Notes from passing the Claude developer exam

I passed Anthropic's Claude Certified Developer – Foundations (CCDV-F) exam with 896 out of 1,000 after four days of preparation. The pass mark is 720. There are no exam questions in this post. It covers how the exam is weighted and what kind of answer it tends to reward.

The exam has 53 questions and a 120-minute limit, and it's scored from 100 to 1,000. The score report breaks your result into 25 test objectives across eight domains. The official exam guide publishes a weight for each domain, so read it first and plan your time around those numbers.

Applications and Integration is the biggest domain at 33.1%. After it come Model Selection and Optimization (16.8%), Agents and Workflows (14.7%), Prompt and Context Engineering (11%), Tools and MCPs (10.6%) and Security and Safety (8.1%). Claude Code gets 3.1%, and Eval, Testing, and Debugging gets 2.6%.

You don't write any code, and you don't need syntax, parameter names or prices. Each question describes a situation and asks what a developer should do. Models are described by role, such as the fast, low-cost option or the strongest reasoning tier, instead of by name. Time spent memorizing API fields is better spent learning when each feature is the right call.

Many questions have two answers that both sound right, and the better one often puts the steps in a more sensible order: settle the requirements before picking a model, find the cause before applying a fix, and test a new model on your own workload before switching to it. If you can't choose between two options, ask which step should happen first.

Answers that include a check tend to beat answers that don't. "Pick the cheapest model tier that meets the bar" is a decent answer. "Pick the cheapest tier that meets the bar and keep evaluating it" is a better one. Watch for options that are correct but skip validation.

If something has to happen every time, like a check before a command runs, a fixed output format or a human approval, put it in a hook, a permission rule or an approval step. Asking for it in the prompt isn't enough. When a scenario says the model "occasionally skips" a step, the answer is usually to enforce that step in code.

When an agent's history grows too long, give it less to carry: prune old tool results, compact older turns, or hand self-contained work to subagents that have their own context. A bigger context window only delays the problem.

Much of Applications and Integration is general software engineering. Alongside the Claude API, it covers requirement types (functional, non-functional and infrastructure), the system life cycle, version control and code review, and configuration and secrets management. To sort a requirement, try deleting it: if the system still works, just worse, it was non-functional.

Some Claude API rules come up often enough to learn well. If nobody is waiting on the result, use the Message Batches API, which runs requests asynchronously at half the price. If someone is waiting, make a normal request and stream long responses. Prompt caching reuses an identical prefix of the prompt, so it lowers the cost of repeated input but never of output. The API doesn't remember earlier turns or include claude.ai's built-in instructions, so you send the conversation history and the system prompt yourself. Use a workflow when the steps are fixed, and an agent when you can't predict them.

My advice for preparing: start with the exam guide and its weights, practice scenario questions rather than memorizing facts, and review your mistakes by objective instead of by total score. Much of my four days was hands-on, in a playground I built for the Claude API, the Agent SDK, Claude Code and MCP, organized around the exam's objectives.

The certification track from that playground is now a free study site, CCDV-F Study Lab, at johncarl-30.github.io/claude-code-playground. It has knowledge checks, mock exams, a mistakes deck and a study plan, and you don't need an API key. I wrote the questions from the official docs, and none of them come from the real exam.

Good luck if you're taking it soon. If you try the Study Lab and spot a question that's wrong or out of date, tell me and I'll fix it.