LLM Certification: The CCDV-F Developer Exam Study Guide

If you are searching for an LLM certification that tests real production skills rather than trivia, the Claude Certified Developer, Foundations exam (CCDV-F) is the most technically demanding entry-level credential in the Claude Partner Network. Launched 12 March 2026 and priced at $125 per attempt, it covers eight domains across 53 scenario-based items in 120 minutes. The pass mark is a scaled score of 720 on a 100-to-1000 scale. This guide walks through every domain, the questions candidates are actually asking, and the most efficient path to that score.

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

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

The domain weights are the single most important input to your study plan. Unlike many exams that round weights to the nearest five percent, the CCDV-F publishes fractional percentages, which signals that Anthropic is deliberate about emphasis.

Domain 2 (Applications and Integration) alone accounts for 33.1% of the exam. Combined with Domain 5 (Model Selection and Optimisation, 16.8%), those two domains represent roughly half the test. A candidate who masters those two areas and achieves average performance everywhere else is already close to the 720 pass mark.

Domains 3 and 4 together total only 5.7%. That does not mean you skip them entirely; a single hard item in a low-weight domain can still swing a borderline score. It does mean you should not spend equal time on every domain.

What does Domain 2 (Applications and Integration) actually test?

Domain 2 is the exam's centre of gravity at 33.1%. It tests how you build, integrate, and operate Claude-powered applications in realistic production contexts.

Expect scenario questions around:

• Choosing between synchronous API calls and the Message Batches API for high-volume, latency-tolerant workloads.
• Structuring multi-turn conversations correctly, including how to maintain full conversation history when the API is stateless.
• Integrating Claude into existing software stacks: webhooks, queues, and streaming responses.
• Handling rate limits, retries, and back-off strategies without cascading failures.
• Designing output schemas that downstream systems can parse reliably.

The exam consistently rewards candidates who think about the full request-response lifecycle. Our Messages API Request-Response Cycle concept covers the mechanics that underpin most Domain 2 scenarios.

One pattern that trips up many candidates: confusing a valid empty result with an access failure. Both return without an exception, but they require completely different handling logic. The exam will present you with a scenario and ask which response is correct.

How much of the exam is about MCPs, tools, and agent orchestration?

More than a quarter of the exam, when you add the relevant domains together. Domain 8 (Tools and MCPs) is 10.6%, Domain 1 (Agents and Workflows) is 14.7%, and portions of Domain 2 overlap with integration patterns that involve tool use. Together these areas reward candidates who understand how Claude selects tools, how errors propagate through multi-agent pipelines, and how the Model Context Protocol structures server-client communication.

Key concepts to master for this cluster:

• How tool descriptions function as the primary selection mechanism. A poorly worded description causes misrouting; the fix is almost always a description rewrite, not a system-prompt patch.
• The isError flag in MCP responses and when to set it versus returning a structured error payload.
• Agentic Loop Anti-Patterns: premature termination, infinite loops, and the conditions that trigger each.
• Parallel Subagent Spawning: when it reduces latency, and when it creates attribution and synthesis problems.

The exam does not test MCP configuration syntax as rote recall. It presents a broken integration and asks you to identify the root cause and the proportionate fix. Candidates who have built real MCP servers will recognise these patterns immediately; those who have only read documentation will need to work through practice scenarios carefully.

What should I know about model selection and optimisation (Domain 5, 16.8%)?

Domain 5 is the second-largest domain and the one where candidates most often lose marks they should not. The core skill is matching model capability to task requirements and cost constraints, not memorising benchmark numbers.

The exam tests three broad judgement calls:

1. Haiku vs Sonnet vs Opus

The exam does not ask you to recite context window sizes. It presents a production scenario with latency, cost, and quality constraints and asks which model family is appropriate. The heuristic the exam rewards: use the smallest model that reliably meets quality requirements for the task. Haiku is the right answer for high-volume, low-complexity classification or extraction. Sonnet handles most reasoning and generation tasks. Opus is warranted only when the task genuinely requires extended reasoning and the cost is justified.

2. Sampling parameters

Candidates sometimes arrive expecting questions about temperature, top_p, and token-based thinking budgets as they existed in older API versions. The exam tests current API patterns. Know what parameters are available, what they control, and when adjusting them is the right lever versus when the problem is actually a prompt or schema issue.

3. Prompt caching and cost optimisation

For workloads with large, repeated system prompts, prompt caching can reduce both latency and cost substantially. The exam will present a scenario where caching is the correct answer and ask you to identify it.

What kinds of scenario questions appear around security, safety, and prompt injection?

Domain 7 (Security and Safety) carries 8.1% of the exam. The scenarios here are less about Anthropic's internal safety research and more about the practical security posture of production systems.

Prompt injection is the most commonly tested attack vector. The exam presents a scenario where user-supplied input is concatenated into a prompt that also contains privileged instructions, and asks you to identify the vulnerability and the correct mitigation. The answer is almost always structural: separate untrusted input from trusted instructions at the architecture level, not by filtering keywords.

Other Security and Safety scenarios cover:

• Validating tool outputs before passing them back into the context window.
• Designing human-in-the-loop checkpoints for high-stakes agentic actions. The exam rewards High-Stakes Enforcement Decision Rules: when the cost of an irreversible action is high, require explicit human confirmation rather than relying on model judgment alone.
• Recognising when a model's refusal is appropriate versus when it signals an overly restrictive system prompt.

The exam does not ask you to implement cryptographic controls or write firewall rules. It tests whether you can design Claude-based systems that are robust to adversarial inputs and that escalate appropriately when they encounter situations outside their authorised scope.

How does Domain 6 (Prompt and Context Engineering, 11.0%) differ from what I already know?

If you have used Claude extensively, you likely have intuitions about prompt engineering that are mostly correct. Domain 6 tests whether those intuitions hold up under production constraints.

The exam distinguishes between:

• Goal-based prompts that describe the desired outcome and let the model determine the steps, versus step-based prompts that enumerate a procedure. The exam tests when each is appropriate. See our Goal-Based vs Step-Based Prompts concept for the decision framework.
• Few-shot examples as a calibration tool. The exam rewards candidates who know that few-shot is the highest-leverage technique for ambiguous edge cases and extraction quality, and who can identify when adding examples is more effective than rewriting instructions.
• Context window management under real constraints. The Attention Dilution Problem is a tested concept: as context grows, model attention to early instructions weakens. The exam presents scenarios where this is the root cause of degraded output and asks for the correct architectural response.

Domain 6 also overlaps with Domain 2 in structured output design. Candidates who understand how schema design prevents fabrication and how to write output format instructions that survive edge cases will find these questions straightforward.

What are the exact exam logistics for CCDV-F?

One important structural difference from the Architect exam (CCAR-F): CCDV-F does not draw from a scenario bank. Each sitting presents items written directly against the domain skills. This means you cannot predict which specific scenarios will appear, but it also means thorough domain coverage is the most reliable preparation strategy.

The score report you receive after the exam shows pass or fail, your scaled score, and percent-correct by domain. That domain breakdown is valuable: if you sit the exam and do not pass, it tells you exactly where to focus before your next attempt.

How should I structure my preparation to avoid wasting time?

Given the domain weights, a rational study allocation looks like this:

1. Start with Domain 2 (Applications and Integration, 33.1%). This is non-negotiable. Build or review at least one end-to-end Claude application that handles streaming, error states, and multi-turn conversation correctly.

2. Move to Domain 5 (Model Selection and Optimisation, 16.8%). Work through model selection scenarios until the Haiku/Sonnet/Opus decision feels automatic. Understand prompt caching mechanics.

3. Cover Domain 1 (Agents and Workflows, 14.7%) and Domain 8 (Tools and MCPs, 10.6%) together. These domains share conceptual infrastructure. Tool descriptions, error propagation, and agentic loop design all appear in both.

4. Study Domain 6 (Prompt and Context Engineering, 11.0%). If you have production Claude experience, this domain will feel familiar. Focus on the edge cases: few-shot calibration, context window management, and structured output reliability.

5. Allocate remaining time to Domain 7 (Security and Safety, 8.1%), Domain 3 (Claude Code, 3.1%), and Domain 4 (Eval, Testing, and Debugging, 2.6%). Do not skip these; do not over-invest.

Our platform's adaptive engine uses Bayesian Knowledge Tracing with a 0.90 mastery threshold, which means it will automatically surface the concepts where your probability of mastery is lowest. For CCDV-F, the practice exams mirror the real format: 53 questions, scored 100 to 1000 with 720 as the pass mark. You can track your domain-level readiness before you book the real sitting.

For candidates also considering the Architect track, our Agentic Architecture and Tool Design and MCP Integration concept libraries cover the CCAR-F domains in depth.

Is CCDV-F the right LLM certification for me, or should I sit CCAR-F instead?

The two credentials test different skill profiles.

If your day-to-day work involves writing application code that calls the Claude API, integrating Claude into existing systems, or selecting models for production workloads, CCDV-F is the more directly relevant credential. If your work involves designing multi-agent orchestration, defining system boundaries, and making architectural decisions about how Claude components interact, CCAR-F is the better fit.

Both credentials are valid for 12 months from the award date and are part of the Claude Partner Network, a $100M programme that as of 3 June 2026 had more than 40,000 partner applicant firms and 10,000 certified individuals.