What you will be able to do
- Judge whether a task suits Claude's strengths, and say where its known limits make it a poor fit
- Tell a use where Claude's output informs a person's decision apart from one where it acts on people directly
- Check a proposal against the Usage Policy's prohibited uses, then apply the human-review and disclosure requirements for high-risk areas
- Vet a third-party Skill or plugin before it touches real data
- Run an evaluation in order and record the result as appropriate, appropriate with safeguards, or inappropriate, with a written rationale
Key concept
Delegating implementation, not judgment — A use case is appropriate when Claude does the producing work and an accountable person still makes the decisions that affect people. The safeguards a use case needs depend on how much of that judgment it tries to hand over, and to whom the output goes.
1.Matching the task to what Claude does well
Evaluating a use case starts with an ordinary question: is this the kind of work a language model does reliably? Claude produces text by predicting it one token at a time. That makes it strong on well-worn tasks such as summarising, reformatting, drafting and explaining something several ways, and it can use tools. Tasks that depend on recent facts, rare or niche knowledge, or long chains of exact reasoning are weaker. Those are where fabrication and stale information concentrate.
Weak spots don't automatically rule a task out. They tell you what has to be checked. A first draft of a policy memo plays to Claude's strengths, and a quick read catches its errors. A list of cited case law plays to its weaknesses, because every citation has to be confirmed before anyone relies on it. That is where hallucination checks come in. Neither task is forbidden, but the second carries a verification cost that belongs in the evaluation.
| Task feature | Leans toward | What it means for the evaluation |
|---|---|---|
| Summarising, reformatting, explaining to different audiences | Strength (versatility, fluency) | Usually suitable; review proportionate to the stakes |
| Recent events, new library versions, recent rule changes | Limit (knowledge cutoff) | Supply current sources or verify against them |
| Specific citations, packages, references | Limit (hallucination) | Spot-check every reference before relying on it |
| Multi-step complex reasoning | Limit (unreliable complex reasoning) | A qualified person checks the reasoning, not just the conclusion |
Checkpoint 1 of 8· Match them up
A team proposes four Claude tasks. Match each task to the capability or limit it relies on and what the evaluation must then require.
Tap a term, then the definition that fits it.
Claude is strong at familiar tasks like summarising for different audiences. Recent facts, specific references and long chains of reasoning hit its known limits. Those limits don't rule a task out, but each one adds its own verification step to the evaluation.
“Current limits include knowledge cutoffs, hallucinations, and unreliable complex reasoning.”Source: academy.claude.com
2.Delegation limits: informing a decision versus making it
The most useful line in appropriateness review is between output that informs a person and output that acts on people directly. Delegation works by breaking a problem into parts and deciding what role AI plays in each. There are three modes: Automation (AI does it, you check), Augmentation (you and AI work on it together), and Agency (AI acts with latitude inside boundaries you set). Every mode keeps a person somewhere in the loop: checking, collaborating, or setting the boundaries.
Compare two proposals. A paralegal asks Claude to flag unusual clauses in a 40-page contract, then reads every flagged clause and decides which ones go to the attorney. Claude speeds up the reading, and the judgment stays with her. In the other proposal, an insurer has Claude deny weak claims and send the denial letters without anyone looking at them. Now Claude's output is the decision, and it lands on a claimant. Grades published straight to transcripts and sales emails sent overnight unread have the same shape. In each of these, the fix is a review step before release. Someone accountable reads the output and decides whether it goes out.
Building more autonomy into a system doesn't change who answers for it. The best results pair human judgment, creativity and oversight with AI's speed and scale. A proposal that removes the person rarely needs a better prompt. It needs that person put back at the point where the output reaches someone. How deep that review should go is covered in review depth and escalation.
Keep the drafting and drop the automatic send. Claude writes the follow-ups into a queue, and a rep reads each one and sends it, or edits it first. The time saving comes from drafting, and the risk to tone and accuracy comes from sending unread.
Checkpoint 2 of 8· Check yourself
Under an HR proposal, Claude screens incoming job applications and emails rejection notices to the weaker applicants overnight, and nobody reads them first. Which change makes this use case appropriate?
Claude's output is acting on applicants directly, so the fix is to put an accountable person back at the point of release, not to write a better prompt. A sample audit after the emails go out comes too late. Dropping Claude gives up the drafting and screening work it can do safely.
“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”Source: support.claude.com
3.The Usage Policy: prohibited uses first, then high-risk requirements
Anthropic's Usage Policy asks two separate questions, and the order matters. The first is whether the use is prohibited. If it is, no safeguard rescues it. The policy applies to agents and agentic features just as it does to chat. Its agentic examples include tracking people's activities or movements without notification or consent, building profiles from protected attributes, using facial recognition, phishing, impersonating someone without consent, and coordinated inauthentic behaviour. They also cover installing malware, using someone's stored credentials without authorisation, and unauthorised or fraudulent financial transactions. A proposal that falls in one of these categories is classified inappropriate. Adding a reviewer doesn't change that.
The second question applies only to permitted uses: is the use high-risk? The policy names high-risk use cases in legal, healthcare, insurance and finance, among other areas. A high-risk use is allowed, but it comes with two requirements: human-in-the-loop measures and direct disclosure to the people affected. In practice, a qualified professional reviews the content before it reaches consumers, and anyone on the receiving end of an external-facing product is told they are dealing with an AI system.
Disclosure is proactive, and it concerns the fact that AI was used. The bank tells recipients that AI was involved in the recommendation. It doesn't wait to be asked, and it doesn't skip the notice because a credit officer reviewed the result. What the person needs to know is that an AI system was involved. Crediting a particular vendor or model is a separate matter, and a notice that AI helped produce the recommendation meets the requirement. The wider ethics of transparency and accountability are covered in policy, ethics and diligence.
An inappropriate classification needs its rationale written down, not just the verdict. The record names the prohibited category the proposal falls into, states why no safeguard changes the outcome, and says what would have to be different before the proposal could be reviewed again. 'Rejected: violates policy' tells the requester nothing. 'Inappropriate: the agent builds profiles from protected attributes, a prohibited agentic use, and human review does not alter that' can be audited, reused for the next similar proposal, and challenged if the facts turn out to differ.
Checkpoint 3 of 8· Check yourself
Classify this proposal. A recruiting lead launches a Claude agent that pulls applicants' public social media posts and builds a profile of each one, sorted by inferred religion and health conditions. A senior HR officer reviews every profile before it is used, and the lead argues that the human launch and the review keep it within policy.
The first question is whether the use is prohibited. Profiling people by protected attributes is a prohibited agentic use, and no safeguard rescues a prohibited use. The high-risk requirements of review and disclosure only apply to uses that are permitted to begin with.
“create profiles based on individuals’ protected attributes, sensitive characteristics, or personal circumstances”Source: support.claude.com
Checkpoint 4 of 8· Exam question
A bank compliance officer asks whether employees who receive a credit recommendation produced with help from Claude must be told that AI was used, and whether the bank must name Claude or Anthropic. What does the Anthropic Usage Policy require?
Correct answer: A — Tell them AI was used in the decision, but identifying Anthropic or Claude by name is not required.
- A. Correct. High-risk recommendations require disclosure that AI was used to the people affected, without naming the vendor or model.
- B. Incorrect. Disclosure applies to high-risk recommendations even when a qualified human reviews them.
- C. Incorrect. The policy does not make naming Anthropic or Claude mandatory.
- D. Incorrect. Credit is a high-risk area, and the notice is required rather than left to the customer to request.
4.Trusting third-party Skills, plugins and connectors
Skills, plugins and connectors are part of the use case, because they decide what Claude can read, run and send. A plugin can run code on your computer with access to your files and shell. Its Skills enter Claude's context as instructions, so they shape what Claude does with every tool it already has. Anthropic doesn't control what third-party plugins contain and can't verify that they will behave as intended or stay unchanged. That makes the trust decision yours, and it should come before any real data is involved.
Start with who publishes it and where it comes from. Claude Code sorts plugin marketplaces into three tiers. Even inside a tier, check the marketplace's actual source. Claude Code itself stops loading a marketplace that uses an official or community name when its source sits outside github.com/anthropics/.
| Tier | Which marketplaces |
|---|---|
| Official | The official marketplace names, such as claude-plugins-official |
| Community | claude-community, claude-plugins-community, and healthcare |
| Third-party | Every other marketplace |
A tier labels the marketplace, not the contents of any plugin in it, so no tier removes the need for review. Official and community names are protected by the source check above: a marketplace using one of those names from outside github.com/anthropics/ is refused, which tells you the name is genuine. Third-party covers every other marketplace, and nothing vouches for its name, so treat its plugins as unreviewed code and also ask whether you trust the marketplace's owner. If you don't, remove the marketplace, which uninstalls every plugin you took from it. Claude Code's own reporting follows the same split: official and community plugins appear by name, and everything else is logged only as third-party.
Next, read everything: SKILL.md, every file it references, and any bundled scripts. Then work out what the Skill can reach. Some features raise the risk sharply, and two harmless-looking features can be dangerous together. A Skill that reads files and can also make network calls is a possible route for your data to leave. Add an external connector and its reach grows further.
That last item needs a word of explanation. MCP (Model Context Protocol) servers are the external connectors that give Claude extra tools. When a plugin declares one, Claude Code connects to it and hands Claude its tools, and a stdio MCP server runs as a process started on your machine. A Skill that names an MCP tool in the form ServerName:tool_name therefore steers Claude toward systems and data that sit outside the Skill's own files, and the admin decision about which connectors are enabled is also the decision about which data the agent can reach.
| Risk indicator | What to look for | Concern level |
|---|---|---|
| Code execution | Scripts in the Skill directory (*.py, *.sh, *.js) | High: scripts run with full environment access |
| Instruction manipulation | Directives to ignore safety rules, hide actions from users, or alter Claude's behavior conditionally | High: can bypass security controls |
| MCP server references | Instructions referencing MCP tools (ServerName:tool_name) | High: extends access beyond the Skill itself |
| Network access patterns | URLs, API endpoints, fetch, curl, or requests calls | High: potential data exfiltration vector |
| Hardcoded credentials | API keys, tokens, or passwords in Skill files or scripts | High: secrets exposed in Git history and context window |
| Filesystem access scope | Paths outside the Skill directory, broad glob patterns, path traversal (../) | Medium: might access unintended data |
Only then try the Skill. Run its scripts in a sandbox on non-sensitive sample files and confirm the results match what the Skill claims to do. If it passes, roll it out with its purpose, owner and version documented, so the team knows what it is for. The person who wrote the Skill shouldn't be the one who reviews it.
A review doesn't stay valid forever. When auto-update is on for a marketplace, its plugins update in the background, and the files you reviewed can change on disk. On Claude Enterprise, automated Skill scanning blocks Skills that fail or haven't finished scanning. It complements the manual checklist but doesn't replace it, and it doesn't cover Skills uploaded through the Claude API.
Checkpoint 5 of 8· Check yourself
Your organisation runs Claude Enterprise with Skill and plugin security scanning turned on. A security lead suggests a new rule: any Skill that passes the scan can go live without a manual review, including the Skills your developers upload through the Claude API. How should you assess this rule?
Automated scanning complements the manual review checklist, so passing the scan doesn't remove the need for a human review. Scanning also doesn't cover Skills uploaded through the Claude API, so those depend entirely on the checklist.
“If scanning is available to your organization, turn it on. Scanning complements, but doesn’t replace, the review checklist.”Source: platform.claude.com
Checkpoint 6 of 8· Exam question
An office manager wants to enable a third-party Skill for her team. Which sequence of steps best prepares her to adopt it safely?
Correct answer: A — Identify the Skill's publisher and download source → Read the Skill's instructions and bundled scripts → Check the files and connections it reaches → Trial the Skill on non-sensitive sample files → Roll out to team with a usage note
- A. Establishing who published the Skill comes first, then reading its contents reveals what it does, which lets its reach be compared with the task. Only then is it trialled on harmless sample files before the team gets it.
- B. Reading the instructions and scripts before identifying the publisher means the contents are judged without knowing who wrote them or where they came from. Provenance is the baseline for deciding how much trust the contents deserve, so it should come first.
- C. Checking what the Skill can reach before reading its instructions and scripts means the reach is assessed without knowing what the Skill actually does. Its behavior has to be known before its access can be compared with what the task needs.
- D. Trialling the Skill on sample files before checking what files, data and connections it can reach means the test runs without knowing its exposure. Even sample files could be touched or sent somewhere unexpected, so reach should be checked before any run.
5.Running the evaluation and recording the classification
An appropriateness review is a sequence, and each step depends on the one before it. A classification decided before the earlier checks finish is a guess. Safety is a shared responsibility here: Anthropic's own safeguards aren't failsafe, and the team deploying Claude acts as a second line of defence. The review is where that team does its part.
Run the steps in this order:
- Describe the task, the data involved and the people the output affects. Every later check depends on this description.
- Check the proposal against the Usage Policy's prohibited uses. A prohibited use stops the review here, because no safeguard fixes it.
- Decide whether the output falls into a high-risk area, such as legal, healthcare, insurance or finance, that needs human review.
- Define the safeguards: who reviews, what gets verified, and what disclosure is given.
- Record the classification and the reasoning behind it in the review log.
| Classification | Typical pattern | Example |
|---|---|---|
| Appropriate | Permitted, low-stakes, and a person uses the output to inform their own work | A paralegal has Claude flag contract clauses, then reads each flagged clause herself |
| Appropriate with safeguards | Permitted but high-risk, or output reaches people; becomes acceptable once qualified review and disclosure are in place | Claude drafts credit recommendations; an officer reviews each one and recipients are told AI was used |
| Inappropriate | A prohibited use, or a design that removes the accountable person and won't put them back | An agent tracks individuals' movements without notification or consent |
The written rationale is what makes the classification defensible. 'Appropriate with safeguards' means little on its own. 'Credit is a high-risk area; an officer reviews each recommendation before release; recipients are told AI was used' tells a later reviewer exactly what has to stay true for the classification to hold. If one of those conditions disappears, for example the officer stops reviewing, the use case has to be reviewed again. Sensitive data questions raised in step 1 are handled in data protection in practice.
Checkpoint 7 of 8· Match them up
Three proposals reach step 5 of the review, where the classification is written down. Match each proposal to the classification and rationale that belong in the review log.
Tap a term, then the definition that fits it.
The review checks for prohibited uses before anything else. Using stored credentials without authorisation is prohibited, so a reviewer can't rescue it. A permitted healthcare output that reaches patients is high-risk and needs both review and disclosure. Output that only informs the analyst's own work is appropriate as it stands.
“Access or modify another person's account using their stored credentials without authorization”Source: support.claude.com
Checkpoint 8 of 8· Exam question
A project owner evaluates a new Claude use case for the team. The owner must settle whether the use is permitted before judging how risky it is, and may record the classification only once every earlier check is finished. Which sequence of evaluation steps is correct?
Correct answer: D — Describe the task, data and affected people → Check against prohibited uses in the policy → Assess whether output needs human review → Define reviewer, verification and disclosure safeguards → Record classification and rationale in the log
- A. Assessing the high-risk area before checking prohibited uses swaps two adjacent steps. A proposal that is prohibited outright would get a risk review and safeguards it should never have needed, because the prohibition ends the evaluation.
- B. Defining safeguards before assessing the high-risk area sets reviewer, verification and disclosure requirements without knowing the risk they are meant to address. The safeguards would be guesses and might miss the need for qualified human review.
- C. Checking prohibited uses before describing the task means the proposal is screened without a clear picture of the task, data and affected people. Recording the log before defining safeguards also leaves the decision record incomplete, since it would not capture the safeguards.
- D. The task must be described first, prohibited uses are checked next because they end the evaluation, the high-risk assessment follows for permitted uses, safeguards are derived from that risk, and the classification is recorded last once all checks are done.
Sources
(9)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://academy.claude.com/courses/ai-fluency-for-builders/ai-capabilities-and-limitationsOfficial docs
“Current limits include knowledge cutoffs, hallucinations, and unreliable complex reasoning.”
↩︎ Matching the task to what Claude does well“The best applications pair your judgment, creativity, and oversight with AI’s speed and scale.”
↩︎ Delegation limits: informing a decision versus making it - 2.
“why the model is excellent at well-worn paths like summarizing or reformatting”
↩︎ Matching the task to what Claude does well - 3.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“Automation (AI does it, you check), Augmentation (you and AI work it together), or Agency (AI operates with latitude inside boundaries you set)”
↩︎ Delegation limits: informing a decision versus making it“Explain why delegating implementation is safe and delegating judgment is not”
↩︎ Key concept - 4.https://support.claude.com/en/articles/12005017-using-agents-according-to-our-usage-policyOfficial docs
“Monitor or track individuals’ online activities, behaviors, or movements without notification or consent”
↩︎ The Usage Policy: prohibited uses first, then high-risk requirements“Engage in unauthorized, illegal, or fraudulent financial transactions (such as brokerage or investment advisory activities) or payment processing”
↩︎ The Usage Policy: prohibited uses first, then high-risk requirements“All uses of agents and agentic features must continue to adhere to Anthropic’s Usage Policy.”
↩︎ Exam trap 2“create profiles based on individuals’ protected attributes, sensitive characteristics, or personal circumstances”
↩︎ Checkpoint“Access or modify another person's account using their stored credentials without authorization”
↩︎ Checkpoint - 5.
“including high-risk use cases around legal, healthcare, insurance, finance and other areas”
↩︎ The Usage Policy: prohibited uses first, then high-risk requirements“When Software is providing a high-risk use case, Developers must comply with the human-in-the-loop measures and provide direct disclosure to consumers”
↩︎ Exam trap 1“When Software is providing a high-risk use case, Developers must comply with the human-in-the-loop measures and provide direct disclosure to consumers”
↩︎ Prediction - 6.
“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
↩︎ The Usage Policy: prohibited uses first, then high-risk requirements“For external-facing products, disclose to your users that they are interacting with an AI system.”
↩︎ The Usage Policy: prohibited uses first, then high-risk requirements“safety is a shared responsibility. Our features are not failsafe, and committed partners are a second line of defense.”
↩︎ Running the evaluation and recording the classification“For sensitive information and decision making, have a qualified professional review content prior to dissemination to consumers.”
↩︎ Checkpoint - 7.https://code.claude.com/docs/en/plugins/securityOfficial docs
“Plugins can run code on your computer, with access to your files and shell.”
↩︎ Trusting third-party Skills, plugins and connectors“Skills, commands, and agents: these enter Claude’s context as instructions, so they influence what Claude does with the tools it already has.”
↩︎ Trusting third-party Skills, plugins and connectors“Make sure you trust a plugin before installing, updating, or using it.”
↩︎ Trusting third-party Skills, plugins and connectors“when a marketplace uses an official or community name but its source is outside github.com/anthropics/”
↩︎ Trusting third-party Skills, plugins and connectors“if you don’t trust the marketplace’s owner either, remove the marketplace too, which uninstalls every plugin you installed from it.”
↩︎ Trusting third-party Skills, plugins and connectors“Claude Code connects to the MCP servers an enabled plugin declares and gives Claude their tools.”
↩︎ Trusting third-party Skills, plugins and connectors“so the files you reviewed can change on disk”
↩︎ Exam trap 4 - 8.https://claude.com/blog/ciso-guide-to-agentic-aiSecondary source
“The admin decision about which connectors to turn on is also the decision about which data the agent can reach.”
↩︎ Trusting third-party Skills, plugins and connectors - 9.
“Read all Skill directory content. Review SKILL.md, all referenced markdown files, and any bundled scripts or resources.”
↩︎ Trusting third-party Skills, plugins and connectors“Consider the combined risk when a Skill uses both file-read and network tools together.”
↩︎ Trusting third-party Skills, plugins and connectors“Run scripts in a sandboxed environment and confirm outputs align with the Skill’s description.”
↩︎ Trusting third-party Skills, plugins and connectors“Establish separation of duties: Skill authors should not be their own reviewers.”
↩︎ Trusting third-party Skills, plugins and connectors“Document the Skill in your internal registry with purpose, owner, and version.”
↩︎ Running the evaluation and recording the classification“Scanning complements, but doesn’t replace, the review checklist.”
↩︎ Exam trap 3“If scanning is available to your organization, turn it on. Scanning complements, but doesn’t replace, the review checklist.”
↩︎ Checkpoint