You are a developer who builds with Claude, and now there is an exam that says so. The Claude Certified Developer, Foundations exam, code CCDV-F, has a clean-looking blueprint and a $125 fee, and almost nothing written about it tells you what the questions actually feel like or where people lose points. Most pages restate the exam guide. You can read the exam guide yourself.
I want to do something different here. I do not hold the CCDV-F. What I have is a back office that runs on more than 30 scheduled Claude Code jobs across the products I build, and a long list of those jobs breaking in the specific ways this exam asks about. When I read the blueprint, almost every domain matched something that had already cost me money, a night’s work or a burst of failure emails. So this is the study plan I would follow, built around those lessons, with practice questions I wrote from them.
On this page
The short answer
The CCDV-F is 53 questions in 120 minutes, multiple choice and multiple response, delivered by Pearson VUE, with a pass mark of 720 on a 100 to 1,000 scale. It costs $125 before any partner discount and the credential lasts 12 months. Right now you can only register if your company is in the Claude Partner Network, using your company email.
To pass it: spend a third of your time on Applications and Integration, because it is a third of the exam. Refresh your general engineering too, because about a fifth of the blueprint is software engineering that has nothing to do with Claude. Build four or five small things rather than reading about them. And sit full-length timed practice exams, because Anthropic retired its own practice exam when certification moved to Pearson VUE, and the guide gives you only three sample questions.
What the exam is, in one table
Everything here is from Anthropic’s official CCDV-F exam guide (v1.0, July 2026) and the Partner Academy certification FAQ, which I read in full in September 2026.
| Item | Fact |
|---|---|
| Credential | Claude Certified Developer, Foundations (CCDV-F) |
| Questions | 53, multiple choice plus multiple response (“select TWO”, “select THREE”) |
| Time | 120 minutes, about 2 minutes 15 seconds per question |
| Pass mark | Scaled 720 on a 100 to 1,000 scale (not 72%) |
| Fee | $125 USD; 50% off for Select, Preferred and Global Premier partners, free for Global Premier through Dec 31, 2026 |
| Validity | 12 months; on-time renewal is a free, non-proctored assessment |
| Retakes | Wait 14, then 30, then 90 days; max 4 attempts in 12 months; full fee each time |
| Who can sit it | People at Claude Partner Network organizations, 18+, company email |
| Recommended background | 1 to 5 years of software engineering, 6+ months hands-on with Claude, Python or TypeScript |
One discrepancy worth knowing before you book: the exam guide says you can reschedule up to 24 hours ahead, but the newer FAQ says 48 hours. I would plan around 48.
The blueprint, read by someone who ships this stuff
| Domain | Weight | About this many questions |
|---|---|---|
| Agents and Workflows | 14.7% | 8 |
| Applications and Integration | 33.1% | 17 to 18 |
| Claude Code | 3.1% | 2 |
| Eval, Testing, and Debugging | 2.6% | 1 |
| Model Selection and Optimization | 16.8% | 9 |
| Prompt and Context Engineering | 11.0% | 6 |
| Security and Safety | 8.1% | 4 |
| Tools and MCPs | 10.6% | 6 |
Three things jumped out at me.
First, the headline topic is not the heavy one. Claude Code is 3.1% as a domain, roughly two questions. But its pieces are scattered through the rest: CLAUDE.md and settings.json sit under Configuration Management (4.1%), hooks under Security and Safety (1.0%), and the choice between Skills, MCP and custom tools under Agentic Customization (4.1%). If you use Claude Code every day, you are ahead on more of the exam than the Claude Code line suggests.
Second, the exam wants an engineer, not a prompt writer. Software Engineering Foundations (7.4%), Technical Fundamentals (6.1%), Understanding Requirements (3.4%) and Systems Life Cycle (2.8%) add up to 19.7%. That is REST, JSON, async, version control and code review, about ten questions a Claude-only study plan walks straight past.
Third, evals are one question. Eval, Testing, and Debugging is 2.6%. It matters in real life far more than that, but on this exam a week on eval frameworks buys you one item.
Every domain I would study, and what it already cost me
This is the part no exam guide can give you. For each area, here is what the blueprint tests and the real incident from my own setup that taught me the same lesson. None of these was an agent doing something dangerous. Every one was the layer around the agent: the scheduling, the retries, the exit codes, the limits. I wrote up the full set in Claude Code best practices from running it unattended; here they are mapped to the exam.
Model Selection and Optimization (16.8%). In September 2026 an SEO audit job fanned out into 30 Opus subagents and drained my entire five-hour usage window in about 15 minutes, which killed the article job scheduled after it. The fix was not a cleverer prompt. A plain script now does the checks in seconds, and a smaller model, Sonnet, only reads the output and writes the report. The exam version of this lesson: pick the smallest model that can do the job, and never pay a model to do deterministic work that code can do. Study the model comparison, effort levels and prompt caching, and if you want the real cost picture behind these tradeoffs, I broke down what I actually pay for Claude Code across eight products.
Applications and Integration: error handling (inside the 33.1%). A few days earlier, my job supervisor kept relaunching a run that had hit its session limit. One day produced 13 launches and 9 failure emails, all retrying against a limit that was not going to reset in the next hour. The rule I adopted: never retry an error that will fail the same way next time. The API version is on Anthropic’s errors page: 429 and 529 are worth backing off and retrying, a 400 or 401 is not, and a 429 caused by a spend cap has no retry-after header and keeps failing until access resumes.
Security and Safety: hooks and exit codes (1.0% plus 2.3% guardrails). A guardrail script of mine exited with code 1 on purpose after sending an alert. The supervisor read that as a crash and reran it 23 times in 8 hours: 46 emails in one night. My rule since then is that any job that alerts must exit 0. Claude Code hooks have the same trap in reverse, and the exam likes it: for a PreToolUse hook, exit code 2 is the one that blocks the action, while exit code 1 is a non-blocking error and the tool call goes ahead. Read the hooks reference until that is automatic.
Claude Code: headless mode and sessions (3.1% plus Configuration Management 4.1%). One overnight run edited two live articles, then ended its turn while it thought it was waiting on something in the background. The headless process exited cleanly, and the morning email said zero articles published. Three days later a run published and verified an article, lost its connection, and again reported that nothing had happened. The lesson in both: an unattended run needs to report from what it actually did, and a dropped run should resume the same session rather than start again. Know headless mode, the CLAUDE.md hierarchy and settings.json scopes cold. My CLAUDE.md conventions file is a worked example, and the free CLAUDE.md analyzer will score yours line by line.
Security and Safety: secrets (1.6%). I keep every key in one local secrets file and run commands through a small injector, so the agent finds a secret by its name and description and code puts the value into the process. The key never enters the model’s context. That one decision is what lets me leave automations running without supervising them. It is exactly what Identity, Secrets, and Key Management is testing.
Security and Safety: approvals. Once, during an SEO task, a subagent made production writes and deploys nobody had asked for. Nothing broke, but the standing rule now is explicit confirmation before any production write or deploy. On the exam, the right answer to “how do I make sure the agent asks first” is almost always something enforced in code, a permission rule, a hook or a check in the tool handler, not a stronger sentence in the prompt.
Tools and MCPs (10.6%). I run Murkuz, a Search Console connector for Claude, whose MCP server holds 92 tools while customers see 32. When I trimmed the list for launch, the trim applied to every caller, and my own blog jobs started failing because the tool they called no longer existed for them. The fix was to decide which tools each account sees at the moment of the call. That is tool set construction and MCP server design, two things this domain covers. The MCP build-a-server guide is the fastest way in if you have never written one.
Agentic Customization (4.1%). I write five blogs with Claude, and I keep five separate Skills, one per blog, each with its own voice rules and banned phrases, because one shared prompt flattened five voices into one. When the exam asks whether something should be a Skill, an MCP server, a custom tool or a CLAUDE.md line, this is the reasoning it wants: Skills load on demand, CLAUDE.md loads every session, MCP reaches external systems. If you want real Skills to study, I publish 23 of mine free at locul.ai/skills.
Three practice questions I wrote from those incidents
These are my own questions in the exam’s style, not real exam items. The candidate agreement makes live exam content confidential, so anyone selling “real CCDV-F questions” is selling something they should not have.
Question 1 (Model Selection and Optimization). Select TWO.
A nightly Claude Code job runs a 7-point SEO audit across 14 sites. It spawns one Opus subagent per check per site and regularly exhausts the team’s usage limit before other scheduled jobs run. Which two changes most reduce cost without losing the report?
- A. Raise the effort level so each subagent finishes in fewer turns
- B. Move the deterministic checks into a script and have the model only read its output
- C. Run all the subagents in parallel so the job finishes sooner
- D. Use a smaller model such as Sonnet or Haiku to write the report from the script’s output
- E. Add “be efficient with tokens” to the system prompt
Answer: B and D. The waste is paying a large model to do work code can do, many times over. Parallelism (C) finishes sooner but burns the same tokens faster. Higher effort (A) usually spends more, not less. A prompt instruction (E) is a request, not a control.
Question 2 (Applications and Integration)
A scheduled job calls the Messages API and gets a 429 rate_limit_error with no retry-after header. The response indicates the organization has reached its spend cap. What should the job do?
- A. Retry immediately in a loop until it succeeds
- B. Retry with exponential backoff for up to 24 hours
- C. Stop, record the run as blocked by a spend limit, and alert a human
- D. Switch to the Message Batches API, which is not subject to spend limits
Answer: C. Anthropic documents that a spend-cap 429 keeps failing until access resumes, so retries only generate noise, as my 13 launches in one day proved. Batches (D) are cheaper, not exempt from limits. Backoff (B) is right for an ordinary 429 with a retry-after, not this one.
Question 3 (Security and Safety). Select TWO.
A team uses Claude Code to prepare releases. It must never run the production deploy command without a human’s go-ahead. Which two controls enforce that reliably?
- A. A line in CLAUDE.md saying “never deploy without asking”
- B. A
denypermission rule for the deploy command in settings.json - C. A PreToolUse hook matched on Bash that blocks the deploy command by exiting with code 2
- D. Lowering the temperature so the model behaves more predictably
- E. A system prompt that lists the consequences of a bad deploy
Answer: B and C. Both are enforced by Claude Code itself, whatever the model decides. A and E are instructions the model can be talked around or forget in a long session; D does nothing for permissions. Sources: permissions and the hooks guide.
What people on Reddit get wrong about this exam
I read through the threads where people discuss the Claude certifications. Five points come up again and again.
“I’ll just sign up and take it.” You cannot, unless your employer is in the Claude Partner Network. Registration rejects personal email addresses, and linking an extra company domain takes Anthropic 7 to 10 days. Check this before you study, not after.
“The mocks were nothing like the real thing, and two hours was not enough.” That complaint shows up about the Associate exam, where 60 items share the same 120 minutes. The CCDV-F is more generous at about 2 minutes 15 seconds a question, but only if you have practised at that pace. A practice set of short quizzes does not train it; full-length timed exams do.
“I have the Anthropic Academy certificate, so I’m certified.” The free courses at academy.claude.com give completion certificates. They are good, and they are not the proctored credential.
“It’s mostly about the API.” Partly true. A third of the exam is Applications and Integration, and people who passed describe many questions about how you would structure real API calls. But “the API” here includes streaming, batching, caching, error recovery and configuration, not just messages.create.
“Is it even worth it?” If your company is a partner, yes: it counts toward the firm’s tier, the fee is often covered, and the blueprint drags you through parts of the platform you would otherwise skip. If you cannot register yet, learn the material anyway. Everything in the plan below made my own systems cheaper and more reliable before any exam existed.
My 3-week study plan
This assumes you already build with Claude and can give it about 8 hours a week. The idea is simple: build first, then practise under the clock, and study only what the practice exams show you are missing.
Start with a baseline. PrepClubs has a free 25-question CCDV-F diagnostic that samples all eight domains at the official weights in 57 minutes, and it tells you your two weakest domains before you spend a single study hour.
CCDV-F: 3-WEEK PLAN (~8 hrs/week)
DAY 1
[ ] Confirm company is in the Claude Partner Network + my email domain is recognized
[ ] Read the exam guide end to end, including the 3 sample questions
[ ] Free 25-question diagnostic -> write down my 2 weakest domains
WEEK 1: BUILD (covers ~60% of the blueprint)
[ ] Build 1: a CLI that streams a Claude answer and calls one tool
[ ] Build 2: the same prompt run through the Message Batches API; compare cost
[ ] Build 3: add prompt caching; confirm cache reads in the usage fields
[ ] Build 4: a tiny MCP server with one tool, connected to Claude Code
[ ] Build 5: a PreToolUse hook that blocks one dangerous command (exit 2)
[ ] Refresh: REST, JSON, async/await, git branching, code review etiquette
WEEK 2: PRACTISE UNDER THE CLOCK
[ ] Full-length timed exam (53 q / 120 min) x 3
[ ] After each: log every miss in the mistake log below
[ ] Re-read the docs page behind every miss the same evening
[ ] Drill multiple-response items: read "select TWO/THREE" before the options
WEEK 3: CLOSE THE GAPS, THEN BOOK
[ ] Model tiers, effort, adaptive thinking, token budgeting
[ ] Context bloat: tool-result clearing, compaction, subagents for isolation
[ ] Prompt injection: untrusted input, approvals enforced in code
[ ] Headless mode, CLAUDE.md hierarchy, settings.json scopes, Skills vs MCP
[ ] 2-3 more timed exams; book when two in a row clear ~80%
[ ] 48h before: Pearson system test, ID name matches registration exactly
And the mistake log, which is the single most useful habit I know for any multiple-choice exam. Copy it into a spreadsheet:
date | exam # | question topic | domain | my answer | right answer | why I missed it (knowledge / misread / time) | docs page to re-read
The “why I missed it” column is the point. If most of your misses are misreads, you need slower reading on multiple-response items, not more studying. If they are knowledge gaps clustered in one domain, that is where your next evening goes.
Where I would practise, and why
For the timed practice in week 2, I would use PrepClubs’ Claude Certified Developer practice exams. I should be clear that PrepClubs is one of my products, so read this with that in mind. The reason I would point you there anyway is what the bank is: 10 full-length exams at 53 questions and 120 minutes, 555 original questions written against the v1.0 exam guide, every exam weighted to the official blueprint within one question per domain, and about one item in five a “select two or three” question that says so, as the real exam does. Roughly a quarter of the questions put a request body, tool schema, settings.json file, hook or MCP config in front of you. Every explanation says why each wrong option is wrong and names the Anthropic docs page it relies on, and each exam ends with a percent-correct-by-domain breakdown like Anthropic’s own score report.
The honest test of any prep bank is whether it makes you sit the whole exam against the clock. Short quizzes feel productive and train nothing about pacing. Whatever you use, make sure it does that.
FAQ
Is the Claude Certified Developer certification worth it?
For engineers at a Claude Partner Network company, yes: it counts toward the firm’s partner tier, the fee is modest and often covered, and preparing for it forces you through batching, caching, hooks and MCP properly. For individuals who cannot register yet, the knowledge is worth it and the credential can wait.
How hard is the CCDV-F?
It is practical rather than academic. Most questions describe a build decision and ask for the best approach, so daily hands-on use is the biggest advantage you can have. One widely shared review of all four Claude exams ranked the Developer exam second of four for technical difficulty. The points people lose tend to be on multiple-response items, cost and caching details, and the general engineering questions.
How long should I study for the Claude Certified Developer exam?
If you already build with Claude weekly, three weeks at about 8 hours a week is realistic. If you have used Claude only through chat, give it six weeks and do the builds in week 1 slowly.
Can I take the CCDV-F without working at an Anthropic partner?
Not currently. Certification is open only to organizations in the Claude Partner Network, and registration needs a company email on a recognized domain. Joining the network is free for a company.
What is the passing score for the Claude Certified Developer exam?
A scaled 720 on a 100 to 1,000 scale. Anthropic set it through a standard-setting study and does not publish a raw-score conversion, so treat it as a comfortable 80% in practice rather than 72%.
Is there an official CCDV-F practice test?
No. Anthropic’s FAQ says the practice exam on the previous platform was retired in the move to Pearson VUE. The exam guide has three sample questions with rationale, and Partner Academy has a free 5-course prep path for partners.
How is the Claude Certified Developer different from the Claude Certified Architect?
The Developer exam is for people who write the code: API, agents, tools, Claude Code, cost. The Architect Foundations exam (60 questions, scenario-based) is for people designing the system and owning tradeoffs across teams. If you are not sure, take the Developer exam first. For all four Claude certifications side by side, including the Associate and Architect Professional, this comparison of the four Anthropic exams lays out who each one is for.
Where this comes from
I build and run eight products on my own, and my back office runs on more than 30 scheduled Claude Code jobs, from daily articles across several sites to content upkeep and SEO audits. Most of the incidents in this article happened in that setup in September 2026, and every rule above came out of it. Before that I spent about two and a half years in cybersecurity presales at GFI Software, running international webinars and training reseller and distribution partners, so I have taught a lot of material to people facing an assessment without holding the certificate myself. I do not hold the CCDV-F. This is the plan I would follow to pass it, and the reasoning I would want in my head on exam day.