A third of the Claude Certified Developer: Foundations (CCDV-F) exam is one domain: Applications & Integration. If you study evenly across the eight domains in Anthropic's exam guide, you'll underprepare for the part that decides your score.
Here are the weights from the exam guide:
| Domain | Share of exam |
|---|---|
| Applications & Integration | 33.1% |
| Model Selection & Optimization | 16.8% |
| Agents & Workflows | 14.7% |
| Prompt & Context Engineering | 11% |
| Tools & MCP | 10.6% |
| Security & Safety | 8.1% |
| Claude Code | 3.1% |
| Evaluation, Testing & Debugging | 2.6% |
The exam is 53 items in 120 minutes, and you pass at 720 on a 1,000 scale. Check those numbers against the exam guide before you book, since Anthropic can change them.
Three questions follow, two from the biggest domain and one from the second biggest. Each one names the doc page it comes from.
1. Rate limits (Applications & Integration)
A batch worker runs parallel requests against Claude Sonnet 5. Under traffic spikes some requests come back HTTP 429 with an error.type of rate_limit_error and a retry-after header. What should the client do with those requests?
- Retry on a fixed 250 ms delay of the client's own, ignoring the retry-after header
- Sleep for the retry-after value, then retry with exponential backoff to a ceiling
- Cut max_tokens from 4096 to 1024 on every worker request and hold concurrency at 40
- Route the failed requests to a dead-letter queue and replay them in the nightly sweep
Answer: sleep for retry-after, then back off exponentially to a ceiling. A fixed 250 ms retry lands inside the window the header just reserved. Cutting max_tokens frees nothing, because max_tokens doesn't count toward the output-token rate limit. A dead-letter queue is for a poisoned request, and this one is transient. Source: Rate limits.
2. Files API isolation (Applications & Integration)
A SaaS product lets each customer upload PDFs through the Files API and later ask questions about them. The frontend posts the customer's chosen file_id straight into the Messages request. What should the team change to keep customer files isolated?
- Sign each file ID before sending it to the browser and verify the signature on use
- Hold file IDs server-side and upload each tenant's files into its own workspace
- Namespace uploads by tenant in the filename and filter the Files list by prefix
- Set expires_in_seconds to 3,600 on every upload and re-upload on each session
Answer: keep file IDs on the server and give each tenant its own workspace. Any API key with access to a workspace can reach its files, so the workspace is the isolation boundary. Signing IDs stops forgery but leaves every file in the shared workspace readable. A filename prefix is only a naming convention, because access goes by file_id. A one-hour expiry shortens the exposure without closing it. Source: Files API.
3. Model routing (Model Selection & Optimization)
A developer routes support requests with a classifier: routine inquiries go to a smaller model, and complex investigations go to a stronger model to meet a strict quality threshold. The budget assumes most traffic goes to the smaller model. An evaluation shows the classifier labels requests correctly only 70% of the time. What should the developer do before launch?
- Improve the classifier's accuracy before routing requests across the two model tiers
- Send every request to the stronger model and log the classifier's outputs for analysis
- Raise the effort parameter on the smaller model so it absorbs misrouted requests
- Deploy the routing pipeline as designed, since the misrouting cost stays manageable
Answer: fix the classifier first. Routing only pays off when classification is accurate. At 70%, complex requests land on the smaller model and miss the quality bar before any savings show up. Sending everything to the stronger model breaks the budget, and more effort on the small model spends tokens without adding the capability. Source: Building effective agents.
Where these come from
I built a free practice site for this exam and eight others. The CCDV-F set has 105 questions across all eight domains, each with an explanation and a source link like the ones above. You can take five timed questions at a time or a 53-question exam-style form, with no account: CCDV-F practice questions.
Some caveats. I wrote the questions with AI help from Anthropic's public docs and the exam guide, and none are real exam questions. The site is independent and not affiliated with Anthropic. My bank has the same weak spot most people's studying has: Applications & Integration is 33.1% of the exam but only 23 of my 105 questions. Everything is single-answer, while the real exam also uses multiple response.
The exam itself is only open to people at Claude Partner Network organizations, according to Anthropic's Partner Academy FAQ. Check that before you plan around it.
![Real Time Operating Systems — Ch.12: RTOS Case Studies [EN]](https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9f8b1u4hxrwc1uq4gfe2.png)