Frequently asked questions

Anthropic's foundational certification for engineers who build with Claude. It tests practical knowledge across five domains: agentic architecture and orchestration (27%), tool design and MCP (18%), Claude Code configuration and workflows (20%), prompt engineering and structured output (20%), and context management and reliability (15%). It's scenario-driven, not trivia-based — most questions describe a production situation and ask you to pick the best move.

Engineers and AI builders who are shipping (or about to ship) products that use Claude — directly via the API, through the Agent SDK, with Claude Code, or via MCP. If your day job touches LLM prompts, tools, agents, or any of the surrounding plumbing, the exam validates that you understand the production patterns Anthropic recommends.

No formal prerequisite. In practice you'll do better if you've shipped at least one small project that calls Claude, written a tool the model invokes, or used Claude Code on a real codebase. The exam is designed to test working knowledge, not theory.

60 questions in 90 minutes. Scoring is weighted by domain on a 1000-point scale; 720 is the passing line. Our mock exam mirrors the weighting and pass threshold so you can calibrate before sitting the real test.

Anthropic runs the official cert through Skilljar at https://anthropic.skilljar.com — there's a deep link in our header (Official exam ↗). The cert is separate from this prep site; we're not affiliated.

Foundations is the starting tier. Anthropic has signaled more specialized certifications over time. We'll add coverage here as new tiers come out and link to them from the dashboard.

No. claudecert.com is an independent community resource. We're not affiliated with, endorsed by, or sponsored by Anthropic. The official source of truth for the exam and its objectives is Anthropic itself.

Yes. Every lesson, the diagnostic, drill mode, the audiobook, the mock exam, and the 14-Day Plan are free with no signup required. You don't need to create an account to use any of it.

Locally in your browser's localStorage under keys prefixed ccp.* — your completed lessons, drill state, mock attempts, voice preference, and 14-day plan all live there. Clearing site data resets your progress.

Cross-device sync is on our roadmap behind optional sign-in. The Auth.js scaffolding is in place; we'll switch it on once we wire a provider. Today, progress lives per-browser.

No third-party trackers, no analytics scripts, no ads. Server logs (HTTP method, path, status, IP) are retained for 14 days for debugging purposes only.

Most pages will load once cached, but the diagnostic, drill, mock exam, and audiobook depend on local state plus (for the audiobook and read-aloud) a network call to ElevenLabs. Treat the site as online-first.

Continuously. Every code or content change lands an entry on the public /changelog page. Major Anthropic releases (new model versions, MCP changes, Claude Code updates) get a journal post and curriculum updates within days.

Take the 15-minute diagnostic first (/learn/diagnostic). It surfaces your two weakest domains and links straight into them. From there, either follow the 14-Day Plan for a structured cadence, or jump around the curriculum on your own.

Most engineers who've shipped with Claude before are ready in 8–15 hours of focused study. People newer to Claude typically need 20–30 hours. The 14-Day Plan caps you around an hour a day to fit into a working week.

It builds an initial schedule weighted by exam domain percentages, then rebalances every visit based on your current BKT mastery — weak Knowledge Points float to the front, mastered ones drop out. Day 7 is a mini mock; Day 14 is the full mock as a final dress rehearsal.

Stretch the 14-Day Plan across more days — your remaining days won't disappear if you skip a date. Or just use Drill Mode (/learn/drill) for spaced-repetition recall on the highest-yield concepts.

Both, in this order: lesson body → embedded quick check → drill cards in the same domain over the following week. Reading without retrieval practice rarely sticks; retrieval without context is shallow. The combination is what compounds.

Take the full 60-question mock exam (/mock-exam) under timed conditions. If you score above 800/1000 with under 70 minutes elapsed and your domain breakdown has no glaring weak spots, you're solidly in the pass zone.

An atomic, testable sub-concept inside a lesson. "Prompt cache TTL is 5 minutes" is a KP. "Static-first ordering maximizes cache hits" is a different KP. The site has ~165 KPs across the 30 lessons.

A standard education-research model from Corbett & Anderson (1995) that estimates a per-KP probability you've mastered it. Every correct/wrong answer on a lesson check, the diagnostic, mock exam, or drill nudges the estimate up or down. A KP crosses 90% to register as "mastered."

Each cell's intensity tracks its BKT probability. Untouched KPs are gray; learning ones progress from a faint tint to the domain's full color; mastered KPs glow softly. Open /plan/map and use the All / Weak / In-progress / Mastered filter pills to focus.

Yes — there's a Restart button on the plan overview. It wipes the plan state (your KP mastery stays). The next visit rebuilds the schedule from Day 1.

Truly adaptive. Days 1–N are locked, but days N+1 through 14 reshuffle every visit. Weakest remaining KPs move forward in the queue; KPs you've mastered drop out, freeing time for review or harder material on the back end.

Nothing breaks. The plan's currentDay rolls forward by calendar date, but no KPs are dropped. Days you skipped just have unfinished tasks you can revisit any time — they'll also factor into the next rebalance.

Weighted: each domain contributes its exam-percentage share of the 1000-point total. Within a domain, points are split evenly across questions. A perfect score is 1000; 720 passes. The results card breaks down your performance by domain so you can see where to study next.

As many times as you like. Each attempt gets a unique URL, your timer persists if you refresh, and the per-attempt review screen stays around. Your best attempt is what gets surfaced on the dashboard readiness score.

The diagnostic is a 15-minute gap-finder — 12 mixed-domain questions that point at your weakest two domains. Drill mode is spaced-repetition over the highest-yield ideas (SM-2 algorithm, Again/Hard/Good/Easy buttons). The mock exam is the real-test simulation — 60 timed scenario questions weighted by domain.

No. Flagging is just a bookmark to come back to during the same attempt. It has no effect on scoring.

Yes — after submitting, the review section shows every question with your A/B/C/D choice next to the correct A/B/C/D and an explanation paragraph. Each question also displays the domain it belongs to via the colored dot.

Tap it to have the question read aloud by Amelia or Roger (your choice from the footer voice picker, powered by ElevenLabs). It's surprisingly useful for accessibility and for catching things your eyes skim over.

As of May 2026: Opus 4.7, Sonnet 4.6, and Haiku 4.5 are the current production lineup. The exam tests model-selection scenarios (when to step up to Opus, when to step down to Haiku, when to mix) using these as the reference set.

Opus 4.7 supports an optional 1M-token input window in addition to the standard 200K. Pricing tiers differ above 200K input. The exam covers when the 1M window is the right reach and how compaction strategy adapts.

Yes — the official exam includes questions that turn on the precise identifier. Memorize: claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5-20251001. They're sprinkled through our drill cards on purpose.

Still supported in production deploys; still fair game on scenario questions where the prompt specifies them. The exam default, when a scenario doesn't name a model, is Sonnet 4.6.