Anthropic Academy: CCA-F Exam Prep Guide

What is Anthropic Academy, and is it free?

Anthropic Academy is Anthropic's own free learning portal, available at academy.anthropic.com. It offers self-paced video and text modules covering AI fundamentals, the Claude API, prompt engineering, tool use, MCP, and Claude Code. Every course is free to start; you do not need a paid Claude plan or a partner account. Completion certificates are issued per course, but they are not the same credential as the Claude Certified Architect, Foundations (CCA-F) exam, which costs $99 per attempt and requires a passing scaled score of 720 out of 1000.

The distinction matters. Anthropic Academy is study material. The CCA-F is the professional credential. Treating them as interchangeable is the single most common misconception we see among candidates who arrive under-prepared.

How does Anthropic Academy map to the CCA-F exam blueprint?

The CCA-F is divided into five weighted domains. Before investing time in any Academy track, it helps to know how each one maps to exam weight.

Domain 1 alone accounts for 27% of the exam. If you spend most of your Academy time on the AI Fluency introductory track, you are optimising for the lowest-weighted material. We recommend inverting the default course order: start with Tool Use & Agents, then MCP, then Claude Code, and use AI Fluency only to fill conceptual gaps.

Which Academy tracks matter most for solution architects?

Solution architects preparing for the CCA-F should prioritise tracks in this order, based on domain weight and the scenario-based question format.

1. Tool Use & Agents (Domain 1, 27%). This track covers the agentic loop, multi-agent orchestration, subagent spawning, and error routing. These are the concepts that appear most frequently in exam scenarios. Pair the Academy content with our Agentic Architecture & Orchestration concept library for deeper coverage of topics such as hub-and-spoke architecture and coordinator dynamic subagent selection.

2. MCP Track (Domain 2, 18%). The Model Context Protocol track explains server scoping, tool discovery, and the isError flag pattern. The Academy modules are a solid introduction, but they stop short of the debugging scenarios the exam tests. You will want to supplement with hands-on MCP server work and our Tool Design & MCP Integration concepts.

3. Claude Code Track (Domain 3, 20%). This track covers the three-level configuration hierarchy, CLAUDE.md files, hooks, and CI integration. The exam tests configuration decisions at the project, user, and enterprise levels. The Academy content here is among the most directly exam-relevant of any track.

4. Prompt Engineering Track (Domain 4, 20%). The Academy covers chain-of-thought, few-shot examples, and structured output. The exam goes further, testing schema design, output validation loops, and the interaction between system prompts and tool descriptions. Our Prompt Engineering & Structured Output concepts cover the delta.

5. AI Fluency (Domains 4 and 5, background). Useful for candidates who are new to large language models, but it contributes little marginal value for experienced engineers. Treat it as optional.

6. Cloud Integration (AWS Bedrock, Google Cloud Vertex AI). The Academy touches on deploying Claude via Amazon Bedrock and Google Cloud Vertex AI. The CCA-F does not test cloud-provider-specific configuration deeply; it tests Claude's behaviour and architecture patterns. Do not over-index here.

Do Academy completion certificates count toward the CCA-F credential?

No. Academy completion certificates are issued by Anthropic's learning portal and confirm that you watched or read the course material. The CCA-F credential is issued only upon passing the proctored exam with a scaled score of 720 or above. The two are separate systems.

That said, Academy certificates are not worthless. They signal structured study to hiring managers and partners within the Claude Partner Network, which as of 3 June 2026 includes over 40,000 applicant firms. Listing them alongside the CCA-F on a CV is reasonable; listing them instead of the CCA-F is not.

What are the hardest concepts the Academy under-covers?

Based on the exam blueprint's 30 task statements, these are the areas where Academy content is thinnest relative to exam weight.

Prompt caching and token efficiency. The Academy explains what prompt caching is, but the exam tests when to apply it, what the cache invalidation rules are, and how to architect prompts to maximise cache hits across a multi-turn session. This falls under Context Management & Reliability.

Extended thinking. The Academy introduces extended thinking as a feature. The exam tests the trade-offs: latency, token cost, and when extended thinking genuinely improves accuracy versus when it adds cost without benefit.

Agentic loop anti-patterns. The exam consistently rewards candidates who can identify failure modes before they occur. The Academy describes the happy path; it does not dwell on agentic loop anti-patterns such as premature termination, infinite retry, or narrow decomposition failure.

Structured error metadata in MCP. The Academy covers the isError flag. The exam tests the full error taxonomy: access failure versus valid empty result, structured error metadata fields, and how errors propagate across a multi-agent pipeline.

High-stakes enforcement decisions. A recurring exam pattern asks whether a constraint should be enforced via a system prompt or programmatically. The Academy does not address this distinction explicitly. Our concept on high-stakes enforcement decision rules covers the exam's preferred reasoning pattern.

How much hands-on building experience does the exam expect?

The CCA-F is scenario-based: 60 multiple-choice questions, each presenting a realistic architecture or debugging situation. You are not asked to write code during the exam. However, the scenarios assume you have built with the API, not just read about it.

Specifically, the exam expects you to reason about:

• What happens when a tool call returns an error mid-loop
• How to structure a CLAUDE.md file for a monorepo with path-specific rules
• When to use tool_choice: {"type": "any"} versus {"type": "auto"}
• How to pass structured context between a coordinator and subagents without losing attribution

None of these questions are answerable from Academy videos alone. Candidates who have built at least one end-to-end agentic workflow using the Messages API, even a simple one, consistently report that the scenarios feel familiar rather than abstract.

A minimal hands-on checklist before sitting the exam:

1. Make a raw Messages API call with a tool definition and inspect the stop_reason field.
2. Build a two-agent system where a coordinator delegates to a subagent and handles a tool error.
3. Create a CLAUDE.md file with at least one path-scoped rule and test it in Claude Code.
4. Write a prompt that returns structured JSON and validate it against a schema.

# Minimal Messages API call with a tool definition
curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-4-5",
    "max_tokens": 1024,
    "tools": [{
      "name": "get_order_status",
      "description": "Returns the current status of a customer order by order ID.",
      "input_schema": {
        "type": "object",
        "properties": {
          "order_id": {"type": "string", "description": "The unique order identifier."}
        },
        "required": ["order_id"]
      }
    }],
    "messages": [{"role": "user", "content": "What is the status of order ORD-9821?"}]
  }'

After running this, inspect the response body. If stop_reason is "tool_use", Claude is requesting the tool. If it is "end_turn", Claude answered without calling it. Understanding that distinction is foundational to Domain 1.

What is the fastest credible path from Academy to CCA-F pass?

We suggest a four-week structure for candidates who already have some API experience.

Our practice exams use the same 100 to 1000 scale and 720 passing bar as the real exam. The adaptive engine, which uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, surfaces the specific task statements where you are weakest so you can return to the right Academy module or concept page rather than re-reading everything.

AI Skill Certs is an independent prep platform; we are not affiliated with or endorsed by Anthropic. We built our 174-concept library specifically to cover the gaps between Academy content and the full 30-task-statement blueprint.

What does Anthropic recommend for MCP debugging and tool discovery?

Per Anthropic's MCP documentation, the recommended debugging approach for MCP servers is to inspect the tool list returned at connection time, verify that tool descriptions are unambiguous and non-overlapping, and use structured error responses with the isError flag rather than raising exceptions that the model cannot interpret.

For tool discovery in production, Anthropic recommends scoping MCP servers to the narrowest permission set that satisfies the use case, using environment variable expansion for credentials rather than hardcoding them in configuration files, and testing tool descriptions by asking Claude to explain which tool it would use for a given task before running it. These patterns appear directly in exam scenarios under Domain 2.

The exam does not test AWS Bedrock or Google Cloud Vertex AI configuration syntax. It tests Claude's behaviour and the architectural decisions that surround it. Cloud integration knowledge is useful context, but it will not move your scaled score.