file or a large number of files may need to be trimmed, summarized, or split into a follow-up question rather than one giant ask.
Practical habit If a long-running chat starts giving inconsistent or repetitive answers, that's a signal to summarize progress, start a new conversation with that summary, and drop stale context — not to keep piling on more instructions.

2.3 · The Model Lineup

Anthropic offers Claude in a small family of models, generally described along a speed/cost vs. depth-of-reasoning spectrum. Exact model names, versions, and specific limits change over time — check Anthropic's current documentation for what's live today — but the three-tier shape is stable and is explicitly named in the exam blueprint:

TierBest for
Haiku (fastest, cheapest)High-volume, low-complexity tasks: quick replies, simple classification or extraction, short drafts where speed matters more than depth.
Sonnet (balanced)The everyday default for most business writing, analysis, and drafting — a strong balance of quality, speed, and cost for typical knowledge work.
Opus (most capable)Complex reasoning, high-stakes analysis, long or nuanced documents, and tasks where getting it right matters more than getting it fast or cheap.
Common trap "Always use the most capable model for everything" is a wrong answer pattern on this exam. Matching model to task — cheap and fast for simple, high-volume work; capable for complex reasoning — is a named exam objective, not just a nice-to-have.

↑ back to top

3 · Prompting & Task Execution Domain 1 · 14%

Objectives: create effective prompts for business and technical tasks, apply task decomposition, iterate to improve output, and adapt strategy to task type.

3.1 · The Seven Prompting Moves

These are drawn from Anthropic's own published prompt-engineering guidance, translated for a non-technical audience. None of them require special syntax — they are ways of thinking about what you ask for.

  1. Be clear and direct. State the task, the audience, the format, and the length up front, the way you'd brief a new hire who can't read your mind. Vague asks get vague answers.
  2. Give it a role. "You are an experienced grant writer reviewing this proposal for clarity" focuses tone and priorities far better than no framing at all.
  3. Show examples. One or two examples of the input/output pattern you want ("here's a sample of the tone we use") teach a format faster than a paragraph of description.
  4. Let it think in steps. For anything with real reasoning (comparing options, working through a policy question), ask Claude to reason through the problem before giving a final answer — this improves quality on non-trivial tasks.
  5. Structure a long or complex prompt. When a request has several distinct parts (background, constraints, the actual ask), separate them clearly — headings or numbered sections — so nothing gets missed or blended together.
  6. Give it the material, not just the question. Paste in the actual document, data, or draft rather than describing it from memory; Claude works from what's in front of it, not what you assume it already knows.
  7. Tell it what "done" looks like. Specify the acceptance bar: "flag anything you're not confident about," "keep it under 200 words," "match this style guide" — this is what turns a first draft into a usable one.
MoveWhen it helps mostOne-line example
Give contextClaude doesn't know your team, product, or situation"We're a 12-person nonprofit; this goes to first-time donors."
Be specificThe task has more than one reasonable interpretation"Trim to 150 words, keep the second paragraph's numbers."
Give examplesYou want a particular tone or format matched exactly"Match this style: [paste one sample sentence]."
Assign a roleYou want a specific lens or priority applied"You're reviewing this as a skeptical procurement officer."
Ask for a formatThe output feeds a next step (comparison, import, forwarding)"Return this as a table with columns Name, Date, Amount."
Set constraintsThere's a real limit — length, audience, what to avoid"Under 200 words, no jargon, don't mention pricing."
DecomposeThe ask has several distinct deliverables"First just outline it. I'll approve before you draft it in full."
IterateThe first draft is close but not right"Keep the structure, but make the tone warmer."

3.2 · Task Decomposition

A single sprawling request ("reorganize our onboarding process and also draft the new email templates and also summarize stakeholder feedback") reliably produces a worse result than the same work broken into steps. Task decomposition means splitting a complex ask into smaller, ordered pieces, often across separate turns in the same conversation:

Before

"Fix our customer onboarding process." (Too broad — Claude has to guess what "fix" means and for whom.)

After

Turn 1: "List the current onboarding steps from this document and flag where users seem to drop off." Turn 2: "For the two worst drop-off points, propose two alternatives each." Turn 3: "Draft the email copy for the option we picked."

Decomposition also makes output easier to evaluate — you can check each small step before building on it, instead of discovering a flawed assumption buried inside a huge final answer.

3.3 · Iterating on a Prompt

The first draft of a prompt is rarely the best one. Treat it like editing: read the actual output, name specifically what's wrong ("too formal," "missed the Q3 numbers," "assumed the wrong audience"), and revise the prompt rather than manually rewriting the output every time.

Before

"Summarize this report." → gets a generic, oddly-focused summary.

After

"Summarize this report in 5 bullet points for a VP audience who cares about budget impact and timeline risk, not methodology." → gets a summary that's actually usable.
Keep a prompt journal Save prompts that worked well and the ones that needed three tries to get right, with a note on what changed. This is the single highest-leverage exam-prep habit for Domains 1, 2, and 7 — you're building the exact judgment the exam tests.

3.4 · Matching Strategy to Task Type

Task typeWhat changes about the prompt
AnalysisGive the source material directly, state the specific question, ask for reasoning steps before the conclusion.
ResearchBe explicit about scope and recency needs; if current information matters, use a Claude surface with web search rather than relying on training knowledge alone (see §5.1).
DraftingProvide examples of tone/format, specify audience and length, ask for a first pass rather than a "final" version.
BrainstormingAsk for a wide, unfiltered set of options first, then a separate follow-up to narrow or rank them — mixing generation and judgment in one prompt narrows the ideas too early.

A more granular breakdown, by the nine task types the exam blueprint touches on most:

Task typeWhat to give ClaudeWhat to ask forHow to check it
SummarizeThe full source documentLength, audience, and what to prioritizeSpot-check 2-3 claims against the source
ExtractThe source and the exact fields you needA table or list with consistent fieldsConfirm every row against the source, not just a sample
DraftGoal, audience, tone or a sampleA first pass, not a final versionRead for factual claims and tone fit, then iterate
RewriteThe original text and what should changeWhat to preserve (facts, structure) vs. what to change (tone, length)Diff the meaning against the original, not just the wording
AnalyzeThe data or documents and the specific questionReasoning steps before the conclusionSanity-check the reasoning, not just the final number
CompareBoth items and the criteria that matter to youA structured table, not a paragraphVerify each criterion independently rather than trusting the summary row
BrainstormThe constraint and the goal, minimal filteringA wide list first, ranking as a separate stepCheck that ideas are actually distinct, not restatements
TranslateThe text and the target audience/registerA draft translation, flagged for idioms or ambiguityHave a fluent human reviewer check anything customer-facing or legal
ClassifyThe item and the exact categories allowedOne category per item, with a stated reasonReview borderline cases by hand; don't assume 100% consistency

3.5 · Common Prompt Failures

Most weak prompts fail in one of a handful of predictable ways. Recognizing the pattern in the output tells you which part of the prompt to fix.

FailureSymptom in the outputThe fix
Vague askGeneric, unfocused, technically responsive but not usefulAdd audience, purpose, and what "good" looks like
Missing contextReasonable-sounding but wrong for your actual situationPaste the real document, data, or background instead of describing it
Too many asks at onceShallow on every part, misses at least one entirelyDecompose into ordered steps (§3.2)
Wrong format requestedCorrect content, unusable shape (prose when you needed a table)Name the exact structure you need up front
Leading questionOutput agrees with the framing rather than giving an independent readAsk neutrally first ("what are the risks here?"), not "don't you think this is a bad idea?"
Leading questions Framing a question to imply the answer you want ("this vendor seems overpriced, right?") tends to get you agreement, not analysis. For anything you actually need judged fairly, ask neutrally and let Claude reason to its own conclusion before you share your own view.

↑ back to top

4 · Output Evaluation & Validation Domain 2 · 21%

The largest domain on the exam. Objectives: evaluate outputs for accuracy and completeness, identify hallucinations/inconsistencies/bias, apply fact-checking, determine when human review is required, and edit/organize output into the right format.

4.1 · Spotting Hallucinations

A hallucination is fluent, confident, wrong output — a fabricated citation, a plausible-sounding statistic that doesn't exist, a policy detail invented to fill a gap. Confidence in Claude's tone is not evidence of accuracy; a hallucinated fact and a correct one can read identically.

  • Highest-risk spots: specific numbers, dates, named sources, legal or regulatory citations, quotes attributed to a real person or document, and anything about very recent events.
  • Checklist: does the claim cite something checkable? Can you find that source independently? Does the number match a document you already have? Would this claim be embarrassing or costly if wrong — and if so, is it verified yet?
  • Asking Claude to self-rate its confidence and trusting a high score is not a validation technique — self-reported confidence is not a reliable accuracy signal. Verify against an authoritative source instead.
Exam pattern A scenario gives you a confident-sounding output with a specific citation or figure and asks what to do next. The correct answer is almost always "verify it against the original source before using it," never "send it because it sounded confident" or "ask Claude to rate its own confidence."

Hallucinations aren't all the same shape. Recognizing which kind you're looking at speeds up the check:

TypeWhat it looks likeWhy it happensHow to catch it
Invented factsA specific detail — a policy clause, a product feature — that sounds plausible but doesn't existThe model fills a gap in its knowledge with the most statistically likely-sounding textAsk "where does this come from" and check the source directly
Invented citationsA named source, article, or case that doesn't exist or doesn't say what's claimedCitations are a common, high-confidence surface for fabrication when the real source wasn't providedLook up the citation independently before using it
Wrong numbersA statistic, total, or date that's close to right but off, or entirely fabricatedLanguage models predict plausible text, not arithmetic or lookups, unless given the actual dataRecompute or cross-check against the source document
Confident vaguenessAn answer that sounds complete but is actually generic filler dressed in specific-sounding languageFluent phrasing can mask a lack of real substanceAsk "what specifically supports this" and see if a real answer follows
Outdated informationA correct-as-of-training-cutoff fact presented as currentClaude's training knowledge has a cutoff and no built-in sense of "today" unless toldFor anything time-sensitive, use a surface with web search or supply current data (§5.1)
Misread source documentA summary or extraction that quietly contradicts or misstates the document you providedLong or dense documents increase the chance a detail gets blended, dropped, or misattributedSpot-check the summary's specific claims against the original, not just its overall shape
Misread-source risk Even when Claude is given the correct document, it can still misstate what's in it — this is different from inventing a fact from nothing, and it's easy to miss because the summary "sounds" like it came from your document. Treat a summary of your own material with the same spot-checking discipline as a claim from memory.

4.2 · Fact-Checking & When to Escalate to a Human

Not every output needs the same level of scrutiny — match the check to the stakes:

StakesAppropriate check
Low (internal brainstorm, first draft)Skim for obvious errors; refine as you go.
Medium (external-facing draft, internal report)Check factual claims against your own source documents; have a colleague skim before sending.
High (legal, financial, medical, regulatory, anything going to a customer or regulator)Verify every specific claim against an authoritative source, and route through the human expert or process your organization already requires for that category — Claude output does not change that requirement.

Framed the other way — by the actual document you're producing, and who has to stand behind it:

StakesVerification requiredWho signs off
Throwaway draft (personal notes, a first pass you'll rework)None beyond a quick sanity readYou, informally
Internal document (team memo, internal report)Check factual claims that matter to the decision being madeYou, or a colleague on a quick skim
Customer-facing (email, proposal, public post)Verify every specific claim, name, and figure against a sourceYou plus a reviewer before it goes out
Regulated or legal (compliance language, contracts, medical or financial advice)Full verification against authoritative sources, no exceptionsThe qualified human expert or process your organization already requires — unchanged by Claude being involved
Effort should scale with stakes, not with how impressive the draft looks A polished, well-formatted draft is not lower-risk than a rough one — formatting quality and factual accuracy are unrelated. Judge the required verification effort by what the output is for, not by how finished it looks.

Escalating to a human (or to a Claude Architect/Developer for a technical build) is a core, tested skill, not an admission of failure. Signals it's time to escalate: the task requires professional judgment you're not qualified to give (legal, medical, financial advice), the stakes of being wrong are high and unverifiable, or the task has moved from "using Claude" to "building a technical integration."

4.3 · Editing, Refining & Choosing an Output Format

Part of evaluation is deciding how the output should be packaged, not just whether its content is right:

  • Inline chat text — fine for a quick answer, a short explanation, or something you'll copy a sentence or two from.
  • An Artifact — better for anything you'll keep editing, share as a standalone document, or that benefits from a dedicated view (a long document, a table, a simple mockup). See §5.1.
  • Structured data (a table, a list with consistent fields) — better than prose when the next step is comparison, sorting, or feeding it into another tool.

Editing and adapting output for its actual audience — trimming jargon for an executive summary, adding detail for a technical reviewer — is itself a tested skill, not a "nice to have" after the fact.

4.4 · Output Evaluation Rubric

This is the highest-weighted domain on the exam — a working checklist matters more here than anywhere else in the guide. Run any consequential output past these six dimensions before it goes anywhere:

DimensionWhat to checkHow to check it quicklyWhat a failure looks like
AccuracyAre the facts, numbers, and claims actually trueSpot-check specific claims against a source you trustA confident, specific-sounding statement that turns out to be fabricated or wrong
CompletenessDoes it answer the whole question, not just part of itReread the original ask and check each part was addressedA polished answer to only one of three things you asked
RelevanceIs it actually useful for what you asked, not just technically responsiveAsk "would this help the actual audience make a decision"Generic content that could apply to almost any similar request
ToneDoes it match the audience and occasionRead it as the recipient wouldOverly casual for an executive, or stiff and jargon-heavy for a general reader
FormatIs it structured the way the next step needsCheck if you can use it as-is or have to reformat it firstA wall of prose when the next step needed a table or list
SourcingCan every specific claim be traced back to something checkableAsk "where did this come from" for any number, quote, or citationA statistic or quote with no traceable origin
Exam pattern A question describes output that is well-written, well-organized, and confidently delivered, then asks whether it's ready to send. Good writing quality is not one of the six dimensions above — the correct answer checks accuracy and sourcing regardless of how polished the prose reads.

↑ back to top

5 · Product & Model Selection Domain 3 · 12%

Objectives: select the right Claude product feature, differentiate model types, align model choice with task requirements, and manage context/memory considerations.

5.1 · Chat, Projects, Artifacts, Research

SurfaceWhat it's for
A one-off chatA single question or task with no need to reuse context later — quick, disposable, no setup.
A ProjectRecurring work that shares context: instructions, reference documents, and a consistent voice across many conversations. See §7.1.
An ArtifactA standalone piece of output you'll keep, edit, or share separately from the chat — a document, a table, a simple interactive view — rather than scrolling back through conversation text to find it.
Research / web searchTasks that need current information beyond Claude's training knowledge — recent events, current prices, today's status of something. Turn it off when currency doesn't matter and speed or a controlled, closed set of sources does.

The full surface map, including the pieces covered in more depth elsewhere in this guide:

SurfaceWhat it isWhen to reach for itThe mistake people make with it
ChatA single conversation with no persistent shared setupA quick, one-off question or taskRe-explaining the same standing context every single time instead of moving it into a Project
ProjectsA container of shared instructions, knowledge, and chatsRecurring work a team or individual repeats oftenTreating it as "set once, forget forever" instead of reviewing it (§7.3)
ArtifactsA standalone, editable view of a piece of outputAnything you'll keep, edit, export, or share on its ownLeaving finished work buried in chat scroll instead of promoting it to an Artifact
Research / web searchA mode that lets Claude pull current information from outside its training dataAnything time-sensitive: current events, prices, live statusLeaving it on for pure analysis of material you already supplied, adding noise and latency for no benefit
ConnectorsA live link to an external system (drive, mailbox, workspace tool)Recurring need for current data from a system of recordConnecting a whole mailbox or drive when only one folder's worth of content was actually needed (§7.2)
SkillsA saved, repeatable way of doing a specific taskA task pattern your team performs the same way often enough to standardizeBuilding a one-off Skill for something that only happens once, instead of just prompting directly
MemoryCarries relevant detail forward from past conversations automaticallyPersonal convenience across your own sessionsAssuming memory is a substitute for documented, team-visible Project instructions (§5.3)
Surfaces are not access controls Which surface you use (chat vs. Project vs. Artifact) governs convenience and context-sharing, not who else can see sensitive data. A Project shared with the wrong team, or a connector pointed at a mailbox with more than intended in it, is a governance mistake regardless of which surface it happened on — see §8.2.

5.2 · Choosing a Model

Reuse the tiering from §2.3: match model capability to task complexity, not the other way around. A high-volume, low-complexity job (short customer-reply drafts, simple tagging) run through the most expensive, slowest model wastes budget and time without buying meaningfully better output. A complex, high-stakes analysis run through the fastest, cheapest model risks a shallower result than the task deserves.

Tier (conceptual)Typical fitTradeoff
Fast / smallHigh-volume, low-complexity, latency-sensitive work — simple replies, tagging, short extractionCheapest and quickest, but shallower on nuance and multi-step reasoning
BalancedEveryday business writing, analysis, and draftingThe default good-enough choice for most tasks; not the deepest option for genuinely hard problems
Most capableComplex reasoning, high-stakes analysis, long or nuanced documentsSlower and costlier — worth it only when depth actually changes the outcome

Exact model names, generation numbers, and which tier is "current" all change over time — this table is deliberately generic. Check Anthropic's current model documentation for what's live today rather than memorizing a name.

5.3 · Managing Context & Memory

Two related but different things:

  • Context is what's visible to Claude right now — this conversation, any files attached to it, and a Project's instructions/knowledge if you're in one. It resets with a new chat (outside a Project's persistent knowledge).
  • Memory, where available, lets Claude carry forward relevant details from your past conversations so you don't have to restate your role or preferences every time. It is a convenience layer, not a substitute for putting stable, important instructions in a Project where the whole team can see and edit them.

Know when to restart, summarize, or persist: restart a conversation that's grown long and inconsistent; summarize progress before restarting if you need continuity; persist genuinely stable, reusable information (style guides, standing instructions, reference documents) in a Project instead of re-explaining it every session.

Memory is not a source of truth Memory carries forward what you told Claude in past conversations, not verified facts about your company or a guarantee of consistency. Treat anything Memory surfaces the same way you'd treat any other Claude output — worth checking, not worth citing as authoritative. Standing facts your whole team relies on belong in documented Project instructions, not personal Memory.

↑ back to top

6 · Workflow Integration & Solution Design Domain 4 · 16%

Objectives: analyze requirements and use cases, use Claude for research/planning/process optimization, support solution design and iteration, integrate Claude into existing workflows, and communicate Claude's value and limits to stakeholders.

6.1 · Analyzing a Use Case

Before automating or augmenting a process with Claude, ask three questions: Is this task language-shaped? (drafting, summarizing, classifying, extracting, explaining — Claude's strengths). Is the risk of an error tolerable given the review step you'll put around it? Does a person still need to own the outcome? A process where the answer to the third question is "no, it can run unsupervised" is usually a sign the task has moved beyond this certification's scope and needs an engineer to build proper guardrails.

A helpful draft is not the same as a validated process A single good result from Claude on a use case doesn't mean the use case is ready to roll out. Test it on a handful of real, messy examples — not just the clean one you tried first — before treating it as a repeatable workflow.

6.2 · Workflow Recipes

Five patterns cover most business use of Claude. Each is really the same shape: give Claude the material, be specific about the transformation you want, specify the audience and format.

RecipeShape of the prompt
SummarizePaste the source, state the audience and length, name what to prioritize (decisions, risks, numbers — not "everything").
ExtractPaste the source, specify exactly which fields/values to pull and the format to return them in (a table is usually easiest to check).
DraftGive the goal, audience, tone/examples, and length; ask for a first pass, then iterate rather than expecting a final version immediately.
AnalyzeGive the data or documents, ask a specific question, request the reasoning before the conclusion so you can sanity-check it.
TranslateProvide the text and target audience/register (formal vs. casual, technical vs. general); for anything customer-facing or legal, have a fluent human reviewer check it — translation nuance is a real evaluation risk.

The same six recipes again, this time by what triggers reaching for Claude at all and what to check before you actually use the result:

Recurring taskTriggerPrompt patternCheck before you use it
Weekly status summaryA report or thread has piled up and needs distilling for an audience that won't read the whole thingPaste the source, name the audience and what to prioritize, ask for a fixed lengthSpot-check 2-3 claims against the source; confirm nothing decision-relevant was dropped
Data extraction into a tableThe same fields need pulling from many similar documentsPaste one example, specify the exact fields and the output formatConfirm every row, not a sample — extraction errors don't announce themselves
First-draft external emailA routine but customer-facing message needs writingGive goal, audience, tone, and constraints; ask for a first passVerify every fact and figure before sending; check tone reads right for the recipient
Rewriting for a new audienceExisting content needs to work for a different reader (technical → exec, formal → casual)Give the original and name what must stay the same vs. changeConfirm meaning survived the rewrite, not just that the wording changed
Comparing optionsA decision needs weighing two or more choices against criteria that matterGive both items and your criteria, ask for a structured tableVerify each criterion's cell independently — don't trust a single "winner" line
Brainstorming a solution setA problem needs more options than the room has generated on its ownState the constraint and goal, ask for a wide list before any rankingCheck ideas are genuinely distinct, not the same idea worded five ways
Classifying a backlogA queue of items needs sorting into a fixed set of categoriesGive the items and the exact allowed categories, ask for one category plus a reason per itemHand-review borderline or low-confidence items rather than assuming full consistency

6.3 · Communicating Value & Limits to Stakeholders

Part of this domain is explaining Claude accurately to colleagues who haven't studied it: what it speeds up (drafting, first passes, structuring), what it doesn't replace (accountability, verified facts, domain expertise, final sign-off on anything regulated), and where the review step lives in a new workflow. Overselling what Claude can do to a stakeholder creates exactly the kind of ungoverned, unverified use that Domain 6 exists to prevent.

"It'll save us a headcount" is a red flag, not a pitch Framing Claude to stakeholders as a replacement for a role rather than an assist to one tends to produce exactly the unsupervised, unaccountable use this domain warns against. The accurate pitch is faster first drafts and less repetitive work with a human still owning the outcome — not fewer humans in the loop on anything consequential.

6.4 · Escalation Signals

Not every gap belongs to the same fix. Three different escalation targets show up on this exam, and mixing them up is a common wrong answer:

Sign you're seeingEscalate toWhy that target
The task needs an API integration, custom automation, or an agentic system wired into other softwareClaude DeveloperBuilding against the API and designing agentic systems is explicitly outside Associate scope
The task needs enterprise-scale architecture decisions — how Claude fits across many systems and teamsClaude ArchitectArchitecture-level design is a separate, higher-scope credential
The output needs legal, medical, financial, or other specialist judgmentThe qualified human expert your organization already usesNo amount of Claude configuration substitutes for professional accountability
The stakes of an error are high and you genuinely can't verify the claim yourselfA human reviewer with the standing to check itVerification effort has to scale with stakes (§4.2), and some verification requires expertise you may not have

↑ back to top

7 · Configuration & Knowledge Management Domain 5 · 12%

Objectives: configure Projects with instructions and knowledge sources, manage uploaded knowledge and connectors, write effective system-level instructions, and keep configuration current.

7.1 · Configuring a Project

A Project bundles three things: instructions (standing guidance Claude follows in every chat inside the Project — tone, format, constraints, "always do X"), knowledge sources (reference documents everyone in the Project can draw on without re-uploading them each time), and the chats themselves, which share that context. Use a Project the moment a task repeats — see the decision table in §11.

Writing a good system-level instruction

  • State the Project's purpose and audience up front, not just a list of rules.
  • Be specific about tone and format ("always answer in a table," "never use jargon without defining it").
  • Name what's out of bounds as clearly as what's expected — an instruction that only says what to do leaves gaps around what not to do.
  • Keep it short enough that a teammate can actually read and remember it; a wall of rules gets skimmed and ignored.

The most common configuration mistake is putting the right content in the wrong place. A quick sort:

Goes inWhat belongs thereExample
Project instructionsStanding rules that apply to every chat in the Project — tone, format, constraints, scope"Always answer in a table." "Never quote a price without flagging it as an estimate."
Uploaded knowledgeReference material chats can draw on but that isn't a rule about behavior — documents, data, past examplesThe current style guide, last quarter's report, a set of approved product descriptions
The individual promptWhat's specific to this one request and not reusable — this week's numbers, this one recipient, today's exact ask"Using the above, draft this week's update for the Tuesday leadership meeting."
Exam pattern A scenario shows someone repeatedly re-typing the same tone and format rules into every prompt inside a Project that already exists. The fix is moving that standing guidance into the Project's instructions once, not writing a longer individual prompt each time.

7.2 · Uploaded Knowledge & Connectors

Uploading a file puts specific content directly in front of Claude for that Project or chat — good for a fixed document (a policy, a style guide, last quarter's report) that doesn't change often. Connectors link Claude to a live external source (a shared drive, an email account, a workspace tool) so it can pull current information without you manually exporting and re-uploading it every time. See the "file vs. connector" line in §11 for exactly when each fits.

Governance angle A connector to a live mailbox or shared drive can expose far more than the one document you meant to share. Before connecting a data source, know what's actually in it and whether your organization's data policy allows Claude to see all of it, not just the part you care about.

Zooming out to every way information reaches Claude, not just the two above:

Knowledge sourceFreshnessGeneral size limitsWhen to useFailure mode
Pasted textAs current as what you typed, right nowSmall — a paragraph or two before it's easier to upload insteadQuick, one-off context you won't reuseGets lost once the chat ends; nobody else can find it later
Uploaded fileFixed as of upload — stale the moment the source changesFinite and counted against context; large files may need trimming or splitting — check current limits rather than assumingA stable reference document that doesn't change often
Project knowledgeFixed as of upload, same as any file, but shared and visible to the whole ProjectSame general constraints as an uploaded file, aggregated across everything in the ProjectReference material the whole team should draw on consistently
ConnectorLive — reflects the connected system as of the moment it's queriedGoverned by the connected system, not a fixed upload sizeRecurring need for current data from a system of record
Web searchLive, public web as of the searchNot applicable in the same sense — bounded by what a search returnsCurrent events, prices, or status not covered by anything you've provided
Freshness is the variable, not quality None of these five sources is inherently "more accurate" than another — a stale uploaded file and an out-of-date pasted paragraph fail the same way. The question to ask is whether the information is current enough for the task, not which method you used to get it in front of Claude.

7.3 · Maintaining Instructions Over Time

Configuration goes stale: a style guide gets revised, a policy changes, a Project's purpose drifts as the team's work evolves. Treat Project instructions and knowledge sources like any other living document — review them periodically, remove outdated files, and update instructions when you notice Claude following guidance that no longer matches how the team actually works.

↑ back to top

8 · Responsible & Safe Use Domain 6 · 15%

Objectives: identify appropriate and inappropriate use cases, apply data-sensitivity/regulatory/privacy considerations, follow organizational AI policy, and understand the ethical implications of AI use.

8.1 · The Usage Policy in Practice

Anthropic publishes a Usage Policy that sets the outer bounds of acceptable use of Claude — prohibiting things like generating content for illegal activity, weapons development, deliberate deception at scale, and other clearly harmful uses. For this exam, the practical takeaway is simpler than memorizing every clause: Claude is a tool, and the person directing it is responsible for how it's used. Two layers of judgment apply on top of Anthropic's policy: your organization's own AI usage policy (often stricter, and specific to your industry), and plain professional judgment about a use case that seems fine in the abstract but is wrong for a specific, sensitive situation.

Three layers, working together in practice:

LayerSetsWhat happens if it's ignored
Anthropic's Usage PolicyThe outer bound of acceptable use for everyone, everywhereAccess to Claude itself can be restricted — this floor isn't negotiable at the individual or org level
Your organization's AI policyRules specific to your industry, data, and risk tolerance, usually stricter than the floor aboveAn org-level compliance or data-handling incident, even if Anthropic's policy was technically followed
Individual professional judgmentThe specific, situational call neither policy could fully anticipateA use case that's technically permitted but wrong for this particular sensitive situation

8.2 · What Never to Paste In

The single most testable, practical habit in this whole domain: know what categories of information should not go into a prompt, an upload, or a connector without your organization's explicit sign-off.

  • Secrets and credentials — passwords, API keys, access tokens, internal system details.
  • Regulated personal data — health information, financial account details, government ID numbers, anything covered by a privacy regulation your organization must comply with — unless it has been redacted or anonymized and policy allows the remaining use.
  • Other people's personal data you don't have a clear right or need to share, even if it's "just for internal analysis."
  • Confidential business information beyond what the task actually requires — trade secrets, unreleased financials, legal strategy.
  • Anything your organization's policy specifically restricts, even if it seems harmless to you personally — the point of an org policy is that individual judgment isn't the final word on regulated data.
Exam pattern A scenario describes uploading a file with names and account numbers so Claude can "just look for trends." The correct move is to remove or anonymize the identifiers first, not to upload as-is because the analysis is "internal," and not to upload anyway while telling Claude "don't retain this" — that instruction doesn't satisfy a policy control.

The same five categories, with the reasoning and the safer path for each:

CategoryWhy it's a problemSafer alternative
Secrets and credentialsA prompt or upload isn't a secure secrets store, and exposure can compromise a system well beyond the current taskNever include them; reference that a credential exists without pasting its value
Regulated personal dataPrivacy regulations govern how this data can be processed and by whom, independent of intentRedact or anonymize identifiers before sharing, or get explicit policy clearance first
Other people's personal dataYou may not have the right or the need to share it, even internallyAsk whether the task actually requires the personal detail, or just the pattern behind it
Confidential business informationTrade secrets, unreleased financials, or legal strategy carry consequences beyond the immediate task if exposedShare only the minimum needed for the task, not the full underlying document
Unreleased materialPre-announcement product, financial, or legal information can carry disclosure or insider-trading implications if it leaksTreat it with the same handling rules as any other pre-release confidential document

8.3 · Bias, Transparency & Ethics

  • Bias: Claude's output reflects patterns in its training data and can carry the same biases present in that data. Treat outputs touching hiring, performance evaluation, lending, or other consequential decisions about people with extra scrutiny, and keep a human decision-maker accountable for the outcome.
  • Transparency with colleagues and customers: don't present AI-drafted work as entirely human-authored when that distinction matters to the audience, and don't hide from a colleague that a shared document or analysis started as an AI draft — this is both an ethics point and a practical trust issue on a team.
  • Human oversight: the recurring theme across this domain is that Claude assists judgment, it doesn't replace the accountable human — captured concretely in "recognize when human review, validation, or escalation is required" from the Minimally Qualified Candidate profile.
Neutral-sounding is not the same as neutral Claude's fluent, even-handed tone can make a biased output feel objective simply because it reads calmly. Judge outputs touching people-decisions by their actual content and pattern across cases, not by how measured the language sounds.

↑ back to top

9 · Working with Claude in a Team

Not a separately weighted exam domain on its own, but it threads through Domains 4, 5, and 6 (workflow integration, configuration, and governance), and it's where most of the "how do we actually roll this out" real-world value lives.

9.1 · Shared Projects

A Project shared across a team turns individual prompting skill into an organizational asset: one set of instructions, one set of reference documents, one consistent voice, instead of every person reinventing their own approach (and their own mistakes) from scratch. The team member who sets up the Project should document why each instruction exists, not just what it says — that context is what lets someone else maintain it later.

9.2 · Admin & Team Concepts

Organizations using Claude at team or enterprise scale typically have an administrator role responsible for provisioning access, setting organization-wide policy defaults, and managing which connectors and data sources are permitted. Specific plan names, seat limits, and admin controls change over time and vary by current offering — check Anthropic's current plan and admin documentation rather than relying on a fixed number here. The concept that's stable and exam-relevant: administrative and governance controls exist at the organization level precisely so that individual judgment isn't the only line of defense for the concerns in §8.

9.3 · Getting Adoption Without Chaos

  • Start with one well-scoped Project for a real, recurring pain point, not a company-wide mandate on day one — a working example spreads faster than a policy memo.
  • Write down the guardrails once (what data can go in, what needs human review, which model tier for which task) so new team members inherit judgment instead of rediscovering it themselves.
  • Expect uneven skill levels. Some teammates will write excellent prompts immediately; others need the worked examples and decision tables in this guide. Pair people rather than assuming a single training session fixes it.
  • Revisit configuration periodically (§7.3) — adoption chaos usually shows up months in, as instructions drift out of date and nobody owns updating them.

↑ back to top

10 · Troubleshooting & Optimization Domain 7 · 10%

Objectives: identify, diagnose, and resolve issues with underperforming prompts or poor outputs; adjust approach based on feedback and results; optimize workflows for efficiency and effectiveness.

SymptomLikely causeFix
Output is generic or misses the pointPrompt was too vague, or the wrong context was providedAdd specifics: audience, format, length, what to prioritize (§3.1)
Output is confidently wrongHallucination, or a question outside what Claude can actually knowVerify facts independently; provide the source material instead of relying on recall (§4.1)
Claude "forgets" an earlier instructionConversation has grown too long, filling the context windowSummarize and restart, or move standing instructions into a Project (§2.2, §5.3)
Answers are inconsistent across a teamNo shared instructions or reference materialConsolidate into a shared Project (§7.1, §9.1)
Task is taking too long / costing too muchWrong model tier for the task, or task not decomposedMatch model to complexity (§5.2); break the task into steps (§3.2)
Output format is unusable downstreamDidn't specify format needsAsk for the specific structure you need — table, list, Artifact (§4.3)
The general method Diagnose before you re-prompt: is the problem the instruction, the context, the model choice, or the underlying task itself? Fixing the wrong layer wastes the most time — a vague prompt to a more expensive model still produces a vague answer.

↑ back to top

11 · Decision Guides

The exam frequently gives you two plausible-sounding options and expects you to pick the one that actually fits. Memorize these discriminators.

Project vs. one-off chat

SituationChoose
Task repeats weekly with the same tone/format/reference materialProject
One-time question, no reuse expectedOne-off chat
Team needs a consistent voice across many peopleProject
Sensitive, one-off exploration you don't want lingering as reusable contextOne-off chat (and mind §8.2 regardless)

Upload a file vs. paste text

SituationChoose
Long document, spreadsheet, or something you'll reference againUpload
A short paragraph or two, quick one-off contextPaste
Data needs to persist for a whole Project's lifetimeUpload as a Project knowledge source
Content changes frequently and lives in a live systemConnector, if your org permits it, rather than repeated manual uploads

When to use an Artifact

SituationChoose
Output you'll keep editing, export, or share standaloneArtifact
Quick answer you'll read once and move onInline chat text
Structured content (table, list, document) that benefits from its own viewArtifact

When web search helps vs. hurts

SituationGuidance
Question needs current events, prices, or statusHelps — turn it on
Task is pure drafting/analysis of material you already providedHurts or is irrelevant — off, avoid unnecessary latency/noise
Highly sensitive internal question with no public answerHurts — searching won't help and may not be appropriate to run at all

When to break a task into steps

SituationGuidance
Request has multiple distinct deliverablesDecompose (§3.2)
You need to check an assumption before building on itDecompose, verify the first step, then continue
Simple, single-purpose askOne prompt is fine

When to escalate to a developer or engineer

SituationEscalate?
Task needs an API integration, custom automation, or agentic systemYes — Claude Developer/Architect scope
Task needs enterprise-scale architecture decisionsYes — Claude Architect scope
Task is a business workflow using built-in features (Projects, Artifacts, connectors already offered)No — this is Associate scope
Output needs specialist judgment you don't have (legal, medical, financial sign-off)Escalate to the appropriate human expert, not a technical role

↑ back to top

12 · Study Materials

This certification is new — the official Exam Guide is dated July 2026 — so be conservative about anything claiming to be a comprehensive practice-question bank for it. A vendor advertising "guaranteed" real exam questions for a brand-new exam is a red flag, not a shortcut.

1 · Official Anthropic resources (start here)

  • The official Exam Guide PDF, available from the certification's page on the Anthropic Partner Academy (anthropic-partners.skilljar.com). Read it in full — it includes the blueprint, sample questions with rationale, scoring, and every policy you're bound by.
  • Anthropic's help center and documentation (docs.claude.com and support.claude.com) for Projects, Artifacts, Memory, Skills, and connectors — the exam guide itself points here for feature knowledge, not to a third-party course.
  • Anthropic's prompt engineering guidance, the primary source for the techniques in §3.1.
  • Anthropic Academy prep courses, if offered for this credential on the Partner Academy platform at the time you're studying — check what's currently listed, since course offerings change.

2 · Community material — read with caution

Because CCAO-F is new, community write-ups and third-party practice exams for it have little track record yet. Treat any third-party practice set as a rough style guide at best, not a predictor of your score, until it has been through at least a few real exam cycles. Prefer material that explains why an answer is correct over material that just hands you an answer key.

3 · What's usually not needed

You do not need a coding environment, an API key, or SDK familiarity to prepare for this exam — if a resource is teaching you to call the Claude API, it's studying for the wrong certification (that's Claude Certified Developer territory). The best preparation is hands-on daily use of Claude on real work, not passive video-watching.

Sources

  • Official: Anthropic's CCAO-F Exam Guide PDF (v1.0, effective July 2026), retrieved from the Anthropic Partner Academy · the public Claude Certified Associate – Foundations certification page · Anthropic's public documentation and help center.
  • Community: general awareness that third-party practice content for this specific exam is early and unproven as of this writing — verify recency and track record before trusting any of it.

↑ back to top

13 · Study Plan (≈2 weeks)

This is a foundational, non-engineering exam, so the plan is short and deliberately hands-on. Passive reading will not build the judgment Domain 2 and Domain 7 test — daily real use will.

The learning principles this plan uses

PrincipleHow to apply it to CCAO-F
Active recallAfter reading a domain's objectives, close the guide and write out what each one means in your own words before checking.
Spaced repetitionRevisit your prompt journal and the glossary every few days rather than once at the end.
Teach-backExplain "what a hallucination is and how you'd catch one" or "when to use a Project vs. a chat" out loud to a colleague, unscripted.
Deliberate practiceRedo a prompt that didn't work well, don't just move past it — the redo is where the learning happens.

The daily plan

PhaseFocusActions
Days 1–2OrientationRead the official Exam Guide PDF in full. Skim Anthropic's documentation for Projects, Artifacts, Memory, Skills, and connectors. Start a prompt journal.
Days 3–9Daily practiceUse Claude on real work every day: at least one summarize, one draft, and one analyze task across the week. Log one prompt that worked and one that didn't each day; redo the weak one the next day.
Days 10–11Governance & evaluation deep-diveStudy §4 and §8 hard — together over a third of the exam. Write your own "never paste in" list from memory, then check it against §8.2.
Days 12–13Practice & decision tablesWork every worked example in §16 without looking at the answer first. Drill the decision tables in §11 until you can answer them from memory.
Day 14Light reviewSkim the Cram Jam and glossary. No new material. Sleep well.
Readiness bar You're ready when you can explain, from a blank page, what a hallucination is and how to catch one, when to use a Project vs. a one-off chat, what never to paste into Claude, and when a task belongs with an engineer instead of you.

↑ back to top

14 · Exam-Day Tips & Traps

Before you sit

  • Closed book, proctored, no exceptions. No Claude, no documentation, no browser tabs, no external tools, no notes, no second monitor, no phone in the room. If testing from home, clear your desk and be prepared for the proctor to review your space on camera.
  • Bring a valid, unexpired, government-issued photo ID that matches your registration name exactly.
  • You'll accept a confidentiality/non-disclosure agreement before the exam starts — declining ends the session with no refund.

Strategy

  • 120 minutes / 60 questions is 2 minutes a question on average — comfortable pacing. Flag anything you're unsure of and move on rather than stalling.
  • Items are standalone, not shared scenarios — nothing from question 12 carries into question 13. Read each one fresh.
  • Multiple-response items state exactly how many answers to pick — select that number, not more or fewer, even if another option looks tempting.
  • When two options both sound reasonable, check which one matches the "verify, don't assume" pattern that runs through Domain 2, or which one keeps a human accountable rather than delegating judgment entirely to Claude.

Recurring traps

  • Confidence is not accuracy. A fluent, specific-sounding answer can still be a hallucination — the correct move is almost always independent verification, not trusting the tone.
  • "Always use the best/most powerful model" is a wrong-answer pattern. Matching model to task complexity and cost is the tested skill.
  • Uploading sensitive data "because it's just for internal use" is a wrong-answer pattern. Redact or anonymize first, or don't proceed without policy clearance.
  • This exam is not testing API or coding knowledge. If an answer choice requires writing code, calling an API, or building an integration, it is very likely the wrong choice for an Associate-level item — that's Developer or Architect scope.
  • Self-rated confidence from Claude is not validation. Asking Claude "are you sure?" and trusting a confident "yes" doesn't satisfy a fact-checking requirement.
Keyword → intent "confident but unverified" = hallucination risk · "recurring task, shared context" = Project · "standalone document to keep/share" = Artifact · "sensitive data, internal use" = still needs redaction/policy check · "cost and speed matter more than depth" = smaller/faster model · "complex reasoning, high stakes" = most capable model · "beyond your role's scope" = escalate to a human expert or an engineer.

↑ back to top

15 · Glossary

TermPlain-language meaning
LLM (large language model)An AI system trained on large amounts of text to generate fluent, relevant language in response to a prompt. Claude is an LLM.
PromptThe instruction or question you give Claude.
TokenA small chunk of text (roughly part of a word) that Claude reads and writes in; the unit its context window is measured in.
Context windowThe total amount of text — prompt, files, conversation history, and reply — that fits into one exchange with Claude.
HallucinationConfident, fluent output that is factually wrong or fabricated.
ProjectA container in Claude for recurring work: shared instructions, reference documents, and the chats that use them.
ArtifactA standalone piece of output (document, table, simple interactive view) Claude creates that you can view, edit, and share separately from the chat.
ConnectorA link between Claude and an external, live data source (like a shared drive or email account) so Claude can use current information from it.
MemoryA feature that lets Claude carry forward relevant details from past conversations without you restating them each time.
SkillA repeatable, saved way of doing a specific task that Claude can apply consistently, often shareable across a team.
Model tier (Haiku/Sonnet/Opus)Anthropic's spectrum from fastest/cheapest to most capable; match tier to task complexity.
Task decompositionBreaking a complex request into smaller, ordered steps.
Usage policyAnthropic's published rules for acceptable use of Claude, on top of which your own organization's policy usually adds further restrictions.
EscalationHanding a task off to someone better positioned to own it — a human expert, or a Claude Developer/Architect for technical build work.
Standalone itemAn exam question that is self-contained, not part of a linked multi-question scenario — the format CCAO-F uses throughout.
Scaled scoreYour exam result expressed on a fixed 100–1000 scale rather than a raw percentage; 720 is the passing cut score.
Criterion-referencedScored against a fixed standard of required knowledge, not ranked against other test-takers.

↑ back to top

16 · Worked Example Questions

Standalone items in the exam's own style. Each states how many answers to select, gives an answer, and explains why every distractor is wrong.

1 · Domain 2 — Output Evaluation

Claude drafts a client-facing summary of a new regulation and cites a specific subsection number with full confidence. Before it goes to the compliance team, what should you do? (Select 1.)

  1. Send it as-is; Claude sounded certain.
  2. Verify the cited subsection against the official regulation text first.
  3. Ask Claude to rate its own confidence and send it if the rating is high.
  4. Reword it to sound more formal, then send it.

Answer: B. Confident tone is not evidence of accuracy, and specific citations are a classic hallucination risk. A is wrong because tone isn't a validity check. C is wrong because self-reported confidence is not a reliable accuracy signal. D is wrong because rewording doesn't touch correctness.

2 · Domain 3 — Product & Model Selection

You need to generate hundreds of short, simple customer-reply drafts where speed and cost matter more than deep reasoning. (Select 1.)

  1. Use the most capable, highest-cost model for every reply.
  2. Use a faster, lower-cost model suited to high-volume, straightforward tasks.
  3. Turn off all product features to cut cost.
  4. Switch to a different AI platform.

Answer: B. Matching model tier to task complexity is a named skill; the top-tier model wastes budget and latency on simple work. A ignores the cost/speed trade-off. C and D don't address the actual requirement.

3 · Domain 6 — Governance

A colleague wants to upload a spreadsheet of customer names and account numbers so Claude can spot trends. Company policy restricts sharing regulated personal data. (Select 1.)

  1. Upload it as-is; the analysis is internal.
  2. Remove or anonymize the identifiers before uploading.
  3. Upload it but tell Claude not to retain it.
  4. Skip the analysis.

Answer: B. Redacting or anonymizing regulated identifiers lets the analysis proceed within policy. A violates policy outright. C doesn't satisfy a policy control just because you asked nicely. D is unnecessary once anonymization is possible.

4 · Domain 1 — Prompting

A prompt asking Claude to "improve this email" repeatedly returns generic edits that miss the point. What's the most likely fix? (Select 1.)

  1. Ask a more capable model the exact same prompt.
  2. Specify the audience, the goal of the email, and what "improve" should mean (tone, length, persuasiveness).
  3. Repeat the same prompt several times and pick the best result.
  4. Shorten the prompt so Claude has less to process.

Answer: B. A vague instruction produces a vague result regardless of model; specificity about audience and goal is the fix. A treats a prompting problem as a model problem. C wastes effort without addressing the ambiguity. D removes information the task actually needs.

5 · Domain 5 — Configuration

A team keeps getting inconsistent tone and formatting from Claude across five different people doing the same weekly report. What should they do? (Select 1.)

  1. Ask each person to write longer prompts.
  2. Set up a shared Project with standing instructions and the reference material everyone should use.
  3. Have one person do all the reports going forward.
  4. Switch everyone to the same model tier.

Answer: B. A shared Project centralizes instructions and reference material so output is consistent regardless of who's prompting. A doesn't fix inconsistency across people. C avoids the actual problem rather than solving it. D doesn't address tone/format consistency.

6 · Domain 4 — Workflow Integration

A stakeholder asks you to have Claude "just automatically approve routine expense reports without anyone checking." What is the appropriate response? (Select 1.)

  1. Set it up as requested since it will save time.
  2. Explain that unsupervised approval removes human accountability and exceeds this role's scope; propose Claude assist with a flagged draft that a person still approves, or escalate the automation to an engineer if full automation is truly needed.
  3. Refuse to discuss the idea at all.
  4. Have Claude approve reports but only on Fridays.

Answer: B. This combines recognizing an inappropriate use case (Domain 6) with knowing when to escalate technical automation (workflow scope). A removes needed human oversight. C is unhelpfully dismissive rather than constructive. D doesn't solve the actual concern.

7 · Domain 2 — Output Evaluation

Claude produces a market analysis with a statistic you can't locate in any of your source documents or a quick outside check. (Select 1.)

  1. Keep the statistic since Claude is generally reliable.
  2. Remove or flag the statistic until it can be verified against an authoritative source.
  3. Change the number slightly so it seems more plausible.
  4. Ask a different AI tool to confirm it.

Answer: B. An unverifiable specific claim should be flagged or removed, not kept on faith. A relies on general reliability rather than checking this specific claim. C fabricates further rather than verifying. D doesn't constitute checking against an authoritative source.

8 · Domain 7 — Troubleshooting

A long-running chat starts giving inconsistent answers and seems to have lost track of an instruction from earlier. (Select 1.)

  1. Add three more instructions to the same conversation to reinforce the point.
  2. Summarize what matters so far and start a fresh conversation, moving stable instructions into a Project if they'll be needed again.
  3. Assume the model is broken and wait a few hours.
  4. Switch to a less capable model.

Answer: B. This is a context-window symptom; the fix is to reset context, not pile on more instructions into an already-full conversation. A worsens the underlying problem. C misdiagnoses the cause. D doesn't address context length at all.

9 · Domain 3 — Product Selection

You need Claude's answer to reflect a fast-moving news topic from this week. (Select 1.)

  1. Rely on Claude's training knowledge alone.
  2. Use a Claude surface with web search/research enabled so it can pull current information.
  3. Ask Claude what today's date is and trust its answer.
  4. Wait for the next model update.

Answer: B. Current events require a live information source; training knowledge has a cutoff. A will likely produce outdated or fabricated specifics. C doesn't solve the underlying knowledge gap. D isn't a practical near-term fix.

10 · Domain 6 — Governance

An employee wants to use Claude to draft performance review language for their direct reports and have Claude also assign the final numeric ratings. (Select 1.)

  1. Let Claude assign ratings directly to save management time.
  2. Use Claude to help draft clear, well-organized review language, but keep the manager as the accountable decision-maker for ratings, watching for bias in the drafted language.
  3. Avoid using Claude for anything related to performance reviews.
  4. Have Claude assign ratings, then have the manager sign off without reviewing them.

Answer: B. This keeps a human accountable for a consequential decision about people while still using Claude productively, and flags the bias risk named in this domain's objectives. A and D hand a high-stakes, bias-sensitive decision to the model. C is an overcorrection that discards a legitimate, lower-risk use (drafting language).

11 · Domain 1 — Task Decomposition

A single prompt asks Claude to "redesign our onboarding process, write the new emails, and summarize last quarter's feedback," and the result is shallow on all three. (Select 1.)

  1. Ask for the same three things again in one prompt, in a different order.
  2. Split the request into separate steps — first the feedback summary, then the redesign, then the emails — reviewing each before moving to the next.
  3. Pick just one of the three tasks and drop the other two permanently.
  4. Ask Claude to guess which of the three matters most.

Answer: B. This is a textbook decomposition case — one broad ask covering three distinct deliverables. A doesn't change the underlying problem. C abandons legitimate work instead of restructuring the request. D offloads a decision you should make yourself.

↑ back to top