Claude Developer Certification Path: CCDV-F Domain Weight Guide

The claude developer certification path begins with one exam: the Claude Certified Developer, Foundations (CCDV-F), launched 12 March 2026 as part of Anthropic's Claude Partner Network. Fifty-three items, 120 minutes, $125 per attempt, scored 100 to 1000 with 720 as the pass mark. No live coding. No lab tasks. Every item tests applied reasoning about API integration, model selection, agent design, and security. This post maps the eight domains by their official weights so you can study in proportion to what actually appears on the test.

What is the CCDV-F exam and what does it test?

The CCDV-F is a multiple-choice and multiple-response exam delivered online-proctored or at a Pearson VUE test centre. Per Anthropic's official CCDV-F exam guide (published 8 July 2026), items are written directly against the skills in each domain. Unlike the Architect track (CCAR-F), the developer exam draws no scenario bank, which makes the domain distribution more predictable across sittings. Each item states how many responses to select.

The score report returns a pass or fail verdict, a scaled score, and percent-correct by domain. Anthropic does not publish the raw-to-scaled conversion, so a specific raw question count cannot be stated as the pass mark. The credential is valid for 12 months from the date awarded.

As of 3 June 2026, Anthropic's certification programme had exceeded 10,000 certified individuals across all tracks, with the Claude Partner Network attracting more than 40,000 partner applicant firms.

What are the eight domains and how are they weighted?

The table below reproduces the official domain weights exactly as published in the CCDV-F exam guide. No rounding has been applied.

Domain 2 alone accounts for 33.1 percent of the exam. Domains 3 and 4 together account for fewer than six percent. That gap should drive your time allocation from day one.

Where should the majority of your study time go?

Domain 2 (Applications and Integration, 33.1%). This is the dominant domain. It covers integration architecture across three deployment surfaces: Anthropic's direct API, Amazon Bedrock, and Google Vertex AI. Each surface uses different authentication mechanisms and different model-ID conventions. Amazon Bedrock requires AWS IAM authentication and uses region-scoped model IDs such as anthropic.claude-sonnet-4-5-20251001-v1:0, whereas the direct Anthropic API uses bearer-token authentication with IDs like claude-sonnet-4-6. Google Vertex AI uses service-account credentials scoped to a project and region. Exam items probe whether you recognise these differences as architectural constraints rather than incidental details. A developer who has only used the direct API will encounter unfamiliar patterns here; accounting for that gap early is one of the highest-leverage study decisions you can make.

Domain 5 (Model Selection and Optimization, 16.8%). Questions test when to route to a faster, cheaper model versus a larger one, how to budget tokens across a pipeline, and when prompt caching pays off. Prompt caching attaches to static system-prompt prefixes. Knowing the cache TTL, the minimum token threshold for cache eligibility, and when caching introduces latency rather than reducing it is regularly tested. Synchronous requests suit low-latency interactive use cases; the Message Batches API suits non-interactive, cost-sensitive workloads where throughput matters more than per-request speed. The exam regularly presents scenarios where both cost and latency constraints are active and asks which configuration satisfies both without over-engineering.

Domain 1 (Agents and Workflows, 14.7%). The agentic architecture concepts that appear in this domain include multi-turn tool loops, subagent orchestration, and the decision of when to decompose a task programmatically versus letting the model decide dynamically. A consistent exam pattern: high-stakes workflows are tested with deterministic, pre-configured enforcement rather than model-driven judgment. Know when each approach is appropriate and why probabilistic enforcement is insufficient for irreversible actions.

Is the CCDV-F a coding exam or a reasoning exam?

The CCDV-F tests reasoning, not implementation. No item asks you to write syntactically correct Python, debug a stack trace, or produce a working function. Every item asks whether you can read a scenario and select the correct architectural or operational response.

A representative question presents an agent that silently returns empty results instead of raising an error, then asks which of four remediation approaches addresses the root cause. The answer requires understanding tool error propagation and the distinction between a result with is_error: true and a result with an empty payload. That is a reasoning skill, not a coding skill.

This distinction shapes how you should prepare. Memorising API parameter names is low-yield. Working through scenario-based practice, where you must reason about trade-offs and identify root causes, is high-yield. AI Skill Certs is an independent prep platform (not affiliated with or approved by Anthropic) whose adaptive study engine routes you to your weakest concepts using Bayesian Knowledge Tracing with a 0.90 mastery threshold. The practice exams mirror the real 53-item format scored 100 to 1000.

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

Domain 6 carries 11.0 percent of the exam. It spans prompt engineering techniques including system-prompt structure, few-shot example selection, XML tag use for structured output, and context window management under token pressure.

A common scenario type presents an inconsistent structured-output problem and asks which prompt revision resolves it. The diagnostic pattern the exam rewards is root-cause tracing: identify what in the prompt allows the inconsistency, then select the fix that addresses that specific cause rather than patching a symptom.

# Scenario: structured extraction with inconsistent output formatting
system_prompt = """
Extract the following fields as JSON.
Fields: company_name, contract_value, start_date, end_date.
Return only the JSON object. No prose.
"""

# Exam question: which single change most reliably prevents
# prose appearing before the JSON in the response?
# A) Add "You must" to the instruction
# B) Provide a few-shot example showing the exact output format
# C) Lower the temperature parameter
# D) Repeat the instruction at the end of the user message

The correct reasoning: a concrete example is higher-leverage than a repeated instruction, because the model anchors to the format it has seen rather than to a description of what that format should look like. Option B is the deterministic fix. Options A and D restate the constraint in words; option C addresses generation variance rather than format ambiguity.

How should you approach Domain 8 (Tools and MCPs)?

Domain 8 carries 10.6 percent of the exam and rewards targeted preparation. The tool design and MCP integration concepts tested here are concrete: tool description quality as a selection mechanism, tool_choice configuration options, the MCP isError flag pattern, and MCP scoping hierarchy.

A well-written tool description is the primary lever by which Claude selects the correct tool. Exam items frequently show two versions of the same description and ask which reduces misrouting. The answer consistently points to specificity: a description that names the correct use case and explicitly excludes adjacent misuse prevents Claude from calling a tool in the wrong context.

The tool_choice parameter has three meaningful states: auto (Claude decides whether to call a tool), any (Claude must call at least one tool), and a specific tool name (Claude must call exactly that tool). Items present scenarios where integration correctness depends on forcing tool use and ask which configuration achieves that reliably.

The isError flag pattern appears as a root-cause diagnosis scenario. A result with is_error: true signals operation failure; a result with is_error: false and an empty payload is a valid empty result. Conflating them causes silent failures in production.

{
  "type": "tool_result",
  "tool_use_id": "toolu_01XYZ",
  "is_error": true,
  "content": "Database connection refused: timeout after 5000ms"
}

When an item asks what the orchestrating agent should do on receiving this response, the correct answer involves routing the error rather than treating it as a successful empty result.

How does Domain 7 (Security and Safety) appear on the exam?

At 8.1 percent, Domain 7 returns well above average yield per study hour. Questions focus on prompt injection defences in tool-using agents, blast-radius limitation through tool scope restriction, and the distinction between input validation at system boundaries and internal trust assumptions.

The exam does not test cryptography or network security protocols. It tests the architectural decisions that determine how much damage a compromised or malicious prompt can cause in an agentic pipeline. A common item type presents an agent with broad file-system write access and asks which scope-restriction approach provides the most proportionate fix without over-engineering the solution. The exam rewards proportionate fixes over maximal restrictions.

How do Domains 3 and 4 fit into a realistic study plan?

Claude Code (Domain 3, 3.1%) and Eval, Testing, and Debugging (Domain 4, 2.6%) together account for fewer than six percent of the exam. Neither warrants deep preparation relative to the higher-weight domains. For Domain 3, understanding the three-level configuration hierarchy and headless mode for CI pipelines is sufficient. For Domain 4, the exam tests conceptual framing: what makes an evaluation set representative, and when per-category accuracy is a more reliable signal than aggregate metrics. Both domains reward a light pass over deep study.

What study sequence works in practice?

A four-week plan proportional to domain weights:

Context management and reliability concepts surface across Domains 1, 5, and 6 simultaneously, so treat those concepts as a thread running through weeks 2 and 3 rather than isolating them to a single session. Prompt caching, context compaction, and stale context handling appear in model-selection questions, prompt-engineering questions, and agent-design questions. Seeing them from multiple angles before exam day is worth more than a single focused session.

After each practice exam, review percent-correct by domain and reallocate remaining study time toward your lowest-scoring areas. The pattern the exam rewards consistently is the same one that makes a productive study session: identify the root cause of a wrong answer rather than accepting a near-miss and moving on.