AI Developer Certification: The CCDV-F Blueprint Guide

If you are weighing which ai developer certification to pursue in 2026, the Claude Certified Developer, Foundations (CCDV-F) is the most directly production-relevant option for teams building on Claude. Launched 12 March 2026 as part of Anthropic's Claude Partner Network, it costs $125 per attempt, runs 53 items in 120 minutes, and is scored on a 100-to-1000 scale with a passing mark of 720. This guide unpacks the full exam blueprint, the API-pattern decisions the exam tests, and the study approach that matches how the items are actually written.

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 blueprint actually look like?

The CCDV-F spans eight domains with fractional weights published in the official exam guide. Unlike the Architect track (CCAR-F), CCDV-F has no scenario bank; every item is written directly against the skills in each domain, so the weighting is the single most important planning tool you have.

Domain 2 (Applications and Integration) alone accounts for a third of the exam. Combined with Domain 5 (Model Selection and Optimisation) and Domain 1 (Agents and Workflows), those three domains represent roughly 65% of your score. The remaining five domains share the other 35%, so neglecting any of them entirely is risky, but your study hours should reflect the weights above.

The practical implication: if you are spending equal time on Claude Code (3.1%) and Applications and Integration (33.1%), you are misallocating roughly ten-to-one.

How hands-on is CCDV-F compared with a theory exam?

Very hands-on. Every item is scenario-based and tests practical judgement, not recall. The exam consistently rewards deterministic solutions over probabilistic ones when stakes are high, proportionate fixes, and root-cause tracing. You will not be asked to recite a definition; you will be asked what to change in a given architecture to fix a specific failure.

This means rote memorisation of API parameter names is less useful than understanding why a particular pattern exists. For example, knowing that tool_choice can be set to auto, any, or a specific tool name matters less than knowing which setting to choose when a multi-step agent is silently skipping tool calls.

When should you use the Messages API, Message Batches API, or an Agent SDK?

This is one of the highest-yield decision points for Domain 2 and Domain 1 combined. The three patterns serve different production needs, and the exam tests whether you can match the pattern to the constraint.

For an agent that must call tools, inspect results, and decide whether to continue, you need a loop that reads the stop_reason field on each response. When stop_reason is tool_use, the agent appends the tool result and continues; when it is end_turn, the loop terminates. Our concept on stop_reason field inspection covers the exact mechanics the exam tests.

For batch workloads such as nightly document classification across thousands of records, the Message Batches API avoids holding open thousands of synchronous connections. The exam will present a scenario with a cost or throughput constraint and ask you to select the appropriate pattern.

A minimal synchronous Messages API call looks like this:

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Summarise the following support ticket: ..."}
    ]
)
print(response.content[0].text)

A batch submission, by contrast, uses client.beta.messages.batches.create with a list of Request objects, each carrying a custom_id for correlation when you retrieve results later.

How do you choose between Opus, Sonnet, and Haiku?

Domain 5 (Model Selection and Optimisation, 16.8%) is the second-largest domain and almost entirely about this decision. The exam does not ask you to memorise benchmark scores; it asks you to match model capability to a production constraint.

The exam frequently presents a scenario where a team is using Opus for a task that Haiku would handle adequately, then asks what change would reduce cost without degrading output quality. The correct answer is almost always to move the simpler subtask to a smaller model and reserve the larger model for the step that genuinely requires it.

Extended thinking is a related concept: it increases reasoning depth but also increases latency and token cost. The exam tests whether you know when extended thinking adds value (complex multi-step problems with ambiguous paths) versus when it is wasteful (straightforward extraction tasks with a clear schema).

One important note for candidates who have studied older materials: token budget parameters for extended thinking changed in newer Claude models. Configurations that were valid on earlier model versions may be rejected by current APIs. Always verify against the current Anthropic API documentation.

How do you handle context limits and long-running agents?

Domain 1 (Agents and Workflows, 14.7%) and Domain 6 (Prompt and Context Engineering, 11.0%) both test context management, and the failure modes overlap. When an agent runs many tool calls over a long session, the context window fills with tool traces, intermediate results, and conversation history that is no longer decision-relevant.

The three main strategies the exam tests are:

1. Pruning: Remove low-value turns (verbose tool outputs, superseded intermediate results) before they consume the window. This is appropriate when the removed content is genuinely not needed for future steps.
2. Compaction / summarisation: Replace a long history segment with a structured summary that preserves the facts needed for continuation. Our concept on summary injection for fresh sessions covers the exact pattern.
3. Subagent isolation: Spawn a subagent with only the context it needs for its specific subtask, keeping the coordinator's context clean. See subagent context isolation for the architectural rationale.

The exam distinguishes between these strategies by the scenario constraint. If the agent must continue a single long session, summarisation is usually correct. If the task is decomposable into independent subtasks, subagent isolation is usually correct. Pruning is the right answer when the content being removed is provably irrelevant to future decisions.

How important are MCPs and tool integrations for CCDV-F?

Domain 8 (Tools and MCPs, 10.6%) and Domain 2 (Applications and Integration, 33.1%) both touch tool design, making tool-related content appear in roughly 44% of the exam by weight when you account for overlap. The exam tests practical integration decisions, not MCP protocol internals.

Key decision points the exam covers:

• When to build a custom MCP server versus use an existing one: If a standard integration exists and meets your security and scoping requirements, using it is almost always correct. Building custom is justified when you need non-standard data access patterns or tight permission scoping.
• Tool description quality: The model selects tools based on their descriptions. A vague description causes misrouting; a precise description with scope, input format, and expected output reduces selection errors. Our concept on writing effective tool descriptions is directly exam-relevant.
• Error signalling: The isError flag in MCP tool responses distinguishes a genuine access failure from a valid empty result. The exam tests whether you know when to set it and what the downstream agent should do in each case.
• Tool overload: Giving an agent too many tools degrades selection accuracy. The correct fix is usually to scope tools to roles or to split a generic tool into specific ones, not to add more instructions to the system prompt.

For tool design and MCP integration concepts mapped to the exam, our concept library covers the full domain.

What does a realistic CCDV-F study plan look like?

Given the domain weights, a proportionate allocation across a six-week preparation period looks like this:

Our adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, which means it will not move you off a concept until you have demonstrated reliable recall under varied conditions. Practice exams on the platform mirror the real format: 53 questions, scored 100 to 1000, with 720 as the passing bar.

The prompt engineering and structured output concepts and the context management and reliability concepts are both mapped to the CCDV-F task statements and are worth reviewing before your first practice exam.

What score do you need, and what does the score report tell you?

The passing score is 720 on a 100-to-1000 scale. Anthropic does not publish the raw-to-scaled conversion, so we do not state an exact question count as the pass mark. The score report gives you three things: pass or fail, your scaled score, and percent-correct by domain. That domain breakdown is the most actionable output if you need to retake: it tells you exactly which domains to prioritise in a second preparation cycle.

The credential is valid for 12 months from the date it is awarded. As of 3 June 2026, more than 10,000 individuals hold a Claude certification across all tracks, within a partner ecosystem of more than 40,000 applicant firms.

How does CCDV-F compare with the other Claude certification tracks?

The Claude Partner Network currently runs four proctored Pearson VUE exams:

CCDV-F and CCAR-F are the same price but test different skill sets. CCDV-F is the right track if your work centres on writing application code, integrating APIs, and shipping Claude-powered features. CCAR-F is the right track if your work centres on designing multi-agent systems, orchestration patterns, and production architecture. The two tracks are not prerequisites for each other; you can sit either in any order.

Tiered Claude Partner Network partners receive discounted first attempts on all tracks.