CCAO-F Common Misconceptions

Part of the CCAO-F track. Objective IDs come from registry.yaml. Product facts are as of 2026-10-24 and were checked against support.claude.com.

Each entry gives the misconception, the correction and why it matters. Many wrong answer options in judgement-style exams are built from beliefs like these, so learn to recognise them.

Prompting (Domain 1)

1. "Longer prompts always produce better results." Relevant detail helps; padding, repetition and contradictory rules hurt. Aim for complete, not long: goal, context, audience, constraints, format and success criteria. (CCAO-F-1.1)

2. "Assigning a role ('You are an expert…') is the key to a good prompt." A role can set tone and focus, but it does not supply the missing context, audience or format. A role with no task details still gives generic output. (CCAO-F-1.1, 1.4)

3. "If the first answer is poor, keep adding follow-ups until it is right." Small adjustments work well as follow-ups. When the first attempt went off track, editing the original prompt and regenerating is usually faster and avoids anchoring on the bad draft. (CCAO-F-1.3)

4. "Breaking a task into steps is slower, so it is always worse." For multi-phase or high-stakes work, decomposition improves quality and lets you catch errors between steps. It is only overhead for simple tasks. (CCAO-F-1.2)

5. "The same prompting style works for every task." Analysis, research, drafting and brainstorming need different moves: evidence and reasoning, scope and sources, audience and voice, volume and variety. (CCAO-F-1.4)

Evaluation (Domain 2)

6. "If Claude sounds confident, it is probably right." Tone and fluency are not evidence. Specific details you did not provide (figures, dates, clause numbers, quotes) are the first things to verify. (CCAO-F-2.2)

7. "Asking Claude whether it is sure is a good way to check its answer." Self-assessment is not independent verification. Self-critique can surface gaps, but validation must check a source outside the answer. (CCAO-F-2.3)

8. "A citation means the claim is correct." Citations make checking easier, but you still need to open them and confirm they exist and say what is claimed. (CCAO-F-2.3)

9. "A well-written summary is a complete summary." Summaries can be accurate yet omit the most important point, or skip one of several documents. Check completeness against the brief, item by item. (CCAO-F-2.1)

10. "Human review is a sign the AI is not good enough." Review in proportion to risk is part of responsible use at every quality level. External, irreversible, regulated or people-affecting outputs need qualified human review regardless of the model. (CCAO-F-2.4)

11. "Artifacts are just for code." As of 2026-10-24, artifacts cover documents, decks, dashboards, designs and small interactive tools. Choose them for substantial, shareable or reusable work; choose inline answers for quick replies and structured data for anything going into another tool. (CCAO-F-2.6)

Products and models (Domain 3)

12. "Always use the most powerful model to be safe." The best model is the cheapest one that meets the quality bar. More capable models are slower and use more of your allowance; light models are often fine for simple, repetitive tasks. (CCAO-F-3.3)

13. "A better model means I no longer need to fact-check." More capable models make fewer mistakes, but none is free of hallucination. Verification still depends on stakes. (CCAO-F-3.2, 2.3)

14. "Research and web search are the same thing." Web search fetches current information in a normal chat. Research runs a multi-step, multi-source investigation across the web and your enabled connectors and returns a cited report; it takes longer and uses limits faster. (CCAO-F-3.1)

15. "With a huge context window, I never need to restart a chat." Large windows help, and paid plans automatically summarise older messages in long chats, but long chats still drift, carry superseded information and consume more usage. Restart, summarise or persist. (CCAO-F-3.4)

16. "A new chat automatically knows what I discussed yesterday." Only if memory or chat search is enabled and relevant, or if the information is in Project knowledge or instructions. Do not rely on memory for team-critical information. (CCAO-F-3.4)

Configuration (Domain 5)

17. "Uploaded Project files update when the original changes." Uploaded knowledge is a snapshot. Replace files when sources change, or use a connector for content that changes frequently. (CCAO-F-5.2, 5.4)

18. "A connector shows Claude the same thing for everyone." Each user authenticates individually, so Claude sees what that user can access. Results can differ between users. (CCAO-F-5.2)

19. "Project instructions are a security control." Instructions guide behaviour; they do not stop someone uploading restricted data or guarantee the model never errs. Permissions, admin settings, data policy and human review provide the boundaries. (CCAO-F-5.3, 6.3)

20. "More knowledge files make a Project smarter." Irrelevant, outdated or conflicting files dilute and confuse answers. Curate a small set of current, authoritative documents. (CCAO-F-5.1, 7.1)

21. "Any custom connector listed online is safe because it works with Claude." As of 2026-10-24, the Help Center states custom connectors are not verified by Anthropic. Connect only to trusted services, review requested permissions, and follow organisational approval. (CCAO-F-5.2, 6.3)

Governance (Domain 6)

22. "Telling Claude not to remember sensitive data makes it fine to upload." An instruction to the model does not change how the platform processes or retains data, and does not satisfy a policy that prohibits that data in the tool. Minimise first: redact, pseudonymise, aggregate. (CCAO-F-6.2)

23. "If training is switched off in my personal account, I can use it for work data." Approved tools are about contracts, controls and audit, not one setting. Use the organisation's approved workspace. (CCAO-F-6.2, 6.3)

24. "When data is sensitive, the only safe option is not to use AI at all." Often the better answer is to make the data safe: anonymise or aggregate it, or use synthetic data, within policy. Skip the task only if no compliant route exists. (CCAO-F-6.2)

25. "Internal use means data rules do not apply." Data classification and privacy obligations apply regardless of whether the output stays internal. (CCAO-F-6.2, 6.3)

26. "If the policy does not mention it, it is allowed." Silence is not permission. Ask the policy owner and pause the risky part. (CCAO-F-6.3)

27. "AI can make objective decisions about people because it has no feelings." Models can reflect biases in their training data and prompts. Decisions about employment, credit, health or rights need qualified human judgement, with AI as support at most. (CCAO-F-6.1, 6.4)

Workflow and troubleshooting (Domains 4 and 7)

28. "The best first use case is the one with the biggest potential impact." The best first pilot is frequent, checkable, low risk and uses permitted data. High-impact, high-risk ideas come later, with evidence and controls. (CCAO-F-4.1)

29. "Automating a step means removing the human." Automation usually moves the human from doing to reviewing. Removing review from high-stakes steps requires an equal or better control and approval. (CCAO-F-4.4, 2.4)

30. "Poor output means the model is the problem." Check the inputs first: outdated or conflicting files, contradictory instructions, missing context, the wrong feature, or a drifting chat. (CCAO-F-7.1)

31. "Optimisation means making it faster." Optimise quality per unit of effort. A faster workflow that sends more errors out is worse. Track a quality measure alongside time. (CCAO-F-7.3)

32. "An Associate should build whatever integration the team asks for." The credential explicitly includes recognising limits and escalating complex or technical work, such as system integrations or public-facing products, to Claude Architects and Developers. (CCAO-F-4.3, 4.5)

About the exam itself

33. "A good score on unofficial practice tests means I will pass." Unofficial scores (including this repository's mock exams) are not the official scaled score. Use them to find weak domains, not to predict a result. The official passing standard is in registry.yaml.

34. "Multiple-response items are just multiple-choice with extra options." Each multiple-response item tells you how many answers to select. Treat each option as a true/false judgement against the scenario. See exam-day.md.

Continue with revision-notes.md.