What you will be able to do
- Tell apart tasks where a Claude error can be caught and fixed from tasks where an error lands on a person before anyone checks it
- Use Claude's documented capability limits to decide whether a task needs human oversight
- Judge whether an agent's autonomy fits the work, using the guidance on trusted environments, review, and approval gates
Key concept
Inform, don't decide — A use case is appropriate when someone can review Claude's output and fix mistakes before they reach anyone. When the output affects a person's health, money, rights, or standing, a qualified human reviews it first and makes the decision. Claude informs that decision but does not make it.
1.The core test: can an error be caught before it matters?
On the exam, most "is this an appropriate use of Claude?" questions come down to one question: if Claude gets this wrong, what happens next? Claude works by predicting likely text. That makes it very good at familiar tasks like summarising, reformatting, drafting, and classifying. It also means Claude can produce statements that sound right but are false. Anthropic's course on AI capabilities and limitations describes both sides: it explains why the model is excellent at well-worn paths like summarizing or reformatting, and also why it can produce things that sound true but aren't.
That second side doesn't rule Claude out. It tells you where a human needs to check. A first draft of an internal memo, a summary you will compare against the original, or a reformatted spreadsheet you will look over are all recoverable: a wrong answer costs some editing time. Things change when Claude's output goes straight to a customer, patient, applicant, or employee, and acts on them without anyone checking it. Anthropic's guidance for teams building on the API covers this case: for sensitive information and decision making, a qualified professional reviews the content before it reaches consumers. The review happens before the output goes out, not afterwards.
Anthropic also says plainly that its own safety measures don't remove this responsibility. It describes safety as shared, and calls the organisations that deploy Claude "a second line of defense." So the person or team deploying Claude is responsible for deciding where its output gets reviewed.
Checkpoint 1 of 4· Check yourself
A clinic wants Claude to draft plain-language explanations of lab results, which are then sent to patients automatically. Which change would bring this in line with Anthropic's guidance?
This is sensitive health information going to consumers. The guidance is for a qualified professional to review it before it goes out. A disclaimer, sampling after sending, or a better model all still let unreviewed content reach patients.
“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”Source: support.claude.com
2.Documented limits that change the answer
Some tasks are inappropriate because Claude can't do them reliably, or won't do them at all. Anthropic documents these limits, and they work as a checklist. The course on capabilities and limitations gives the general pattern: Claude's knowledge is strong on mainstream topics and popular languages, weaker on anything rare, recent, niche, or contested. A question about a niche regulation or last week's events needs your own sources, not just Claude's answer.
The vision documentation gives the most specific list of limits. Each one tells you whether to decline a use case outright or to add a human check.
| Proposed task | Documented limit | What it means for the use case |
|---|---|---|
| Naming people in images | Claude cannot be used to name people in images and refuses to do so | Not available: plan for Claude to refuse |
| Interpreting CTs or MRIs | Not designed to interpret complex diagnostic scans; not a substitute for professional medical advice or diagnosis | Inappropriate as a diagnostic tool |
| Detecting fake or synthetic images | Cannot determine whether an image is AI-generated and might be incorrect if asked | Do not rely on it for authenticity checks |
| Counting many small objects | Approximate counts, not always precisely accurate | Fine for estimates; verify before relying on exact numbers |
| Low-quality, rotated, or very small images under 200 pixels | Might hallucinate or make mistakes | Improve the input, then verify the output |
The documentation sums these limits up in one rule: do not use Claude for tasks requiring perfect precision or sensitive image analysis without human oversight. That doesn't ban precise or sensitive work. It says such work needs a human reviewer. The same documentation tells you to review and verify image interpretations carefully, especially for high-stakes use cases.
Checkpoint 2 of 4· Check yourself
A fact-checking desk wants Claude to flag which photos submitted by readers are AI-generated, and to publish only the photos Claude marks as authentic. What is the main problem?
Detecting synthetic images is a documented limitation. Using Claude as the deciding authenticity filter depends on a capability Claude doesn't have.
“Claude cannot determine whether an image is AI-generated and might be incorrect if asked.”Source: platform.claude.com
3.Agents: match autonomy to how easily mistakes can be undone
When Claude takes actions instead of just writing text, the same test applies to each action. The computer use documentation points to work where an occasional mistake is cheap. It suggests focusing on use cases where speed isn't critical (for example, background information gathering, automated software testing) in trusted environments. It also warns that Claude might select the wrong tool or take unexpected actions, so you should always carefully review and verify Claude's computer use actions and logs.
That leads to tiered autonomy. Actions that are low-stakes and easy to undo can run automatically. Actions that are consequential or hard to reverse pause for a person to approve them. Anthropic's agent permission settings work this way: the guidance is to configure always_ask on the tools you would not let that end user run without review. A deployment that lets an agent handle routine, reversible work while requiring approval for anything affecting production or customers is the pattern the documentation describes. Giving an agent full autonomy over irreversible actions and saying "we can read the logs afterwards" is not.
Checkpoint 3 of 4· Exam question
A cybersecurity vendor asks Claude to reverse-engineer a competitor's software and write exploit code for a previously undisclosed vulnerability, intending to sell the exploit on a private marketplace. Is this an appropriate use of Claude?
Correct answer: B — No, because the Usage Policy bans creating exploits and malware meant for unauthorized offensive use
- A. Incorrect. Security research is not blanket-permitted; the Usage Policy specifically prohibits developing exploits and malware for unauthorized offensive use, regardless of the requester's field.
- B. Correct. Anthropic's Usage Policy prohibits using Claude to create malware, exploits, or other tools designed to exploit vulnerabilities without authorization, which is exactly what this scenario describes.
- C. Incorrect. Labeling content as AI-generated does not make a prohibited use case permissible; disclosure is not a substitute for complying with the underlying restriction.
- D. Incorrect. The core problem is the unauthorized creation of offensive exploit code, not a copyright dispute over the resale of the finished exploit.
Checkpoint 4 of 4· Check yourself
Which agent deployment best matches Anthropic's guidance on oversight?
The guidance is to put approval gates on the tools you wouldn't let run without review, and to let low-risk work proceed. Archived logs alone come too late. Approving everything defeats the purpose of using an agent. Gating testing but not production is backwards.
“Configure always_ask on the tools you would not let that end user run without review.”Source: platform.claude.com
Sources4
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Claude can reliably tell whether an image is real or AI-generated, so it can be used to screen out fakes.Why is that wrong?
The vision documentation says Claude cannot make this determination and may be wrong if asked. It also says not to rely on it to detect fake or synthetic images.
Covered in Documented limits that change the answer
2.Because Claude's actions are logged, an agent can safely handle sensitive tasks or tasks needing precision without anyone approving them first.Why is that wrong?
Logs are for review, but the documentation also says such tasks should not run without human oversight. Precision-critical or sensitive work needs a human involved.
Covered in Agents: match autonomy to how easily mistakes can be undone
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“why the model is excellent at well-worn paths like summarizing or reformatting”
↩︎ The core test: can an error be caught before it matters?“strong on mainstream topics and popular languages, weaker on anything rare, recent, niche, or contested”
↩︎ Documented limits that change the answer - 2.
“But we believe safety is a shared responsibility. Our features are not failsafe, and committed partners are a second line of defense.”
↩︎ The core test: can an error be caught before it matters?“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
↩︎ Key concept“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
↩︎ Checkpoint - 3.
“Do not use Claude for tasks requiring perfect precision or sensitive image analysis without human oversight.”
↩︎ Documented limits that change the answer“it is not designed to interpret complex diagnostic scans such as CTs or MRIs”
↩︎ Documented limits that change the answer“Do not rely on it to detect fake or synthetic images.”
↩︎ Exam trap 1“People identification: Claude cannot be used to name people in images and refuses to do so.”
↩︎ Prediction“Claude cannot determine whether an image is AI-generated and might be incorrect if asked.”
↩︎ Checkpoint - 4.
“Focus on use cases where speed isn't critical (for example, background information gathering, automated software testing) in trusted environments.”
↩︎ Agents: match autonomy to how easily mistakes can be undone“Always carefully review and verify Claude's computer use actions and logs.”
↩︎ Agents: match autonomy to how easily mistakes can be undone“Do not use Claude for tasks requiring perfect precision or sensitive user information without human oversight.”
↩︎ Exam trap 2
Also cited
“Configure always_ask on the tools you would not let that end user run without review.”
↩︎ Checkpoint