Claude Developer Certification: CCDV-F Complete Exam Guide

The claude developer certification (CCDV-F) is Anthropic's proctored, scenario-based exam for developers who build production applications on the Claude API. It sits alongside the Architect Foundations track (CCAR-F) inside the Claude Partner Network, a $100M programme that had attracted more than 40,000 partner applicant firms and 10,000 certified individuals as of 3 June 2026. If you are deciding whether to sit the exam, or you are already studying and want to know where to focus, this guide gives you the full blueprint, domain-by-domain priorities, and a practical study sequence.

What is the CCDV-F exam format and scoring?

The Claude Certified Developer, Foundations exam (CCDV-F) consists of 53 items delivered in a 120-minute window. Items are multiple-choice and multiple-response; each item states how many answers to select. Every item is scenario-based, which means the exam tests practical judgement rather than definition recall.

Scores run on a 100-to-1000 scale. The passing score is 720. Your score report shows pass or fail, your scaled score, and percent-correct by domain. Anthropic does not publish the raw-to-scaled conversion, so we will not speculate on an exact question count needed to pass.

Unlike the CCAR-F Architect exam, CCDV-F does not draw from a scenario bank. Items are written directly against the skills listed in each domain, so the full 53-item set is in play at every sitting.

What are the eight CCDV-F domains and how are they weighted?

The exam blueprint covers eight domains. The weights below are exact percentages from the official CCDV-F exam guide (published 8 July 2026).

Domain 2 (Applications and Integration) alone accounts for 33.1% of the exam. That is more than three times the weight of any other single domain. Study time should reflect this asymmetry.

How should you study for Applications and Integration, the heaviest domain?

Applications and Integration at 33.1% is the single most important domain on the exam. Answer-first: prioritise this domain above all others, and return to it repeatedly throughout your preparation.

The domain covers the full lifecycle of building a Claude-powered application: structuring API calls correctly, handling streaming and non-streaming responses, managing rate limits and retries, integrating Claude into existing service architectures, and choosing the right model for a given integration pattern. Expect scenario items that present a partially built integration and ask you to identify the correct fix or the most appropriate design decision.

Practical study actions for this domain:

1. Work through the Messages API request-response cycle until you can trace every field in a request and every field in a response without looking at documentation.
2. Practise reading raw API responses, including stop reasons. The stop_reason field inspection concept is directly relevant to integration debugging scenarios.
3. Build at least one end-to-end integration that handles streaming, retries on transient errors, and logs structured output for observability.
4. Study how Claude fits into multi-service architectures: webhook handlers, background workers, and synchronous request paths each carry different latency and error-handling requirements.

How heavily does the exam test agents, workflows, and termination conditions?

Domain 1 (Agents and Workflows) carries 14.7% of the exam, making it the third-largest domain by weight. That translates to roughly 7 to 8 items in a 53-item exam, enough to move your scaled score meaningfully.

The domain tests your ability to design and reason about agentic loops: how a model decides to call a tool, how it processes the result, and how the loop terminates. Termination conditions are a recurring exam angle because incorrect termination logic is a common production failure mode. You should understand the difference between a loop that exits cleanly on task completion and one that exits prematurely due to a misconfigured stop condition.

Key concepts to master:

• Agentic loop anti-patterns: what causes loops to stall, repeat, or terminate early.
• Model-driven vs pre-configured decision making: when to let the model decide the next step versus encoding the decision in code.
• Tool result appending: how tool outputs are correctly returned to the model in the conversation structure.
• Workflow orchestration patterns, including fixed sequential pipelines and dynamic adaptive decomposition.

The exam consistently rewards deterministic solutions over probabilistic ones when stakes are high, so if a scenario involves a high-consequence action (deleting data, sending an external message, making a financial transaction), the correct answer almost always involves a deterministic guard rather than relying on the model's judgement alone.

What prompt and context engineering patterns are most tested?

Domain 6 (Prompt and Context Engineering) carries 11.0% of the exam. The domain is not the heaviest, but it underpins every other domain: a poorly structured prompt or a mismanaged context window will degrade performance in agents, integrations, and evals alike.

The most frequently tested patterns in this domain cluster around three themes:

System prompt design. Large static system prompts are a cost and latency concern. The exam tests whether you know when to use prompt caching to reduce repeated token costs, and how to structure a system prompt so that the most stable content sits at the top of the cache boundary.

Few-shot examples. Few-shot prompting is the highest-leverage technique for improving output consistency on structured tasks. The exam tests your ability to identify when few-shot examples will help (ambiguous edge cases, extraction tasks with specific formatting requirements) versus when they add cost without benefit.

Context window management. Scenarios involving long conversations or large document inputs test your understanding of the attention dilution problem and strategies for mitigating it, including summary injection and structured context passing.

Our Prompt Engineering and Structured Output concept library covers the patterns that appear most often in scenario items.

How deep does the exam go on Tools and MCPs?

Domain 8 (Tools and MCPs) carries 10.6% of the exam. The depth expected is practical rather than theoretical: you need to know how to design tool interfaces that Claude will use correctly, how to handle tool errors, and how MCP servers fit into a production architecture.

Specific skills the exam tests:

• Writing tool descriptions that function as a selection mechanism. A vague description causes tool misrouting; a precise description with scope constraints reduces errors.
• Handling the isError flag in MCP responses and propagating errors correctly through multi-agent systems.
• Deciding when to build a custom MCP server versus using an existing one.
• Scoping tools appropriately so that a given agent role only has access to the tools it needs.

The Tool Design and MCP Integration concept area covers these patterns in depth. Pay particular attention to the MCP isError flag pattern and tool descriptions as a selection mechanism, both of which appear in scenario items that ask you to diagnose why a tool is being called incorrectly or why an error is not being handled.

What security and safety knowledge does the exam require?

Domain 7 (Security and Safety) carries 8.1% of the exam. The domain is not the largest, but the scenarios in it tend to be high-stakes and the correct answers require precise reasoning about trust boundaries.

The most tested area is prompt injection defence, particularly for agents that have access to file systems, shell environments, or code repositories. The core principle is that content retrieved from external sources (files, web pages, database records) must be treated as untrusted input, not as instructions. An agent that executes shell commands based on content it read from a file is vulnerable to injection attacks embedded in that content.

Exam scenarios in this domain typically present an agent architecture and ask you to identify the weakest point in the trust model, or to choose between two designs where one correctly isolates untrusted content and one does not.

Additional security topics that appear in the blueprint:

• Minimal-permission tool design: agents should only have access to tools that their current task requires.
• Human-in-the-loop gates for irreversible actions.
• Safe handling of credentials and secrets in MCP configuration (environment variable expansion rather than hardcoded values).

Are evals, testing, and debugging a real exam angle despite the low weight?

Domain 4 (Eval, Testing, and Debugging) carries only 2.6% of the exam, the smallest weight of any domain. That is roughly one to two items in a 53-item exam. It is a real domain with real items, but it should not consume a disproportionate share of your study time.

The items that do appear in this domain tend to test your ability to reason about what a good evaluation looks like: whether a given metric captures the right signal, how to detect regressions in model behaviour across versions, and how to structure a debugging workflow when a Claude integration produces unexpected output.

The exam does not require you to implement a full eval framework. It requires you to reason about eval design at the level of a developer who is responsible for the quality of a production Claude integration.

How much model selection and cost optimisation knowledge is expected?

Domain 5 (Model Selection and Optimisation) carries 16.8% of the exam, making it the second-largest domain. This is a significant weight and one that candidates sometimes underestimate.

The domain tests your ability to choose the right Claude model for a given task based on capability, latency, and cost trade-offs. It also covers token optimisation strategies: prompt caching, batching, streaming, and the synchronous-versus-batch decision for high-volume workloads.

Key study areas for this domain:

• Understanding the capability and cost profile of each Claude model tier (Haiku, Sonnet, Opus) and when each is appropriate.
• Prompt caching mechanics: which content is cacheable, how cache hits are billed, and how to structure prompts to maximise cache reuse.
• Batch API versus synchronous API: the batch API is appropriate for high-volume, latency-tolerant workloads; the synchronous API is appropriate for interactive use cases.
• Token counting and context window sizing: knowing when a task will exceed a model's context window and how to handle that gracefully.

What is the recommended study sequence for CCDV-F?

Given the domain weights, a rational study sequence prioritises time proportional to exam impact:

1. Applications and Integration (33.1%) -- spend the most time here. Build integrations, read API responses, handle errors.
2. Model Selection and Optimisation (16.8%) -- understand model tiers, caching, and batching thoroughly.
3. Agents and Workflows (14.7%) -- master agentic loop mechanics and termination conditions.
4. Prompt and Context Engineering (11.0%) -- focus on system prompt design, few-shot patterns, and context window management.
5. Tools and MCPs (10.6%) -- tool description quality, error handling, MCP scoping.
6. Security and Safety (8.1%) -- prompt injection defence and trust boundary design.
7. Claude Code (3.1%) -- familiarise yourself with the tool but do not over-invest.
8. Eval, Testing, and Debugging (2.6%) -- understand eval design principles; one to two items at most.

AI Skill Certs' adaptive study engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold to route you to the concepts where you have the most to gain. Practice exams mirror the real CCDV-F format: 53 questions, scored 100 to 1000 with 720 as the passing bar. AI Skill Certs is an independent prep platform and is not affiliated with or endorsed by Anthropic.