CCDV-F Exam Domains: Complete Blueprint and Study Plan

The CCDV-F exam tests eight domains across 53 items in a 120-minute sitting. All eight CCDV-F exam domains and their official percentage weights come from the exam guide published 8 July 2026. Because CCDV-F items are written directly against the skill statements in each domain, unlike the CCAR-F which draws from a scenario bank, the weighting translates predictably into question distribution.

What are the CCDV-F exam domains and their official weights?

The eight domains are listed below with fractional weights exactly as Anthropic published them.

The exam is scored on a 100-to-1000 scale; the passing mark is 720. Anthropic does not publish the raw-to-scaled conversion formula. A $125 attempt with a 12-month credential window makes the cost per point relatively low compared with multi-day classroom training.

The distribution is deliberately skewed. Domain 2 alone represents 33.1% of the exam. Domains 2, 5, and 1 combined represent 64.6%. Candidates who achieve mastery in those three areas and hold competence across the remaining five have a strong structural advantage.

Why does Domain 2 (Applications and Integration) carry one-third of the exam?

Domain 2 carries 33.1% of the exam because the certification validates working developer skill, not theoretical knowledge, and most practical Claude work lives at the integration layer. Expect items on the Messages API request-response cycle, multi-turn conversation construction, streaming responses, handling tool_use and tool_result turns, and embedding Claude into larger application stacks.

The Messages API request-response cycle is one of the most examined mechanics across both developer and architect tracks. Getting that loop right is the prerequisite for everything else.

A canonical Domain 2 pattern is correct message array construction:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "What are the main clauses in this contract?"},
    ],
)
print(response.content[0].text)

Items typically test what changes when you add a second turn, how to append a tool result, and which stop_reason value signals a completed response versus a pending tool call. Candidates who have built at least one end-to-end Claude application before sitting the exam consistently report that Domain 2 questions feel familiar rather than tricky.

How do Domains 5 and 1 account for another 31.5% of the exam?

Domain 5 (Model Selection and Optimisation, 16.8%) and Domain 1 (Agents and Workflows, 14.7%) together add 31.5%, bringing the top-three total to 64.6%.

Domain 5: Model Selection and Optimisation

This domain asks you to reason about which Claude model fits a given task, how to balance latency, cost, and capability, and how to optimise throughput through caching and batching. Items present scenarios such as a high-volume classification pipeline, a real-time support application, or a long-document summarisation task, and ask which model tier fits each demand profile and why. The exam does not require memorising exact pricing, but it does expect familiarity with the capability-cost spectrum and when prompt caching changes the economics of a workload.

Domain 1: Agents and Workflows

At 14.7%, this domain tests agentic loop construction, tool-call sequences, multi-agent orchestration, and error handling in long-running pipelines. The agentic architecture concepts that underpin the CCAR-F exam overlap substantially here. Patterns such as agentic loop anti-patterns, coordinator responsibilities, and subagent context isolation all appear, though framed from an implementation perspective rather than a system-design one.

A recurring Domain 1 judgment: when does a task warrant a multi-agent system versus a single-agent prompt chain? The answer almost always turns on parallelism requirements and context-window constraints.

What do Domains 6 and 8 test, and how much weight do they carry?

Domain 6 (Prompt and Context Engineering, 11.0%) and Domain 8 (Tools and MCPs, 10.6%) together represent about 21.6% of the exam.

Domain 6: Prompt and Context Engineering

This domain covers prompt design, few-shot example construction, structured output schemas, system prompt architecture, and context window management. Items test practical judgment: given a model returning inconsistent JSON, should you add a type-specific schema, a few-shot example, or revise the system prompt first?

A well-formed structured output schema is a common scenario:

{
  "type": "object",
  "properties": {
    "sentiment": {
      "type": "string",
      "enum": ["positive", "neutral", "negative"]
    },
    "confidence": {
      "type": "number",
      "minimum": 0,
      "maximum": 1
    },
    "summary": {
      "type": "string",
      "maxLength": 200
    }
  },
  "required": ["sentiment", "confidence", "summary"]
}

Knowing when to constrain output with a schema versus relying on prompt guidance alone is a recurring theme. The Prompt Engineering and Structured Output concepts in the CCAR-F library cover the same technique hierarchy and are useful supplementary reading.

Domain 8: Tools and MCPs

Domain 8 tests tool definition, tool selection mechanics, MCP server integration, and error propagation when tools fail. Candidates who have studied the Tool Design and MCP Integration material will recognise the core patterns: tool descriptions as selection mechanisms, the isError flag, structured error metadata, and surfacing failures cleanly to the orchestrating agent. CCDV-F frames these as implementation tasks, but the underlying mechanics are identical to what the architect track tests.

How should you approach Domain 7 and the two lightest domains?

Domain 7 (Security and Safety, 8.1%) is not a domain to skip on the assumption that general judgement will cover it. Items test specific patterns: recognising prompt injection vectors, applying least-privilege principles to tool access, designing rate-limiting and abuse-prevention strategies, and understanding which safeguards the API enforces automatically versus which you must implement yourself.

Domain 3 (Claude Code, 3.1%) and Domain 4 (Eval, Testing, and Debugging, 2.6%) are the lightest areas by weight, together accounting for roughly 5.7% of the exam. Domain 3 focuses on Claude Code configuration, CLAUDE.md structure, and hooks. Domain 4 covers evaluation design, test-driven iteration, debugging tool-call failures, and using structured output to track model behaviour over time.

A candidate who ignores both tail domains saves minimal study time at genuine score risk. On a 53-item exam with a 720 scaled passing bar, a single missed cluster in a narrow domain can shift the outcome.

How should you structure a study plan around the CCDV-F exam domains?

The domain weights translate directly into a priority order. A build-plus-practice approach covers the most ground fastest.

1. Applications and Integration (33.1%): Build a working Claude application using the Messages API. Construct multi-turn conversations, handle tool calls, and test streaming. This single project covers more exam surface than any other preparation activity.
2. Model Selection and Optimisation (16.8%): Study the Haiku-to-Opus capability-cost spectrum. Run the same task against two model tiers and compare quality and latency. Understand when prompt caching changes the cost calculus for high-volume workloads.
3. Agents and Workflows (14.7%): Extend your application to call at least one tool and handle the full tool_use and tool_result loop. Add a delegating pattern where one agent hands a subtask to another.
4. Prompt and Context Engineering (11.0%): Practise writing structured output schemas. Write a few-shot example that corrects a recurring model inconsistency. Measure the token cost of different system prompt lengths.
5. Tools and MCPs (10.6%): Connect a local MCP server to your application. Observe how changing a tool description affects which tool the model selects for ambiguous input.
6. Security and Safety (8.1%): Review Anthropic's guidance on responsible use, input validation, and which safeguards are API-enforced versus developer-implemented.
7. Claude Code and Eval (5.7% combined): Spend one focused session on CLAUDE.md configuration and one on writing assertions against model outputs in a CI context.

The AI Skill Certs platform offers adaptive practice, scored practice exams in the real 53-item CCDV-F format with the same 720 passing bar, and Archie, a Socratic tutor that guides you toward the answer without supplying it directly. AI Skill Certs is an independent preparation platform; it is not affiliated with or endorsed by Anthropic.

How does the CCDV-F domain structure compare with the CCAR-F architect exam?

The two Foundations exams share a $125 cost, a 120-minute time limit, and a 720 passing score, but differ substantially in structure.

Both exams reward practical judgment over recall. CCDV-F leans toward API-level implementation; CCAR-F leans toward system-level design decisions. The overlap in underlying concepts is substantial: tool design, prompt engineering, and agentic patterns appear in both blueprints and reinforce each other during study.

For developers planning to sit both exams, taking CCDV-F first is the typical path. The implementation fluency it builds makes the architectural reasoning in CCAR-F more concrete.