Free Anthropic CCDV-F Practice Test Questions and Answers (2026)

Non-functional requirements (NFRs) define system attributes such as performance, security, reliability, and usability, rather than specific behaviors or functions. They dictate how the system should perform rather than what it does. Option A specifies a performance metric (latency at the 95th percentile), and Option C specifies a strict security and data isolation constraint (tenant data privacy). Both are classic examples of non-functional requirements essential for enterprise AI deployments.

B. This describes a specific action or feature the system must perform, making it a functional requirement.

D. Creating a follow-up task in a CRM is a specific system behavior and integration, which is a functional requirement.

E. This is a business objective or Return on Investment (ROI) target, not a system-level non-functional requirement.

1. ISO/IEC/IEEE 29148:2018 - Systems and software engineering — Life cycle processes — Requirements engineering, Section 9.4.2.3 (Non-functional requirements including performance and security).

2. Anthropic Official Documentation - "Security and Privacy" (docs.anthropic.com). Section: Data retention and isolation.

Claude exhibits a "recency bias" (or attention mechanism characteristic) where it pays closer attention to instructions placed at the very end of a long prompt. When processing large contexts like a 20-page contract, placing the specific question or instructions after the long document ensures the model retains the immediate task in its working memory before generating the response. This structural adjustment significantly improves accuracy and adherence to instructions in long-context scenarios.

A: Splitting the request destroys the document's global context, preventing the model from synthesizing information across different pages.

C: Removing the contract forces the model to hallucinate or rely on general knowledge, failing the specific document QA task.

D: Deleting every second paragraph arbitrarily destroys critical information and context necessary for accurate legal analysis.

1. Anthropic Claude Documentation, "Prompt Engineering: Long Context Tips," Section: Put instructions at the end.

2. Liu, N. F., et al. (2023). "Lost in the Middle: How Language Models Use Long Contexts." Transactions of the Association for Computational Linguistics. DOI: https://doi.org/10.1162/tacla00638

For a small startup lacking a dedicated infrastructure team, utilizing a managed deployment (such as consuming the Anthropic API directly or using a managed agent platform) is the most appropriate model. This approach abstracts away the complexities of server provisioning, scaling, and runtime patching. The primary tradeoff is relinquishing granular control over the underlying infrastructure in exchange for a significantly faster time-to-market and reduced operational overhead, allowing the team to focus purely on agent logic, prompt engineering, and application development.

A: Self-hosting requires significant infrastructure management, which slows down the launch process rather than speeding it up.

C: Managed deployments and APIs are fully capable of powering agentic workflows; self-hosting is not a strict requirement.

D: Small teams can responsibly deploy agents using managed services and robust SDKs without needing a dedicated platform team.

1. Anthropic Documentation, "Productionizing your application" - Architecture and Deployment Models.

2. Anthropic Documentation, "Building Effective Agents" - Implementation and Infrastructure.

Disciplined prompt engineering requires an empirical, iterative approach. When multiple variables (instructions, examples, formatting) are changed simultaneously, it becomes impossible to isolate which modification caused the performance shift. Best practices dictate changing one variable at a time and evaluating the result against a consistent, high-quality evaluation dataset (evals). This isolates the impact of each change, ensuring that improvements are measurable, reproducible, and not masking regressions in other areas.

A: Making large combined changes prevents isolating the root cause of improvements or regressions, hindering systematic optimization.

B: Reverting everything discards the partial improvements achieved; iterative refinement should build upon successes, not abandon them entirely.

C: LLMs lack the introspective capability to accurately determine which specific prompt change caused a shift in their own output quality.

1. Anthropic Claude Documentation, "Prompt Engineering: Iterative Refinement," Section: Establish a baseline and iterate.

2. Anthropic Claude Documentation, "Evaluating Prompts," Section: Building an evaluation suite.

Exposing API keys in client-side code (like JavaScript bundles) allows malicious actors to easily extract and abuse them, leading to unauthorized usage, data breaches, and severe billing charges. The only secure remediation is to immediately revoke (rotate) the compromised key and redesign the application architecture. API calls must be routed through a secure, server-side backend where the key is safely injected at runtime via a secrets manager or environment variables. Obfuscation or rate-limiting are insufficient as they do not prevent extraction or quota exhaustion.

B. Obfuscation is easily reversed by determined attackers using automated de-obfuscation tools and does not secure the underlying key.

C. Renaming variables is security through obscurity and provides zero protection against automated scrapers or manual inspection.

D. Rate limiting mitigates the financial impact but leaves the key exposed for attackers to exhaust the quota, causing denial of service.

1. Anthropic API Fundamentals Documentation, "Security and Authentication" section (stating keys must be kept secret and never shared in client-side code).

2. OWASP Top 10 API Security Risks (2023), API2:2023 Broken Authentication (detailing the risks of exposed credentials in client applications).

3. MIT OpenCourseWare, 6.858 Computer Systems Security, Lecture on Web Security (emphasizing server-side secrets management).

Relying solely on a single system prompt instruction is insufficient for a production-grade safety strategy. Large Language Models are susceptible to prompt injection and jailbreak attacks that can easily bypass basic system instructions. A robust security posture requires a defense-in-depth approach. This includes comprehensive prompt engineering, strict input validation, output filtering (such as using moderation models or classifiers), applying the principle of least privilege for any integrated tools, and continuous monitoring of user interactions to detect and mitigate abuse.

A. System prompts are probabilistic and can be overridden or bypassed using sophisticated prompt injection or jailbreak techniques.

B. Placement at the top of the prompt improves attention but does not prevent adversarial inputs from overriding the instructions.

C. Relying solely on base model training is unsafe; production applications require explicit, application-level guardrails to ensure compliance and safety.

1. Anthropic Documentation, "Safety and Security" and "Mitigating Jailbreaks" guidelines (advocating for layered defenses).

2. OWASP Top 10 for Large Language Model Applications, LLM01: Prompt Injection (detailing the necessity of defense-in-depth).

3. Perez et al., "Red Teaming Language Models to Reduce Harms," Anthropic Research, 2022. DOI: https://doi.org/10.48550/arXiv.2209.07858.

Claude 3 models possess native multimodal (vision) capabilities and can process images directly. To analyze an image, such as a scanned paper receipt, the image must be passed to the Messages API as an image content block (typically base64-encoded) alongside text instructions detailing the specific extraction task. This allows Claude to perform multimodal reasoning and extract the merchant, date, and total directly from the visual data without needing a separate, legacy OCR pipeline.

B. Relying on the user to write a plain-text description defeats the entire purpose of automated data extraction from the receipt image.

C. The Claude API is a REST interface and cannot access local file paths on the client's machine; image data must be transmitted directly in the payload.

D. Claude 3 models have native vision capabilities, eliminating the strict requirement for a separate, preliminary OCR service to extract text before processing.

1. Anthropic Official Documentation, "Vision", Section: How to use Vision (Image content blocks), 2024. (https://docs.anthropic.com/en/docs/build-with-claude/vision)

2. Anthropic API Reference, "Messages", Section: Content Block (Image), 2024.

Large Language Models (LLMs) are non-deterministic, and new model versions—even those with higher benchmark scores—can exhibit subtle behavioral shifts. These changes can affect prompt adherence, output formatting, tone, and tool use (function calling). Therefore, newer models are not strict supersets of older ones. Before migrating a production application, it is a mandatory engineering practice to run the application's existing evaluation suite (evals) against the new model to identify and mitigate any breaking changes or regressions in specific use cases.

A. Newer models are not strict supersets; underlying weight changes can alter specific behaviors, potentially breaking brittle prompts.

C. Behavior is not standardized; prompt responses can vary significantly between model versions, requiring re-validation.

D. Waiting a year is unnecessary and prevents leveraging cost and performance improvements; proper evaluation suites allow for rapid, safe adoption.

1. Anthropic Official Documentation, "Model upgrades and migrations" (Section: Evaluating new models).

2. Anthropic Official Documentation, "Prompt engineering interactive tutorial" (Section: Testing and Evals).

This scenario describes a classic Prompt Injection attack. The most effective mitigation is a defense-in-depth architectural approach rather than relying solely on prompt instructions. Inbound user data must be treated as untrusted input. Tools should operate on the principle of least privilege, and sensitive actions (like forwarding emails to external addresses) must be gated behind strict guardrails, human-in-the-loop approvals, or hardcoded domain whitelists. System prompts alone are insufficient because attackers can often bypass them with sophisticated jailbreaks.

A: System prompt instructions are easily bypassed by sophisticated prompt injections; LLMs cannot reliably distinguish between system instructions and injected user commands.

B: Raising the temperature increases output randomness, which does not prevent prompt injection and makes the system's behavior generally less predictable.

D: Blocking the word "ignore" is a brittle, easily bypassed filter (e.g., attackers can use "disregard") and causes false positives in normal emails.

1. Anthropic Claude Documentation, "Security and Compliance: Mitigating Prompt Injection," Section: Defense in depth.

2. OWASP Foundation, "OWASP Top 10 for Large Language Model Applications v1.1," Section: LLM01: Prompt Injection.

Anthropic partners with major cloud providers to offer Claude models through managed services like Amazon Bedrock and Google Cloud Vertex AI. This allows enterprises to consume Claude's capabilities while strictly adhering to their existing cloud provider agreements. By using these third-party platforms, organizations can utilize established billing, governance, VPC network controls, and compliance frameworks without needing a direct first-party relationship or exception with Anthropic.

A. Claude is not restricted to the first-party API; it is officially supported and distributed through major enterprise cloud partners like AWS and Google Cloud.

C. Anthropic typically releases new Claude models on partner platforms like Amazon Bedrock and Google Vertex AI simultaneously or very shortly after the first-party release.

D. Prompt engineering and evaluations are fundamental to LLM application development and remain strictly necessary regardless of the hosting infrastructure or cloud provider.

1. Anthropic Official Documentation, "Intro to Claude", Section: Accessing Claude (Cloud Partners), 2024. (https://docs.anthropic.com/en/docs/intro-to-claude)

2. Amazon Web Services Documentation, "Amazon Bedrock - Supported Foundation Models", Section: Anthropic Claude, 2024.