Claude Developer Study Guide & Cheat Sheet (CCDV-F)

Ready to practice? Take the free CCDV-F practice quiz →

How this guide was made. I have never seen the CCDV-F exam. I can’t, because Anthropic restricts registration to employees of partner companies with 10 or more staff, and I don’t qualify. Everything below comes from Anthropic’s published exam guide, which is free and public. No exam recall. No dumps. Ever.

Exam facts

Exam code	CCDV-F
Price	$125 USD
Questions	53, the fewest of the four
Time	120 minutes
Passing score	720 on a scaled range of 100 to 1,000
Prerequisites	None
Validity	12 months, then a free open-book renewal
Counts toward partner tier	Yes

Fifty-three questions is worth noticing. Fewer questions means each one carries more weight, and there is less room to absorb a bad domain.

Who it is for

Technical professionals who build, integrate and ship production-grade applications on Claude. AI and ML engineers, technical leads, senior software engineers.

Recommended background: one to five years of software engineering, at least six months hands-on with Claude, proficiency in Python or TypeScript, fluency with REST APIs and CLI tools, and a working understanding of LLM fundamentals, agents, context management and MCP.

The domain breakdown

CCDV-F is the only Claude exam that publishes sub-domain weights, to one decimal place. Treat them as a literal budget.

Domain	Weight	Roughly
Applications and Integration	33.1%	~18 questions
Model Selection and Optimization	16.8%	~9 questions
Agents and Workflows	14.7%	~8 questions
Prompt and Context Engineering	11.0%	~6 questions
Tools and MCPs	10.6%	~6 questions
Security and Safety	8.1%	~4 questions
Claude Code	3.1%	~2 questions
Eval, Testing and Debugging	2.6%	~1 question

One third of this exam is a single domain. If you study nothing else, study Applications and Integration.

And note the other end. Claude Code and Eval together are about three questions. Do not spend a week on Claude Code for this exam. That is Architect: Foundations territory, where it is 20%.

Applications and Integration (33.1%), by sub-domain

Sub-domain	Weight	Scope
Claude Application Design	8.6%	How Claude interprets instructions across Claude Code, Desktop, claude.ai, API and SDKs. Content boundaries, schema design, session hygiene, plugin management
Software Engineering Foundations	7.4%	REST APIs, JSON, async programming, version control, SDLC integration, code review, refactoring
Claude API Mechanics	6.8%	Messages, tools, streaming, vision, thinking, caching, third-party vendors, batch API, and realtime versus batch trade-offs
Configuration Management	4.1%	CLAUDE.md files, settings.json, model version pinning, prompt versioning, plugin dependencies
Understanding Requirements	3.4%	Functional and infrastructure requirements from business requirements
Systems Life Cycle	2.8%	Developing, implementing, operating and maintaining IT systems

Two of the largest sub-domains, Application Design and Software Engineering Foundations, are broad engineering competence rather than Claude trivia. Together they outweigh the entire Agents and Workflows domain.

The other domains, briefly

Model Selection and Optimization (16.8%). LLM fundamentals like tokens, context windows, sampling and non-determinism. Model options including fast mode, extended thinking, adaptive thinking and effort levels. Opus versus Sonnet versus Haiku trade-offs across quality, latency and cost. Token tracking, cost modeling, and caching including cache check-pointing.

Agents and Workflows (14.7%). Decision criteria for a workflow versus an agent. Manager and supervisor hierarchies. Subagents. The Claude Agent SDK, custom agent loops, self-hosted versus Anthropic-hosted deployment, and hooks for deterministic actions. Common patterns and frameworks including Strands, LangGraph and PydanticAI.

Prompt and Context Engineering (11.0%). Instruction clarity, few-shot examples, system versus user placement, output constraints, input sanitization. Context drift and bloat, tool output pruning, compaction, and context isolation through subagents. Structured output patterns, response validation, defensive parsing, and skepticism toward confident output.

Tools and MCPs (10.6%). Tool use and function calling, writing tool descriptions, error handling, client-side versus server-side tools, approval patterns. MCP server authoring and deployment, plus its three primitives: tools, resources and prompts. And the trade-offs among built-in tools, custom tools, Skills and MCPs.

Security and Safety (8.1%). Prompt injection mitigation, jailbreak defense, untrusted input handling, data leakage and PII. Guardrail layering and secure-by-design principles. Secrets and API key management. And a dedicated 1.0% on using hooks as guardrails to prevent destructive actions.

Three patterns worth memorizing

Anthropic publishes three sample questions for this exam. Each one encodes a rule that recurs.

Large volume, not urgent, cost matters: the answer is the Batch API. Fifty percent off input and output, up to 100,000 requests or 256 MB, results within 24 hours. You are not billed for errored, canceled or expired requests.
Untrusted fetched content: the answer has three parts. Treat retrieved content as untrusted, keep it structurally separate from trusted instructions, and gate sensitive actions behind guardrails or hooks. An answer doing only one of the three is the distractor. This is indirect prompt injection, which is a different threat model from a direct jailbreak.
Multiple applications need the same capability: the answer is an MCP server. Cross-application reuse is the MCP trigger. Single app, single purpose, is a plain tool definition.

Cheat sheet: numbers that show up

Stop reasons. Seven of them. The two that need a specific recovery are pause_turn, where a server-tool loop hit its limit and you send the content back to continue, and refusal, where you read stop_details and retry on a fallback model.
Tool names must match ^[a-zA-Z0-9_-]{1,64}$.
tool_choice has four types: auto (default), any, tool, none. disable_parallel_tool_use is a sub-field of that object, not a top-level parameter.
Prompt caching allows a maximum of four breakpoints. TTLs are five minutes and one hour. Reads cost a tenth of base input.
Minimum cacheable prompt length varies by model, and below it caching is silently skipped with no error. Detect it by both cache_creation_input_tokens and cache_read_input_tokens reading zero.
Batch API gives 50% off and results are available when all complete or after 24 hours.

How to prepare

Anthropic’s own advice is blunt and worth following: build at least one Claude application that exercises the API, tools and security practices. This exam rewards having shipped something.

Then allocate by the weights above. Spend a third of your remaining time on Applications and Integration, and skim the two smallest domains.

Free official material: Building with the Claude API is a nine-hour course and the core resource here. Add Introduction to MCP and the tool use docs.

Put it into practice

Free CCDV-F practice questions, weighted to the official blueprint, with a plain-English explanation on every one.

Free practice questions →

Keep reading: The CCAR-F architect guide · All four Claude exams explained · What it costs

HOW TO // AI is not affiliated with or endorsed by Anthropic. CCDV-F, Claude Certified Developer and Claude are trademarks of Anthropic PBC; we reference them descriptively. All content is original and built from Anthropic’s published exam blueprint.