Claude Certified Developer CCDV-F: Domain Weights and Study Order

The claude certified developer (CCDV-F) examination is the developer track of Anthropic's Foundations certification programme. It launched on 12 March 2026, costs $125 per attempt, and consists of 53 items to be answered within 120 minutes. Scoring runs from 100 to 1000; the passing mark is 720. The credential is valid for 12 months from the date it is awarded.

What makes CCDV-F distinctive from the companion Architect track (CCAR-F) is its item format: there is no rotating scenario bank. All 53 items are written directly against the task statements within each of the eight domains, which means the blueprint published in Anthropic's exam guide dated 8 July 2026 is your complete syllabus. Study the blueprint; ignore nothing with meaningful weight.

What are the 8 domains on the CCDV-F exam?

The eight domains and their exact percentage weights, per the official exam guide, are:

Source: Anthropic, CCDV-F Exam Guide, 8 July 2026. Approximate item counts are illustrative; Anthropic does not publish domain-level question counts separately from the weights.

The weights are exact fractional percentages, not rounded values. Domain 2 alone accounts for roughly one in three exam points; that concentration has direct consequences for how you should allocate study time.

Which domain carries the most weight, and what does it test?

Applications and Integration (Domain 2, 33.1%) is the dominant domain by a margin that has no equivalent on the Architect exam. The next two by weight are Model Selection and Optimization (Domain 5, 16.8%) and Agents and Workflows (Domain 1, 14.7%). Together, these three domains account for nearly 65% of all exam points.

The practical implication is direct: bring Domain 2 to fluency first, build confident competence in Domains 5 and 1, then allocate remaining time to the five lower-weight domains in proportion to their share. Spreading effort evenly across all eight domains is the most common preparation mistake; it treats a 3.1%-weight domain as equivalent to a 33.1%-weight one.

What Domain 2 actually tests. Applications and Integration covers the full lifecycle of embedding Claude into a production application: constructing valid Messages API requests, handling synchronous and streamed responses, choosing when to apply prompt caching, selecting the appropriate model tier for given cost and latency constraints, and designing application logic around the structure of Claude's responses.

Developers who have shipped production Claude integrations will find much of Domain 2 intuitive. Those whose experience is limited to the API playground will need deliberate preparation against each task statement in the domain.

A representative scenario type: a team makes thousands of API calls per hour using an identical, multi-hundred-token system prompt. The item asks which technique reduces token cost most directly. The answer is prompt caching, applied by marking the static content with cache_control. Understanding when caching is and is not worthwhile (it provides limited benefit for low-volume or short system prompts) is precisely the kind of judgment the exam tests. The Messages API Request-Response Cycle covers the underlying mechanics in detail.

# Applying cache_control to a static system prompt
import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": "You are a financial data assistant. [several hundred tokens of domain context]",
            "cache_control": {"type": "ephemeral"}
        }
    ],
    messages=[{"role": "user", "content": user_input}]
)

The exam does not require you to write code, but understanding what this pattern does and when to deploy it is squarely within Domain 2 scope.

What does Model Selection and Optimization (Domain 5) test?

At 16.8%, Domain 5 is the second-largest domain, covering roughly nine items. The core skill is matching Claude's model tiers to workload requirements across three dimensions: capability, cost, and latency.

Domain 5 asks candidates to reason about scenarios such as a batch classification task processing millions of short documents (which model tier minimises cost while maintaining accuracy above a defined threshold?) or a real-time chat application with a 200 ms response budget (which model and request configuration supports that constraint?). A legal document analysis task requiring deep reasoning over long inputs raises a different question: does the capability premium of a more powerful model justify the cost at this volume?

Beyond model selection, Domain 5 covers inference-side optimisation: batching API calls to reduce overhead, streaming responses to reduce perceived latency, and setting max_tokens to bound output without unnecessarily limiting useful responses. Candidates should also understand the relationship between context window size and cost: longer contexts cost more to process, and knowing when to trim aggressively versus when long-context capability justifies the spend is a practical skill the exam probes.

Candidates who can reason about the cost implications of each choice, not just name the options, will perform well here.

How should Agents and Workflows (Domain 1) be approached?

At 14.7%, Domain 1 is the third-largest domain. For CCDV-F, the focus is narrower than the agentic architecture patterns tested in CCAR-F: this domain tests whether a developer can build functional agentic integrations at the API level, not whether they can design large-scale orchestration systems.

Key skills within Domain 1:

• The agentic loop: model call, tool selection, tool execution, tool result injection, next model call
• Using the stop_reason field to determine whether the model has finished responding or is requesting a tool call
• Correctly appending tool results in the conversation structure before the next model call
• Recognising agentic loop anti-patterns: premature termination, infinite tool-call loops, and context pollution across turns
• Deciding when tool use is the right abstraction versus a simpler prompt chain

Domain 1 items frequently present a described implementation and ask candidates to identify what is wrong and which correction resolves the failure. The exam rewards understanding of the agentic loop mechanics at the API level, not familiarity with high-level orchestration frameworks.

What does Prompt and Context Engineering (Domain 6) cover?

At 11.0%, Domain 6 is the fourth-largest domain. The skills it tests overlap significantly with everyday Claude API work:

• Writing system prompts that shape model behaviour reliably across varied user inputs
• Using few-shot examples to improve output consistency for structured extraction or classification tasks
• Managing context window usage in long-running conversations or document-processing pipelines
• Designing prompts for structured output formats and verifying that the model adheres to the specified schema
• Understanding when and how to use sampling parameters such as temperature

The Prompt Engineering and Structured Output concept area covers these patterns in depth. Domain 6 items test judgment rather than recall: given a described failure mode in a prompt design, which change is most likely to resolve it?

A common Domain 6 item type: a developer reports that a Claude integration produces correctly structured JSON 90% of the time but fails on the remaining 10%, returning malformed or incomplete output. The item asks which technique is most likely to resolve this without increasing cost. The answer typically combines clearer schema specification with a few well-chosen examples that demonstrate the expected format at the edges where the model currently fails. Context window management items test realistic constraints: a document-processing pipeline that naively appends all documents to the context will hit limits quickly, and candidates need to know how to structure retrieval, summarisation, and context pruning strategies.

How much time should developers spend on Tools and MCPs (Domain 8)?

Domain 8 (Tools and MCPs, 10.6%) accounts for approximately six items. Coverage includes:

• Defining tools in the API request using name, description, and input schema
• Writing tool descriptions that the model can reliably interpret to select the correct tool for a given task
• Handling tool errors: distinguishing an access failure from a valid empty result, and returning structured error information the model can act on
• Integrating MCP servers into a Claude application and understanding when MCP-hosted tools are preferable to inline tool definitions

The exam does not require candidates to build an MCP server from scratch, but it does test conceptual understanding of what MCP provides and when to use it. Tool Design and MCP Integration covers these patterns in detail, including tool description quality and error handling strategies.

Domains 3 and 4 are the two lightest domains: Claude Code at 3.1% and Eval, Testing, and Debugging at 2.6%. A developer who uses Claude Code regularly will find Domain 3 items straightforward. Domain 4 tests familiarity with building evaluation sets, measuring model accuracy against ground truth, and diagnosing failure modes systematically. Both domains are manageable within a standard preparation cycle even with proportionally limited study time.

What security and safety knowledge does the exam require?

Domain 7 (Security and Safety, 8.1%) covers approximately four to five items. The scope is practical and threat-focused rather than theoretical.

Prompt injection is the primary threat vector in production Claude applications. User-controlled input that redirects a model's behaviour in an agentic workflow can cause real harm. Domain 7 items test whether candidates can identify injection risks in a described architecture and select effective mitigations, such as structurally separating instructions from user-supplied data so that the model cannot be directed to override its intended behaviour.

Tool blast radius is a closely related concern. A tool definition with overly broad permissions can cause disproportionate damage if the model calls it incorrectly or if a malicious input manipulates the call parameters. The exam tests whether candidates can recognise over-permissioned tool definitions and propose constrained alternatives that limit the damage surface.

Trust hierarchies: distinguishing the authority of system prompts, user messages, and tool results appears in items that ask which actions the model should perform autonomously versus which require explicit user approval before proceeding. The exam consistently rewards candidates who apply the correct trust level to each input source rather than treating all context as equally authoritative.

Domain 7 does not require security engineering expertise. It tests pattern recognition at the application layer. Developers familiar with how agentic systems can be manipulated will find this domain manageable within a standard preparation cycle.

How is CCDV-F preparation structured on AI Skill Certs?

Adaptive study, Archie tutoring, and practice exams for CCDV-F are live on the platform today. The practice exams mirror the real format: 53 questions, scored 100 to 1000 with 720 as the passing bar, covering all eight domains at their published weights. The question bank is built against the 8 July 2026 exam guide.

The adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold. Rather than stopping after a fixed number of questions per domain, it continues probing each area until your response patterns are statistically reliable. Prep time concentrates on domains where uncertainty remains, rather than rehearsing knowledge you have already demonstrated.

One note for developer-track candidates: the concept library at /concepts currently covers the CCAR-F Architect track and its 174 atomic concepts mapped to five domains. A CCDV-F concept library is not yet live. Developer-track candidates use the adaptive practice bank and Archie for primary study.

AI Skill Certs is an independent prep platform and is not affiliated with, endorsed by, or approved by Anthropic.