Crack the Claude Exam

Everything you need to know to earn the Anthropic Claude AI certification.

About the Anthropic Claude certification

Anthropic offers certification programs that validate proficiency with Claude AI. The certification demonstrates to employers that you can use Claude effectively, responsibly, and at an advanced level. As of 2026, certifications cover:

• Claude Fundamentals — core features, plans, interface, basic prompting
• Claude for Developers — API, MCP, system prompts, tools, cost management
• Responsible AI with Claude — safety, ethics, Constitutional AI, Anthropic's policies

Check anthropic.com for current exam availability and registration — the program evolves frequently.

Core knowledge areas (all exam tracks)

1. Claude's model families and capabilities

• Opus (most capable), Sonnet (balanced), Haiku (fastest) — when to use each
• Context window: up to 200K tokens across all current models
• Multimodal: text + images (not video)
• Training cutoff and knowledge limitations
• What Claude can and cannot do (policy boundaries)

2. Constitutional AI and Anthropic's approach to safety

This is on every exam track. Know:

• Constitutional AI (CAI) — Anthropic trains Claude using a set of principles (the "constitution"). Claude critiques and revises its own outputs against these principles. This is how Claude avoids harmful outputs without pure human feedback for every response.
• RLHF — Reinforcement Learning from Human Feedback. Used alongside CAI.
• The three core goals — Claude aims to be Helpful, Harmless, and Honest (the "3 H's")
• Absolute limits — things Claude will not do regardless of context (weapons of mass destruction, CSAM, undermining AI oversight)
• Hardcoded vs softcoded behaviors — some behaviors are fixed; others can be adjusted by operators (companies using the API) or users

3. Operators vs Users in the API

A critical concept for the developer track:

• Anthropic trains Claude and sets absolute limits via the model's training
• Operators are companies/developers using the API. They write the system prompt and can expand or restrict Claude's default behaviors within Anthropic's policies
• Users are the end-users chatting with Claude. They can adjust behaviors within what the operator allows
• The hierarchy: Anthropic > Operator > User

4. Effective prompting

• Be specific, direct, and provide context
• Use XML tags to structure complex prompts
• Chain-of-thought: ask Claude to think step by step
• Few-shot examples improve consistency
• Role prompting: give Claude a persona
• Format instructions: specify output structure
• Iterative refinement: build on previous outputs

5. API fundamentals (developer track)

• Messages API structure: system + messages array
• Model IDs: claude-opus-4-5, claude-sonnet-4-5, claude-haiku-4-5-20251001
• Token counting and cost estimation
• Streaming responses
• Tool use (function calling)
• Vision API (image inputs)
• Batch processing
• Rate limits and error handling

Practice questions

Here are the types of questions that appear on Anthropic exams:

Q: What is Constitutional AI?

A: Constitutional AI is Anthropic's training methodology where Claude is trained against a set of principles (the constitution) using a combination of supervised learning and RL. Crucially, Claude is trained to critique and revise its own outputs against these principles — a form of self-improvement that reduces reliance on human feedback for every potentially harmful output. The result is a model that follows values more reliably than pure RLHF.

Q: What's the difference between hardcoded and softcoded behaviors?

A: Hardcoded behaviors are things Claude will always or never do, regardless of any operator or user instructions — for example, Claude will never help create bioweapons. Softcoded behaviors are defaults that can be adjusted: operators might turn off Claude's safe messaging guidelines for a medical platform, or enable explicit content for an adult content platform that has verified user ages.

Q: An operator sets a system prompt saying "you are a customer service bot and should not discuss any topics unrelated to our product." A user asks Claude for help with a completely unrelated coding question. What should Claude do?

A: Claude should follow the operator's system prompt and decline to help with the coding question, politely noting that it can only discuss topics related to the company's product. The operator's instructions define the scope of the deployment. However, Claude should still provide basic safety information (emergency services) if a user is in danger — that's a hardcoded behavior that overrides operator restrictions.

Q: What is the maximum context window for Claude 4 Sonnet?

A: 200,000 tokens (roughly 150,000 words or 500 pages of text).

Q: What are the three levels of the operator/user hierarchy?

A: Anthropic (training) → Operators (system prompt) → Users (human turn). Anthropic's values are baked in at training and can't be overridden. Operators can customize behavior within Anthropic's policies. Users can further adjust within what operators allow.

Study plan