The eight CCDV-F domains and what each one tests

Here is the whole blueprint, heaviest first, with what each domain actually asks for.

Applications and Integration · 33.1%

A third of the exam sits in one domain, which makes it the whole game. It spans six sub-objectives: reading requirements, the systems life cycle, Claude API mechanics, software engineering foundations, application design, and configuration management. The through-line is that a working developer has to turn a stated business need into an architecture and then keep it running. Expect questions that hand you a constraint and ask which requirement follows from it, rather than questions that ask what a term means. Full breakdown of this domain →

Model Selection and Optimization · 16.8%

Tokens, context windows, sampling and non-determinism, the model tiers, and what everything costs. This is where the exam checks whether you understand the machine underneath rather than just the SDK on top. Sampling behavior is worth real attention, including the fact that the newest models reject non-default sampling parameters outright. Full breakdown →

Agents and Workflows · 14.7%

When to build a fixed workflow and when to hand control to a model, how an agent loop is actually constructed, and the standard patterns. The recurring judgment is restraint: agents cost more, fail in more ways, and are the right answer only when the path genuinely cannot be known in advance. Full breakdown →

Prompt and Context Engineering · 11.0%

Context engineering, prompt engineering, and output handling. Practically: system prompts and structure, few-shot examples, keeping multi-turn sessions inside a fixed token budget, and getting output in a shape your code can parse. The most testable idea here is that the context window is a budget you spend, not a container you fill.

Tools and MCPs · 10.6%

Tool implementation, MCP server development, and agentic customization. The distinction that decides the most questions is where code runs: server-side tools execute on Anthropic's infrastructure during the response, custom tools and MCP-served tools execute on your side of the wire, and an Anthropic-defined client tool still leaves execution with your application.

Security and Safety · 8.1%

Application security, guardrails, hooks, and the handling of identity, secrets and keys. The dominant theme is untrusted input: content a model fetches or reads is data, never instructions, and a system that treats it otherwise is the vulnerability the questions are built around.

Claude Code · 3.1%

Roughly two questions. Permission modes, settings, and where a human gate belongs. Worth a light pass, and worth resisting the temptation to over-study because it is the most familiar material on the exam.

Eval, Testing, and Debugging · 2.6%

Roughly one question. Debugging and error handling. That is not a reason to skip evals entirely, because evaluation thinking shows up inside other domains, but it is a reason not to build a study plan around them.

What the weights mean in practice

Applications and Integration, Model Selection, and Agents and Workflows are 64.6% of the exam between them. Claude Code and Eval together are 5.7%, which is about three questions.

The most common misallocation is studying in the order the material is usually presented, which front-loads fundamentals and leaves the heaviest domain for whenever you get to it. Invert that. Start with the 33.1% and work down.

Practice forms should carry the same proportions the exam does. Every form in our bank matches the published weights to within half a question, which is as close as 53 items allows.