Claude Code Certification: Skills vs MCP for CCDV-F

If you are preparing for the Claude Code certification, one question surfaces in almost every study group: do Claude Skills replace MCP servers, or do they coexist? The short answer is they coexist in a layered architecture, and the CCDV-F exam tests whether you know which layer to reach for in a given scenario. This guide unpacks that distinction, maps it to the official domain weights, and gives you the prompt patterns that matter most in production-grade agents.

What exactly is the CCDV-F exam, and how is it structured?

The Claude Certified Developer, Foundations exam (code: CCDV-F) costs $125 USD per attempt, contains 53 items, and runs for 120 minutes. Passing requires a scaled score of 720 on a 100-to-1000 scale. Per Anthropic's exam guide, the raw-to-scaled conversion is not published, so we do not quote an exact question count as the pass mark.

Unlike the CCAR-F Architect exam, CCDV-F has no scenario bank. Items are written directly against the skills in each domain, which means the exam rewards precise conceptual distinctions over pattern-matching to a handful of pre-set scenarios.

The eight domains and their exact weights are:

Domain 2 (Applications and Integration, 33.1%) and Domain 5 (Model Selection and Optimisation, 16.8%) together account for nearly half the exam. Domain 8 (Tools and MCPs, 10.6%) is where the Skills-versus-MCP distinction is tested most directly, but the question bleeds into Domains 1 and 6 as well.

Do Claude Skills replace MCP servers, or do they coexist?

They coexist. The "MCP is dead" narrative circulating in some developer videos is not a valid exam angle. The official architecture composes Skills, subagents, and MCP servers in distinct layers, each solving a different problem.

A Claude Skill encodes procedural knowledge: a reusable, portable instruction set that tells the model how to behave in a class of situations. Think of it as a prompt-level capability that travels with the agent. A Skill for "write idiomatic React components" or "apply our team's code review rubric" is self-contained; it needs no external call to do its job.

An MCP server provides external connectivity: it exposes tools that reach outside the model's context window to fetch live data, mutate state, or call third-party APIs. A Skill cannot check a deployment pipeline's current status without hallucinating, because that status is not in the training data and changes continuously. That task forces an MCP server implementation.

The exam distinction is clean: procedural knowledge lives in Skills; external connectivity lives in MCP servers. When a scenario describes a task that requires real-time or mutable data, the correct answer almost always involves an MCP server. When a scenario describes a task that requires consistent behaviour across many invocations without external I/O, a Skill is the right layer.

How does the exam test the Skills-versus-MCP boundary in practice?

CCDV-F items are scenario-based and test practical judgement. A typical boundary question might describe an agent that needs to both apply a company's deployment checklist (procedural) and query the current CI/CD pipeline status (live data). The correct architecture composes a Skill for the checklist with an MCP server for the status query.

Two Skill categories appear in exam scenarios:

Capability Uplift Skills extend what the model can do: browser automation, code execution, image analysis. These add a new class of action. Exam scenarios involving Capability Uplift Skills typically ask you to identify the correct tool boundary or the safety constraint that applies.

Encoded Preference Skills encode how the model should behave within an existing capability: React best practices, a house style guide, a security review rubric. These shape output quality without adding new I/O. Exam scenarios involving Encoded Preference Skills typically ask you to identify when a Skill is sufficient versus when an MCP server is also required.

The Tool Design & MCP Integration concept library covers the MCP side of this boundary in depth, including the isError flag pattern and structured error metadata that appear in Domain 8 items.

Which prompt-engineering patterns matter most for CCDV-F?

Domain 6 (Prompt and Context Engineering, 11.0%) tests whether you can design prompts that reduce hallucination, control output format, and handle ambiguous inputs reliably. Three patterns carry the most exam weight.

Few-shot examples as the highest-leverage technique. When a model is producing inconsistent output format or hallucinating in edge cases, the fastest fix is usually a well-chosen set of examples. The exam rewards knowing when to deploy few-shot examples versus when to tighten the schema or adjust the system prompt.

Temperature and sampling configuration for production agents. Lower temperature settings reduce variance and are appropriate for tasks where consistency matters more than creativity: code generation, structured data extraction, compliance checks. Higher temperature is appropriate for brainstorming or draft generation. The exam tests whether you can match the sampling configuration to the task, not whether you can recite a specific temperature value.

Structured output with explicit schemas. Asking the model to return JSON against a defined schema, and validating the response programmatically, is the standard pattern for preventing hallucination in production pipelines. The exam distinguishes between prompts that request structure informally ("return JSON") and prompts that enforce structure with a schema and a validation loop.

{
  "type": "object",
  "properties": {
    "deployment_ready": { "type": "boolean" },
    "blocking_issues": {
      "type": "array",
      "items": { "type": "string" }
    },
    "confidence": { "type": "number", "minimum": 0, "maximum": 1 }
  },
  "required": ["deployment_ready", "blocking_issues", "confidence"]
}

A schema like the one above, passed as the response_format or embedded in the system prompt with a validation loop, is the exam-preferred pattern for any agent that must make a binary decision based on structured evidence. For a deeper treatment of how context position affects reliability, see our guide on Prompt Engineering & Structured Output.

How does Claude Code fit into the CCDV-F exam?

Domain 3 (Claude Code, 3.1%) is the smallest domain by weight, but it is not ignorable: three percent of 53 items is roughly one to two questions, and they are often precise. The domain tests configuration and workflow knowledge rather than general coding ability.

The Claude Code Configuration & Workflows concept library covers the three-level configuration hierarchy and version control implications that appear in Domain 3 items. Key distinctions include:

• The difference between project-level and user-level configuration, and which settings belong in version control
• How CLAUDE.md files interact with the configuration hierarchy
• When to use plan mode versus direct execution, and how that choice affects auditability

# Inspect the active configuration hierarchy
claude config list

# Run in plan mode to review proposed changes before execution
claude --plan "refactor the authentication module to use JWT"

Domain 3 items tend to be short and factual. If you have worked through the three-level configuration hierarchy concept, you are well-positioned for this domain.

How do Agents and Workflows (Domain 1) connect to the Skills-MCP architecture?

Domain 1 (Agents and Workflows, 14.7%) tests multi-agent design: how subagents are spawned, how context is passed between them, and how errors propagate. The Skills-versus-MCP distinction reappears here because a well-designed agent workflow typically uses both layers simultaneously.

A coordinator agent might load a Skill that encodes the orchestration policy (which subagent handles which task class) while each subagent uses MCP servers to perform its actual work. The exam tests whether you can identify the correct decomposition strategy for a given scenario, including when parallel subagent spawning is appropriate versus sequential execution.

Error handling is a recurring exam theme in Domain 1. When a subagent fails, the correct response depends on the error category: a transient network failure warrants a retry with backoff; a schema validation failure warrants returning structured error metadata to the coordinator rather than retrying blindly. The multi-agent error handling and routing concept covers the four error categories and their correct responses.

What does the exam expect on security and safety (Domain 7)?

Domain 7 (Security and Safety, 8.1%) tests whether you can identify prompt injection risks, design appropriate trust boundaries, and apply the principle of minimal capability. In the context of Skills and MCP servers, the security questions tend to focus on:

• Scope of MCP server permissions. An MCP server that can write to a production database should not be accessible to an agent that only needs read access. The exam rewards scoped, role-specific tool configurations over broad permissions.
• Prompt injection via tool results. When an MCP server returns data from an external source, that data could contain adversarial instructions. The correct pattern is to treat tool results as untrusted input and sanitise or validate before passing them back into the model's context.
• Skill provenance. A Skill loaded from an unverified source could encode harmful behaviour. The exam expects you to know that Skills from unverified sources require review before deployment, analogous to reviewing a third-party dependency.

The tool_choice configuration options concept is relevant here: restricting which tools an agent can call at runtime is a programmatic safety control that the exam consistently prefers over prompt-only instructions for high-stakes scenarios.

How should you allocate study time across the eight domains?

Given the domain weights, a rational allocation prioritises Domain 2 first, then Domains 5 and 1, then Domains 6 and 8, with Domains 7, 3, and 4 receiving proportionally less time.

This is a floor allocation, not a ceiling. If your practice exam results show weakness in a lower-weight domain, adjust accordingly. Our platform's adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold to surface exactly those gaps, so you are not spending time on concepts you have already mastered.

The CCDV-F practice exam on our platform mirrors the real format: 53 questions, scored 100 to 1000 with 720 as the passing bar. Running a timed practice exam before your final week of study gives you a calibrated scaled score and a percent-correct breakdown by domain, which is the same report format you will see on exam day.

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