Generative AI Developer Certification: The CCDV-F Skill Map

If you are evaluating a generative ai developer certification in 2026, the Claude Certified Developer, Foundations (CCDV-F) is the credential that maps most directly to what production AI teams actually build. Launched 12 March 2026 as part of Anthropic's Claude Partner Network, it costs $125 per attempt, runs 53 items in 120 minutes, and scores on a 100-to-1000 scale with a passing mark of 720. This post walks through every domain, explains what the exam actually rewards, and tells you where to focus your preparation time.

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

What does the CCDV-F exam actually test?

The CCDV-F is not a trivia exam. Every item is written against a specific skill in one of eight domains, and the weighting is deliberately uneven. The two heaviest domains together account for just over half the exam.

Domain 2 (Applications and Integration, 33.1%) is the single largest domain by a wide margin. If you have limited study time, that is where to start. Domain 5 (Model Selection and Optimisation, 16.8%) is the second priority. Together they represent roughly half the exam.

Unlike the architect track (CCAR-F), the CCDV-F does not draw from a scenario bank. Items are written directly against the skills in each domain, which means breadth across all eight domains matters more than deep scenario rehearsal.

How does Domain 2 (Applications and Integration) break down?

Domain 2 is the exam's centre of gravity at 33.1% of the total score. It tests whether you can wire Claude into real systems, not just call the API in isolation.

The skills tested here include retrieval-augmented generation (RAG) design, chunking strategy, retrieval evaluation, structured output pipelines, streaming integration, and multi-turn conversation management. Candidates who have only read documentation tend to stumble on chunking trade-offs: fixed-size chunking is simple but ignores semantic boundaries; semantic chunking preserves meaning but adds latency and cost. The exam expects you to reason about that trade-off in context, not recite a rule.

Retrieval evaluation is a recurring theme. You should be comfortable explaining why recall@k and mean reciprocal rank measure different failure modes, and when each matters more. If your retrieval pipeline returns the right document 80% of the time but ranks it fifth, your downstream generation quality suffers even though recall looks fine.

Structured output reliability is another tested skill. The exam rewards candidates who know when to enforce schema at the API level versus when a well-crafted prompt with examples is sufficient. Our Prompt Engineering & Structured Output concept library covers the schema-design patterns that appear most often in these items.

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

Domain 5 (16.8%) tests cost and latency reasoning alongside model capability judgement. The exam does not ask you to memorise benchmark numbers. It asks you to reason about which model tier is appropriate for a given task, given constraints on latency, cost, and quality.

Key skills include:

1. Selecting between model tiers (Haiku, Sonnet, Opus) based on task complexity and throughput requirements.
2. Applying prompt caching to reduce cost on repeated context.
3. Choosing between synchronous and batch processing based on latency tolerance.
4. Identifying when a smaller, fine-tuned model outperforms a larger general model on a narrow task.

The exam consistently rewards proportionate solutions. A candidate who routes every task to the most capable model will lose marks to a candidate who routes simple classification to a cheaper tier and reserves the frontier model for synthesis and judgement tasks.

How much weight does agent orchestration carry?

Domain 1 (Agents and Workflows, 14.7%) tests your ability to design, debug, and reason about agentic systems. The skills here overlap meaningfully with the CCAR-F architect exam, but the CCDV-F frames them from a developer implementation perspective rather than an architectural design perspective.

The failure modes the exam targets include:

• Infinite loops: an agent that re-invokes a tool without a termination condition.
• Tool misuse: a model selecting the wrong tool because descriptions are ambiguous.
• Deadlocks in multi-agent pipelines: two subagents waiting on each other's output.
• Context degradation: a long-running session where early instructions are effectively lost.

Our Agentic Loop Anti-Patterns concept covers the loop and deadlock failure modes in detail. For the developer exam, the practical question is always: what is the minimal code change that fixes the failure without over-engineering the solution?

Parallel subagent spawning is also tested. The exam expects you to know when parallelism reduces latency and when it introduces race conditions or inflated cost that outweigh the speed benefit.

What do Tools and MCPs (Domain 8) questions look like?

Domain 8 (Tools and MCPs, 10.6%) tests your ability to design, configure, and debug tool integrations, including Model Context Protocol (MCP) servers. The exam rewards candidates who understand the difference between a tool description that routes correctly and one that causes the model to misfire.

A common item pattern: you are given a tool description and a scenario where the model consistently selects the wrong tool. The correct answer identifies the ambiguity in the description and rewrites it to be specific enough that the model routes correctly. Our Tool Descriptions as Selection Mechanism concept maps directly to this pattern.

MCP-specific skills include:

• Configuring environment variable expansion in MCP server config.
• Distinguishing isError flag semantics from a valid empty result.
• Scoping MCP servers to the correct context level (project vs. user vs. global).

The MCP isError Flag Pattern is a concept that appears in both the developer and architect exams, but the CCDV-F tests it at the implementation level: given a tool response, does this represent an access failure or a valid empty result, and what should the agent do next?

How does Domain 6 (Prompt and Context Engineering) differ from what you already know?

Domain 6 (11.0%) tests applied prompt engineering, not theoretical knowledge. The exam does not reward candidates who can define chain-of-thought prompting. It rewards candidates who can look at a broken prompt, identify why it fails, and apply the minimal fix.

The skills tested include:

• Few-shot example selection: choosing examples that cover edge cases, not just the happy path.
• Context window management: deciding when to summarise, truncate, or fork a session.
• System prompt design: writing instructions that are specific enough to constrain behaviour without being so rigid that they break on edge cases.

The attention dilution problem is a recurring theme. When a context window is very long, the model's effective attention to instructions placed in the middle degrades. The exam expects you to know where to place critical instructions and why.

What does Security and Safety (Domain 7) test in practice?

Domain 7 (8.1%) tests your ability to design guardrails, detect prompt injection, and reason about hallucination mitigation. The exam does not test policy recall. It tests whether you can identify a vulnerability in a system design and propose a fix that is proportionate to the risk.

Key skills include:

• Detecting prompt injection vectors in tool outputs and user inputs.
• Designing output validation layers that catch hallucinated facts before they reach users.
• Choosing between prompt-based and programmatic enforcement for a given constraint.

The exam rewards the principle that safety controls should be layered, not singular. A single system prompt instruction is not sufficient for high-stakes constraints; programmatic enforcement at the output layer is the correct answer when the stakes are high.

How does the CCDV-F compare to the CCAR-F architect exam?

Both exams cost $125, run for 120 minutes, and score on the same 100-to-1000 scale with a 720 passing mark. The credential is valid for 12 months from the award date in both cases. The structural differences are significant.

The CCAR-F draws four scenarios at random from a bank of six per sitting, which means the specific scenarios you face vary. The CCDV-F has no scenario bank; items are written directly against domain skills, so the exam is more predictable in structure but broader in coverage.

If you are a developer who writes production code and wants to validate implementation skills, CCDV-F is the right track. If you are designing systems and making architectural trade-offs, CCAR-F is the better fit. Some candidates sit both.

How should you structure your preparation?

Given the domain weights, a rational study allocation for a candidate with four weeks of preparation time looks like this:

Domains 3 and 4 together account for only 5.7% of the exam. Do not ignore them, but do not over-invest either. A candidate who achieves mastery on Domains 2 and 5 and solid competence on Domains 1, 6, 7, and 8 is well-positioned to pass.

Our adaptive study engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, which means it will not move you past a concept until your demonstrated accuracy justifies it. Practice exams mirror the real format: 53 items, scored 100 to 1000, with 720 as the passing bar.

What is the Claude Code domain (Domain 3) actually testing?

Domain 3 (Claude Code, 3.1%) is the smallest domain by weight, but it is not trivial for candidates who have not used Claude Code in a real project. The skills tested include configuring Claude Code for a team environment, understanding the three-level configuration hierarchy, and knowing when to use plan mode versus direct execution.

The Three-Level Configuration Hierarchy concept is directly relevant here. The exam tests whether you understand which configuration level takes precedence and what the version control implications are for each level.

Given the 3.1% weight, a candidate who has used Claude Code in production will likely find these items straightforward. A candidate who has only read about it may need to spend an hour or two on hands-on practice to build the intuition the exam tests.

Where does the CCDV-F sit in the broader Claude Partner Network?

The Claude Partner Network launched 12 March 2026 as a $100M programme. As of 3 June 2026, it had over 40,000 partner applicant firms and more than 10,000 certified individuals across all tracks.

The four live proctored tracks are:

Anthropic has indicated more tracks are planned for later in 2026, with no dates announced. All four current exams are delivered online-proctored or at a Pearson VUE test centre.

The CCDV-F credential is valid for 12 months from the date it is awarded. Given the pace of change in the field, that renewal cadence is deliberate: the exam is designed to test current production skills, not historical knowledge.

On our platform, adaptive study, Archie tutoring, and practice exams for CCDV-F are available today. If you want to see where your knowledge gaps are before committing to a sitting, a diagnostic practice exam is the fastest way to find out.