CCDV-F FAQ

Claude Certified Developer costs $125 per attempt, as listed on Anthropic's Partner Academy certifications FAQ (before any Partner Network tier discount). A retake is the full exam fee, and your tier discount applies to retakes too.

53 questions in 120 minutes, a mix of multiple choice and scenario-based multiple response, delivered by Pearson VUE online or at a test center. The FAQ advises planning for about 135 minutes of total seat time.

A scaled score of 720 out of 1000. It is a scaled score, not a flat percent-correct cutoff, so the exact number of correct answers required is not published.

Waits grow with each failed attempt: 14 days, then 30 days, then 90 days after the first, second, and third failed attempts. You can sit the exam up to 4 times per rolling 12 months.

No formal prerequisite - no Claude certification requires another. Registration does require a work email on a Claude Partner Network organization's domain.

It is valid for 12 months from the date you earn it. Free, non-proctored reassessment if renewed on time. Let it lapse and you retake the full proctored exam at full cost. No grace period is stated in the Partner Academy FAQ. Treat the 12-month date as a hard deadline and renew before it - if it lapses, the FAQ says you retake the full exam at the full fee.

Its published blueprint has 8 domains: Applications & Integration (33%); Model Selection & Optimisation (17%); Agents & Workflows (15%); Prompt & Context Engineering (11%); Tools & MCPs (10%); Security & Safety (8%); Claude Code (3%); Eval, Testing & Debugging (3%).

Software engineers building applications on the Claude API.

Because it is stateless: there is no server-side conversation object or session id, so each request must resend every prior turn and tool result Claude needs. The upside is that your application owns the history and can edit it - trimming a stale tool result or summarising an old turn are legitimate techniques.

Transient ones: 429 rate-limit errors, 5xx errors such as api_error and overloaded_error, and connection errors, all with backoff. 400, 401, 403, 404, and 413 are deterministic client-side failures - resending the identical request fails identically, so change the request, credential, or model ID first.

In direct injection the person talking to Claude tries to override its instructions - jailbreaks and role-play framings. In indirect injection the malicious instruction arrives inside third-party content the agent processes for someone else, such as a hidden line in an email, web page, or README. The real user never sees it and is not the attacker.

Claude reads the description first to decide whether and when to call the tool at all. A vague description gives it nothing to choose on; one that states a trigger condition and an explicit "do not use for..." case gives it a clear boundary.

Decide which way to err on borderline cases and say so. "When in doubt, don't flag" favours precision (fewer false positives); "when in doubt, flag" favours recall (fewer misses). Leaving it implicit makes the model choose inconsistently case by case, so name the rule and give at least one concrete borderline example.

Why generation ended, and each value needs different handling: end_turn is a natural finish, tool_use means run the requested tools and continue, and max_tokens means the output was truncated mid-generation. Treat it as the control signal for your loop rather than inspecting the text.