Governance, Risk, and Responsible Use

Knowledge of

• Anthropic's Usage Policy (AUP) as the outer boundary of acceptable use, with heightened requirements for high-risk and agentic scenarios
• That organizations layer their own AI policies on top of the AUP, and both must be respected
• What counts as appropriate use: language/knowledge work with human review such as drafting, summarizing, analysis, research, brainstorming, and process support
• What counts as inappropriate or restricted use: violating the AUP, breaching organizational policy, mishandling regulated data, or placing unverified AI output into high-stakes decisions without human oversight
• The four delegation criteria used to screen a whole use case (not just a single request): reversibility, consequence of error, need for human creativity/empathy, and accountability
• That the four criteria interact rather than acting as independent gates, and that naming the load-bearing criterion is what makes a classification defensible
• The three use-case classifications -- fully appropriate, appropriate with human review, and inappropriate -- and what distinguishes them
• That 'appropriate with human review' requires a gate defined in WHO/WHAT/WHEN form, and that a restated label ('keep a human in the loop') is not itself a gate

Skills in

• Recognizing when a request falls outside the AUP or organizational policy rather than assuming it's fine because it wasn't explicitly forbidden
• Finding the compliant way to accomplish a borderline task (e.g., anonymizing data first) instead of abandoning the task or ignoring the rule
• Rejecting the 'it's just internal' rationalization for bypassing data and governance rules
• Identifying the AUP plus organizational policy as the authoritative source for what is appropriate use
• Screening a whole use case against the four delegation criteria and identifying which one is load-bearing for the specific scenario
• Classifying a use case as fully appropriate, appropriate with human review, or inappropriate, with a stated rationale tied to the load-bearing criterion
• Writing a defined human-review gate in WHO/WHAT/WHEN form rather than a vague 'human in the loop' restatement
• Recognizing when a high-consequence task becomes appropriate specifically because a defined gate restores accountability and reversibility

Knowledge of

• The data classification spectrum: public, internal, confidential, and regulated (personal data, financial records, health information)
• Minimization and anonymization as the practice of removing or masking personal identifiers before sharing data with Claude when policy restricts it
• That data-protection laws (e.g., GDPR/CCPA-style regimes) constrain what data may be processed and how, and usage must align with them and with contractual obligations
• Why 'don't retain this' instructions to the model are not a substitute for anonymization or a real policy control
• Incognito mode as a feature-level privacy control that excludes a standalone chat from Memory and from your visible history, and why it never substitutes for the classification question on regulated data

Skills in

• Classifying a given dataset or request along the public/internal/confidential/regulated spectrum
• Removing or masking personal identifiers (names, account numbers, PII) before sharing regulated data with Claude
• Recognizing when a task can proceed compliantly after minimization rather than being blocked outright
• Aligning data-handling choices with applicable data-protection regulation and contractual obligations
• Using Incognito mode appropriately for sensitive or exploratory standalone chats, without treating it as clearance to process regulated data that hasn't been approved for that entry point

Knowledge of

• That organizations layer their own governance on top of the AUP: approved tools and plans, allowed data types, required review steps, and escalation paths
• That responsible use means knowing and following those standards, not improvising, and escalating to the right owner when policy is unclear or a case is genuinely novel
• Prompt injection and manipulated inputs as a security-adjacent governance risk: instructions hidden inside a document or web page that try to redirect the model
• Why external content (uploaded documents, fetched web pages) must be treated as untrusted, with outputs validated rather than embedded instructions followed
• Skill trust evaluation as a repeatable governance check -- source, reach, and appropriateness -- before enabling a Skill, and the enable/escalate/decline outcomes that follow from it
• The least-privilege principle generalized beyond Skills to any feature, connector, or tool: hand over only the access a job actually requires

Skills in

• Identifying the organization's approved tools, plans, allowed data types, and required review steps for a given task
• Escalating to the appropriate policy owner when governance is unclear or the use case is novel, instead of guessing or improvising
• Recognizing hidden or embedded instructions in an uploaded document or fetched page as a prompt-injection attempt rather than a legitimate clarification
• Treating external content as untrusted and validating outputs before acting on them
• Evaluating a Skill's source, reach, and appropriateness before enabling it, and choosing to enable, escalate, or decline based on that evaluation
• Applying least-privilege reasoning to any feature/connector/tool decision, not just Skills

Knowledge of

• Bias and fairness: AI output can reflect or amplify bias and requires deliberate review, especially in decisions about people
• Transparency: being honest about when and how AI was used, per organizational norms
• Accountability: the human stays responsible for decisions and published output; AI assists, it doesn't absolve
• The AI Fluency Diligence competency: using AI effectively, ethically, and safely, with the user owning the outcome
• The need for human oversight where outcomes materially affect people (high-stakes use)

Skills in

• Reviewing AI output for skewed framing or unfair treatment, especially in decisions about people
• Disclosing AI use honestly per organizational norms rather than presenting AI-assisted work as purely human-authored without disclosure
• Taking ownership of AI-assisted decisions and output rather than treating the AI's involvement as an accountability shield
• Keeping a human in the loop for high-stakes use where outcomes materially affect people
• Applying the AI Fluency Diligence competency: effective, ethical, and safe use with the human owning the outcome