Skip to main content

Claude Certified Developer — Foundations (CCDV-F)

Study site for the Claude Certified Developer – Foundations certification, built against Exam Guide v1.0 (effective July 2026).

The credential validates that a developer can build, integrate, and ship production-grade applications, agents, and workflows on the Claude platform — through the Claude API and client SDKs, the Claude Agent SDK, Claude Code, custom tools, and MCP servers.

Where CCAR-F tests architectural judgment across scenarios, CCDV-F tests implementation mechanics: which parameter, which endpoint, which stop reason, which layer. Items are shorter, more numerous per domain, and far more likely to hinge on an exact API shape.

Exam details at a glance​

AttributeValue
CredentialClaude Certified Developer – Foundations
Exam codeCCDV-F
Number of items53
Item formatMultiple-choice and multiple-response; each item states how many responses to select
Time limit120 minutes (~2.3 min/item)
DeliveryProctored — online proctored and/or Pearson VUE test center
Passing scoreScaled score of 720 on a 100–1,000 scale
Exam fee$125 USD
Validity12 months from the award date
Result reportingPass/fail with scaled score, plus percent-correct by domain

Registration, retakes, ID rules, and recertification: Exam logistics & policies.

Blueprint and domain weights​

Eight domains, 25 skills. The weights are lopsided — Domain 2 alone is a third of the exam, and Domains 2 and 5 together are half of it.

#Content domainWeightApprox. itemsSkills
1Agents & Workflows14.7%~83
2Applications & Integration33.1%~186
3Claude Code3.1%~21
4Eval, Testing & Debugging2.6%~11
5Model Selection & Optimization16.8%~94
6Prompt & Context Engineering11.0%~63
7Security & Safety8.1%~44
8Tools & MCPs10.6%~53
Total100%5325

Skill weights, ranked​

The exam is written against skills, not domains. Ranked by share of scored items, the top eight skills are 47% of the exam:

SkillWeightPage
Claude Application Design8.6%2.5
Software Engineering Foundations7.4%2.4
Claude API Mechanics6.8%2.3
Technical Fundamentals6.1%5.2
Agent Construction with Claude5.3%1.2
LLM Fundamentals5.2%5.1
Agent Patterns & Frameworks4.9%1.3
Prompt Engineering4.6%6.2
Agent Architecture4.5%1.1
Tool Implementation4.4%8.1
Agentic Customization4.1%8.3
Configuration Management4.1%2.6
Context Engineering3.8%6.1
Understanding Requirements3.4%2.1
AI Application Security3.2%7.1
Claude Code Operation3.1%3.1
Cost & Token Management2.8%5.4
Systems Life Cycle2.8%2.2
Model Selection & Tradeoffs2.7%5.3
Debugging & Error Handling2.6%4.1
Output Handling2.6%6.3
Guardrails & Safe Deployment2.3%7.2
MCP Server Development2.1%8.2
Identity, Secrets & Key Management1.6%7.4
Claude Hooks1.0%7.3
Where the study time goes

Domain 2 plus Domain 5 is 49.9% of the exam and both are mechanics-heavy — API shapes, SDK behaviour, tokens, caching, cost. Domain 3 (Claude Code) is only 3.1%: worth one careful pass, not a week. Domain 7 is 8.1% split across four thin skills, so breadth beats depth there.

Intended audience​

The minimally qualified candidate is a hands-on engineer who ships Claude-powered systems. Recommended experience per the guide:

  • One to five years in software engineering
  • At least six months hands-on with Claude or comparable LLM-based systems
  • Proficiency in Python and/or TypeScript
  • Fluency with REST APIs and CLI tools
  • Working understanding of LLM fundamentals, agents, context management, and MCP

There are no mandatory prerequisites or courses. The credential is awarded on exam performance alone.

Explicitly out of scope as an audience: non-technical users, people without hands-on development experience, and roles limited to prompt writing without broader application-development responsibility.

The seven developer decision rules​

Most items are "which implementation is correct / most appropriate". Seven rules resolve the bulk of them.

SituationAnswer
A loop must know when to stopBranch on stop_reason — tool_use continues, end_turn terminates, max_tokens is truncation, refusal needs stop_details. Never parse prose for completion.
The workload is large and latency-tolerantMessage Batches API — 50% cost, results within 24 h, keyed by custom_id. Not parallel synchronous calls; not a smaller model.
The same big prefix is resent on every requestPrompt caching — stable content first, cache_control breakpoint after it, verify with usage.cache_read_input_tokens.
The model picks the wrong toolRewrite the tool description — purpose, inputs, boundaries, when not to use it. Not a routing classifier.
Output must be machine-consumableStructured outputs — output_config.format, or strict: true on the tool schema. Then still validate defensively before use.
A rule must hold every single timeA deterministic gate in code — a hook, a prerequisite check, a server-side authorization check. Never a stronger system prompt.
Content came from outside your trust boundaryTreat it as untrusted data, delimited and labelled, with least-privilege tools. Injected instructions must be unable to reach a sensitive action.

Underneath all seven: put the guarantee in the layer that can actually guarantee it. Prompts shape probability; code enforces invariants.

Distractor smells: raise the temperature · ask politely in the system prompt · train a classifier · use the model's self-reported confidence as a gate · switch to a bigger model instead of fixing the layer that is wrong · a parameter or flag that does not exist (budget_tokens on a current model, output_format, assistant prefill).

What to build before the exam​

The guide's own preparation advice is to build and operate at least one Claude application end to end. Concretely:

FocusWhat to actually doDomains
A custom agent loopHand-write the while stop_reason === "tool_use" loop with parallel tool execution, is_error results, and retry on transient failures. Then rebuild it with the SDK tool runner and compare what you gave up.1, 8
API mechanics drillOne script that exercises streaming, vision, PDF input, adaptive thinking, prompt caching, structured outputs, token counting, and the Batches API against the same prompt. Read usage on every response.2, 5
An MCP serverExpose 3–4 tools over stdio, plus one resource and one prompt. Connect it to Claude Code and to the API MCP connector.8
Claude Code configurationCLAUDE.md hierarchy, settings.json, a custom slash command, a skill, a subagent, and one hook that blocks a destructive command. Run the same task in headless mode.3, 7
Cost and token accountingModel the per-request cost of a real workload from usage, then cut it with caching, effort, and batch — and measure the difference.5
Failure injectionForce a 429, a 529, a schema-invalid response, a truncated max_tokens response, and a prompt injection. Write the recovery path for each.4, 7

Sample questions​

The guide publishes three sample items (Domains 2, 7 and 8). All three are in the mock exam bank and on the matching skill pages, labelled Official sample question in their rationales:

SampleDomainTestsPage
12 — Applications & IntegrationBatch vs realtime for a latency-tolerant 10,000-document job2.3 Claude API mechanics
27 — Security & SafetyPrompt injection from user-submitted web pages7.1 AI application security
38 — Tools & MCPsMCP server for a reusable internal REST capability8.2 MCP server development

Last-minute assets​

  • Cheat sheet — domain-filterable cram entries: limits, parameter names, and syntax.
  • Mock exam — 53 items, 120-minute countdown, scaled scoring against the 720 cut score, per-domain breakdown.