1. Exam at a Glance
| Item | Value |
|---|---|
| Exam code | CCAO-F |
| Number of items | 60 |
| Item format | Multiple-choice and multiple-response. Each item states how many responses to select. |
| Time limit | 120 minutes |
| Passing score | Scaled score of 720 on a scale of 100–1,000 |
| Scoring | Criterion-referenced. You are measured against a fixed standard, not against other candidates. |
| Result reporting | Pass/fail with scaled score, plus percent-correct by domain |
| Validity | 12 months; free non-proctored renewal assessment if you renew on time |
| Retakes | 14 days after the first fail, 30 after the second, 90 after the third; max four attempts per rolling 12 months |
| Prerequisites | None. No software-development or API experience needed. |
2. The Seven Domains
Item counts assume a 60-question exam. Domain 2 alone is more than a fifth of your score.
| Domain | Weight | Approx. items | Lab |
|---|---|---|---|
| 1. Prompting and Task Execution | 14% | 8 | Lab 1 |
| 2. Output Evaluation and Validation | 21% | 13 | Lab 2 |
| 3. Product and Model Selection | 12% | 7 | Lab 3 |
| 4. Workflow Integration and Solution Design | 16% | 10 | Lab 4 |
| 5. Configuration and Knowledge Management | 12% | 7 | Lab 5 |
| 6. Governance, Risk, and Responsible Use | 15% | 9 | Lab 6 |
| 7. Troubleshooting and Optimization | 10% | 6 | Lab 7 |
3. Prompt Anatomy
An effective prompt removes the guesses Claude would otherwise make. Specificity, not length.
| Element | Question it answers | Missing it causes |
|---|---|---|
| Task | What specific verb, on what? | Claude picks a deliverable for you |
| Context | Who is this for, what has been tried, what constrains it? | Generic output that ignores your situation |
| Format | What shape should the answer take? | Right content, unusable form |
| Criteria | What makes the answer good? | Arbitrary ordering, indefensible conclusions |
| Failure instruction | What to do when it cannot answer? | Confident fabrication |
4. Prompting by Task Type
| Task type | Needs most | Move |
|---|---|---|
| Analysis | Data plus the criteria for judging it | Supply the standard; require conclusions to cite supporting data |
| Research | Scope boundaries and a stance on uncertainty | Say what is in and out of scope; require flagging of anything unverified |
| Drafting | Audience, tone, length, a model of good | Name the reader; show the voice you want or one you do not |
| Brainstorming | Range, plus constraints that force novelty | Volume first, judgment later; rule out the obvious category |
5. Choosing the Product Surface
| Use | When |
|---|---|
| Chat | Quick, exploratory, one-off work you will not repeat |
| Projects | Recurring work of the same kind that needs persistent instructions and knowledge |
| Research mode | Multi-source investigation requiring synthesis across material |
| Artifacts | Substantial, self-contained deliverables you iterate on and reuse |
| Structured data (tables, lists) | When prose would obscure comparison |
6. Model Selection
| Model | Profile | Fits |
|---|---|---|
| Haiku | Fastest, lowest cost | Straightforward, high-volume work where speed and cost dominate |
| Sonnet | Balanced | The default for most professional work |
| Opus | Most capable, higher cost and latency | Genuinely complex reasoning that cheaper models get wrong |
The tested insight: defaulting to the most capable model for everything is a mistake, not a safe choice; it wastes the cost and latency budget. Using the cheapest model for nuanced reasoning is equally wrong, because you pay for it in rework.
7. Context: Restart, Summarize, or Persist
| Do this | When |
|---|---|
| Restart | The thread is anchored to bad early turns, or the topic genuinely changed |
| Summarize | You need continuity but not verbatim history; carry forward a condensed state |
| Persist into a Project | The information is needed repeatedly across sessions |
Symptoms of an overgrown conversation: earlier details dropped, contradictions with constraints set far above, degrading focus.
8. Validation Checklist
Domain 2 is the heaviest on the exam. Fluency is not accuracy.
- Identify the checkable claims. Which statements could be true or false?
- Rank by consequence of being wrong. Not by how confident the output sounds.
- Verify against an authoritative source. Especially citations, statistics, case names, quotes, and URLs; specific-looking details are the highest-risk fabrications.
- Check quantities independently. Recompute; do not trust arithmetic presented fluently.
- Check completeness. What was omitted? An answer can be entirely true and still misleading.
- Check for bias, including bias inherited from the source material you supplied.
- Decide on human review using the table below.
| Signal | Reliable? | Why |
|---|---|---|
| Claude’s self-reported confidence | No | Self-rated confidence is not an accuracy signal. Heavily tested. |
| Confident, fluent tone | No | Fluency and accuracy are unrelated. |
| Asking Claude to check its own work | Weak | Can surface some errors; never substitutes for an authoritative source. |
| Verification against the source document | Yes | The only thing that establishes a factual claim. |
9. When Human Review Is Mandatory
The rule: the cost of being wrong determines the review requirement, never the model’s tone or confidence.
| If the output is… | Then |
|---|---|
| Going to a regulator, auditor, or compliance team | Verify every factual claim against the authoritative source first |
| Published externally under your name or the organization’s | Human review before release |
| Legal, medical, or financial in nature | Review by a qualified person; never present as licensed advice |
| Driving an irreversible or high-consequence decision | Human accountability required; Claude may inform, not decide |
| A consequential decision about an individual (hiring, credit, discipline) | Never the sole basis. Watch for disparate impact. |
| Low-consequence, reversible, internal, easily checked | Spot-check is proportionate |
10. Data Sensitivity: Before Anything Is Uploaded
- Classify the data. Public, internal, confidential, or regulated?
- Check organizational policy. Policy is the floor. If policy is silent on a novel case, escalate; do not improvise. If a manager’s request conflicts with policy, follow policy and escalate.
- Identify identifiers: direct (names, account numbers, emails) and quasi-identifiers that re-identify in combination (ZIP + birthdate + role).
- Minimize. Does the task actually need those fields? Usually not.
- Remove or anonymize what remains, then proceed.
| Tempting action | Verdict | Why |
|---|---|---|
| Upload as-is because the analysis is internal | Wrong | Violates policy regardless of intent |
| Upload, but instruct Claude not to retain it | Wrong | An instruction is not a control. A prompt is neither a technical nor a contractual safeguard. |
| Skip the analysis entirely | Unnecessary | Anonymization usually enables the work |
| Remove or anonymize identifiers, then upload | Correct | Satisfies policy and still delivers the analysis |
11. Projects, Instructions, and Knowledge
| Rule | Detail |
|---|---|
| Instructions describe how to work | Role, audience, standing conventions, criteria, and the failure instruction. Not a single task. |
| Adjectives do nothing | “Be helpful, be accurate, be thorough” is not a criterion Claude can measure itself against. |
| More knowledge is not better | Contradictory or superseded documents degrade output; Claude cannot know which you consider authoritative. |
| State precedence | When overlap is unavoidable, say in the instructions which source wins. |
| Uploads | Frozen, bounded, auditable snapshot. Good for a quarterly-frozen manual. |
| Connectors (Drive, Gmail) | Live and current, but widen what Claude can see. Good for constantly-changing material. |
| Maintain it | Review on a cadence and on trigger events. A stale Project is worse than no Project, because it produces confidently outdated answers people trust. |
12. Troubleshooting: Symptom → Cause → Fix
Name the failure mode before you change anything. Then change one variable.
| Symptom | Cause | Fix |
|---|---|---|
| Factually wrong or fabricated | Validation / sourcing | Ground in supplied knowledge, require citations, add the failure instruction |
| Answers a different question | Task specification | Add the task and the criteria |
| Correct but unusable | Format / audience | Specify shape, length, tone, reader |
| Shallow across every part | Scope — the request is compound | Decompose into checkable steps |
| Drifts, contradicts earlier turns | Context outgrown | Summarize and restart |
| Inconsistent between runs | Under-constrained | Tighten criteria |
| Ignores your instructions | Instruction conflict | Find the contradicting earlier turn or Project instruction |
Stop iterating after roughly three deliberate corrections with no movement. The problem is then the thread, the knowledge, or the task, not the prompt. Efficiency without effectiveness is a trap: measure end-to-end including rework and review, not the Claude step alone.
13. The Escalation Boundary
Recognizing the limits of the Associate role is itself an exam objective. Hand off to a Claude Architect or Developer when the work requires:
| Signal | Why it is out of scope |
|---|---|
| Running over thousands of records programmatically | Requires building against the platform, not using the product |
| Integrating with internal systems, or automated triggers | System integration work |
| Operating without a human in the loop | Changes the accountability model and the governance posture |
| Governance enforced technically (access control, retention, audit logs) | An instruction is not a control; enforcement is engineering |
| Systematic evaluation or regression testing of prompts at scale | Requires tooling beyond the product surface |
14. Exam Traps to Memorize
| Tempting answer | Why it is wrong |
|---|---|
| “Ask Claude to rate its confidence” | Self-reported confidence is not an accuracy signal |
| “Regenerate until it improves” | Rerolling is not iterating; it does not address why it failed |
| “Tell Claude to be more thorough” | Not a criterion it can measure itself against |
| “Use the most capable model for everything” | Wastes cost and latency; does not fix a bad prompt |
| “Instruct Claude not to retain the data” | An instruction is not a control |
| “It is internal, so policy does not apply” | Policy applies regardless of intent or audience |
| “Upload everything, more context is better” | Contradictory and stale sources degrade output |
| “Change several things at once since it improved” | You learn nothing about which change mattered |
| “The AI step got faster, so the workflow improved” | Measure end-to-end, including rework |
| “Promise full automation to win the budget” | Overselling autonomy creates a governance problem |