## Why 30 Days is the Right Window for CCA-F The Claude Certified Architect Foundations (CCA-F) exam is a 60-question, 120-minute scenario test. It rewards depth in five tightly defined domains rather than encyclopaedic recall. That is exactly the kind of exam that responds well to a focused, time-boxed plan -- long enough to build real hands-on muscle, short enough that momentum does not bleed away. Thirty days at 1 to 2 hours per day is around 45 to 60 study hours. Candidates who pass on the first attempt consistently report that number, give or take ten hours, with the variance driven mostly by how much production Claude work they have already shipped. The plan below is built around four weekly themes that mirror the official domain weighting: - **Week 1** -- Foundations: models, Messages API, prompt engineering, prompt caching (the 20% prompt engineering domain plus the 15% context management domain start here) - **Week 2** -- Tool use and MCP (the 18% tool design and MCP integration domain) - **Week 3** -- Claude Code and the Agent SDK (the 20% Claude Code domain plus the largest 27% agentic architecture domain) - **Week 4** -- Review, mock exams, and weak-domain repair If you have already shipped one or more Claude integrations, you can compress this into 21 days by trimming Week 1. If you are newer to agentic systems, stretch it to 42 days by doubling Weeks 2 and 3. Do not skip the mock-exam week -- it is where the score lift actually happens. ## How to Use This Plan A few ground rules before the daily breakdown: - **Build, do not just read.** Every day has a hands-on task. Skipping it converts the plan from a study schedule into wishful thinking. - **Keep a study journal.** A plain text file with one entry per day -- what you did, what confused you, what you want to revisit -- is worth more than any highlighting tool. - **Cite the docs.** When something surprises you, paste the docs.claude.com URL into your journal. You will reread these the night before the exam. - **Use the published weighting as your compass.** When you only have an hour, spend it on the largest-weighted domain you have not touched today. - **One mock exam per week from Week 2 onwards.** Diagnostic, not punitive. All daily blocks assume 1 hour on weekdays and 2 hours on weekends. Adjust to your reality, but keep the ratio of reading to building roughly 1:2. ## Week 1 -- Foundations (Days 1-7) Goal: be fluent in the Claude API, the three current models, structured output, and prompt caching by the end of the week. ### Day 1 -- Orientation and model overview - Read the official CCA-F exam blueprint and write down all five domain weightings from memory at the end - Read the Claude model overview at https://docs.claude.com/en/docs/about-claude/models - Memorise the three current models: **Claude Opus 4.7** (`claude-opus-4-7`) for deep reasoning, **Claude Sonnet 4.6** for balanced agent work, **Claude Haiku 4.5** for fast, cheap, high-throughput - Write a one-paragraph decision tree for picking a model. Save it -- you will refer back to this constantly ### Day 2 -- Messages API and the request lifecycle - Read the Messages API reference end to end - Set up your API key, install the Python or TypeScript SDK, and send your first request - Print the full response object. Note `usage.input_tokens`, `usage.output_tokens`, and any cache-related fields - Send a second request that uses the `system` parameter and a multi-turn `messages` array ### Day 3 -- Tool use basics - Read the tool use docs at https://docs.claude.com/en/docs/build-with-claude/tool-use - Define a single tool with a JSON schema -- something boring but real, like `get_weather(city: str)` - Wire up the tool-use loop manually: send messages, detect `tool_use` blocks, execute, send `tool_result` back - Force a failure (return an error string) and watch how the model recovers ### Day 4 -- Structured output and JSON schemas - Read the structured output guidance - Build an extractor agent: feed it a paragraph of unstructured customer feedback and return a strict JSON object with `sentiment`, `topics`, and `actions` - Test it against ten varied inputs and log every schema deviation - Read the **citations** docs for a glimpse of how retrieval grounding is tested on CCA-F ### Day 5 -- Prompt engineering fundamentals - Read the prompt engineering overview at https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview - Rework your Day 4 extractor to use few-shot examples and measure the schema-conformance lift - Practise the four prompt levers: role priming, structure, examples, and chain-of-thought scaffolds - Skim the extended-thinking documentation -- know when it is worth paying for ### Day 6 -- Prompt caching deep dive - Read https://docs.claude.com/en/docs/build-with-claude/prompt-caching twice - Map the four cacheable blocks in order: **tools, system, messages, conversation history** - Add a `cache_control` breakpoint to your extractor and confirm `cache_read_input_tokens` shows up in the next request - Deliberately edit an earlier message and watch the cache invalidate. Note the rules in your journal. ### Day 7 -- Mini-project and weekly review - Build a small command-line agent that takes a freeform question, picks the right model from your Day 1 decision tree, and answers with citations - Reread your journal entries. Flag anything you cannot explain in one sentence. - Do not move to Week 2 until you can explain prompt caching, structured output, and model selection in plain English without notes ## Week 2 -- Tool Use and MCP (Days 8-14) Goal: understand the full Model Context Protocol -- transports, primitives, server design -- and write tools that the model actually picks correctly. ### Day 8 -- MCP overview and primitives - Read the MCP specification overview at https://modelcontextprotocol.io and https://docs.claude.com/en/docs/agent-sdk/mcp - Memorise the four MCP primitives: **resources, prompts, tools, sampling** - Memorise the four transports: **stdio, HTTP, SSE, WebSocket** - Draw a one-page diagram showing which transport fits which deployment (local CLI vs hosted service vs streaming) ### Day 9 -- Build an MCP server (part 1) - Pick a small but real domain: a notes server, a Postgres reader, or a CSV explorer - Implement at least one resource and one prompt - Run it over stdio and connect it to Claude Code via `.mcp.json` - Confirm Claude Code lists your server with `/mcp` ### Day 10 -- Build an MCP server (part 2) - Add two tools to the server you built yesterday - Write deliberate, unambiguous tool descriptions. Test them by listing all tool names side by side and asking yourself which one a confused user would pick - Return structured error responses (a string the model can actually use to recover) on every failure path - Write a one-page checklist of what makes a good tool description ### Day 11 -- Tool selection and disambiguation - Add a deliberately ambiguous second tool to your server (similar verbs, overlapping scope) and watch the model misroute - Tighten the descriptions until routing accuracy hits 95% - Read about `defer_loading` and **tool search** -- know what they do, when to use them, and how they keep large tool catalogues out of the active context - Practise the failure mode: too many tools cause selection regression ### Day 12 -- MCP in production patterns - Switch your server from stdio to HTTP transport - Add basic auth or an API key check - Connect it to a second Claude Code session and confirm it still works - Read how MCP servers can request **sampling** from the host (the host is the model client, not the server, this trips people up) ### Day 13 -- Built-in tools and Anthropic-hosted tools - Read about the built-in tools: **web search, code execution, computer use** - Build a tiny script that combines your own MCP tool with web search in a single agent - Note which built-in tools are billed separately and which count against your standard usage ### Day 14 -- Week 2 mock exam and review - Sit a domain-targeted mock exam covering Tool Design and MCP Integration (the 18% domain) plus Prompt Engineering (20%) - Review every wrong answer against the official docs - Update your study journal with one-sentence summaries of every gap If you do not have a mock exam, this is the cleanest moment to grab the CCA-F practice exam at https://www.prepmycert.com/course/ccaf-claude-certified-architect-foundations and start using it as your weekly diagnostic. ## Week 3 -- Claude Code and the Agent SDK (Days 15-21) Goal: master Claude Code configuration end to end, then build a multi-agent system with the Agent SDK. This is the heaviest week and covers the two largest-weighted domains (27% + 20% = 47% of the exam). ### Day 15 -- CLAUDE.md and the configuration hierarchy - Read the Claude Code memory docs - Set up a project with three layers of `CLAUDE.md`: project-level, user-level, and one imported file via `@path/to/file.md` - Watch how Claude Code merges them. Note the order of precedence. - Add one enterprise-policy-style instruction and confirm it appears in every conversation ### Day 16 -- Settings, hooks, and slash commands - Read the settings reference - Add a `.claude/settings.json` with a `PreToolUse` hook that blocks `rm -rf` and a `PostToolUse` hook that runs your linter after every Edit - Write one custom slash command that wraps a frequent prompt - Memorise the full hook event list: `PreToolUse`, `PostToolUse`, `Stop`, `SubagentStop`, `UserPromptSubmit`, `SessionStart`, `SessionEnd`, `PreCompact`, `Notification` ### Day 17 -- Agent Skills and plan mode - Read the Agent Skills documentation - Author two skills with descriptive frontmatter that triggers correctly without you having to invoke them by name - Practise plan mode -- enter it, watch Claude propose a plan, approve, and observe what changes - Note the difference between an Agent Skill (discoverable behaviour) and a slash command (explicit invocation) ### Day 18 -- Agent SDK basics - Install the Agent SDK in your language of choice (Python or TypeScript -- pick one and stick with it) - Read the SDK overview at https://docs.claude.com/en/docs/agent-sdk - Build the simplest possible agent: one query, no tools, print the streamed output - Add `allowedTools` and `disallowedTools` and observe how they constrain behaviour ### Day 19 -- Subagents and the agentic loop - Define an orchestrator agent that spawns two subagents -- a researcher and a writer - Confirm that subagents do NOT inherit parent permissions -- they re-request tool access - Add a `SubagentStop` hook that logs per-subagent token usage to a JSON file - Read the hook reference at https://docs.claude.com/en/docs/agent-sdk/hooks and note every event payload ### Day 20 -- Sessions, forkSession, and compaction - Build an agent that runs a long task. Force it to hit a context limit on purpose. - Observe compaction. Add a `PreCompact` hook that snapshots state to disk. - Use `forkSession` (or its language equivalent) to branch a session for an experiment without polluting the main thread - Resume a session from a saved session ID and confirm state survives ### Day 21 -- End-of-week build and mock - Combine everything: a Claude Code project with two custom skills, one slash command, a `PreToolUse` hook, an MCP server you wrote in Week 2, and an SDK-powered subagent triggered by a slash command - Sit a second mock exam, this time spanning all five domains - Score it, calculate your per-domain percentages, and write down the two weakest domains -- those drive Week 4 ## Week 4 -- Review, Mocks, and Final Polish (Days 22-30) Goal: convert knowledge into exam-day reflex. No new material from here on. Repair weak domains, rehearse under time pressure, rest before the exam. ### Day 22 -- Weak-domain repair (round 1) - Pick your weakest domain from yesterday's mock - Rebuild one canonical example from scratch without notes (e.g. an MCP server, a hook-instrumented Claude Code project, a subagent system) - Reread the official doc for that domain start to finish - Update your journal with three things you got wrong last time and will not get wrong again ### Day 23 -- Weak-domain repair (round 2) - Repeat Day 22 for your second-weakest domain - If both weak domains were related (e.g. orchestration + Claude Code hooks), build a single integrated example that exercises both ### Day 24 -- Full-length timed mock - Sit a full 60-question, 120-minute mock under strict exam conditions: no notes, no docs, no breaks, no phone - Score immediately - Write a one-page post-mortem before reviewing answers in detail - Distinguish careless mistakes (re-read the question) from knowledge gaps (study the topic) ### Day 25 -- Targeted review - Go through every wrong answer from Day 24 - For each, find the authoritative doc URL and bookmark it in a single notes file titled `final-revision.md` - Group misses by domain and recount the totals -- they should track your weak-domain assessment from Day 21 ### Day 26 -- Multi-agent patterns and edge cases - Review the three multi-agent patterns: **hub-and-spoke, orchestrator-worker, peer-to-peer handoff** - Walk through a scenario for each in your head: when does the answer change, and why? - Reread the model-selection cheatsheet from Day 1 -- Opus 4.7, Sonnet 4.6, Haiku 4.5 -- and practise picking the right one under realistic constraints (latency budget, cost ceiling, accuracy floor) ### Day 27 -- Second full-length timed mock - Sit another full mock. Same conditions as Day 24. - Compare scores. You want at least +50 scaled points improvement, with no single domain below 65%. - If a domain is still under 65%, schedule a 30-minute repair block tomorrow ### Day 28 -- Final repair and `final-revision.md` - Address any remaining domain under 65% - Walk through `final-revision.md` start to finish - Build nothing new today. Reread, summarise, consolidate. ### Day 29 -- Light review, logistics, and rest - Spend at most 60 minutes reviewing `final-revision.md` - Confirm exam logistics: proctoring software installed, ID ready, room cleared, second monitor unplugged, water bottle filled, phone in another room - Eat well, sleep at your normal time, do not study after dinner ### Day 30 -- Exam day - Light breakfast, light exercise, no last-minute cramming - Arrive at the proctoring software 15 minutes early - Read each scenario twice before considering the answer options - Flag-and-move on anything that takes more than three minutes - Trust the work ## Recommended Resources The right resources, used in the right order, are what compress a vague intent to certify into a 30-day march. The list below is the minimum -- not the maximum -- of what you need. ### Primary sources (read first) - **Anthropic Academy preparation courses** on Skilljar -- 13 free, self-paced modules that mirror the exam domains - **Claude API documentation** at https://docs.claude.com - **Claude Code documentation** at https://docs.claude.com/en/docs/claude-code - **Agent SDK reference** at https://docs.claude.com/en/docs/agent-sdk - **Hooks reference** at https://docs.claude.com/en/docs/agent-sdk/hooks - **MCP specification** at https://modelcontextprotocol.io - **Prompt caching guide** at https://docs.claude.com/en/docs/build-with-claude/prompt-caching ### Hands-on milestones (do all of these) - Ship one MCP server you wrote yourself, with at least two tools, one resource, and one prompt - Configure a Claude Code project with a layered `CLAUDE.md`, two Agent Skills, one slash command, and a hook in `settings.json` - Build one multi-agent system with explicit subagent boundaries and a `SubagentStop` cost-logging hook - Instrument prompt caching on a real workload and observe a cache hit rate above 80% on the cached prefix ### Practice exams The single biggest score lift comes from realistic, domain-weighted practice questions. Reading docs creates the illusion of readiness; questions reveal the truth. PrepMyCert's CCA-F practice course has 360 questions spread across six full-length mock exams that mirror the official blueprint -- https://www.prepmycert.com/course/ccaf-claude-certified-architect-foundations. ## Mock Exam Strategy Mock exams are not just diagnostics. Used well, they are your single most efficient study tool in the back half of the plan. Use this protocol: 1. **One mock per week from Week 2 onwards.** Three mocks total before exam day -- Day 14, Day 21, Day 27 -- plus one more on Day 24 if your schedule allows. 2. **Strict time conditions only.** No pausing, no looking up answers, no notes. The point is to surface real gaps under real pressure. 3. **Score immediately, then post-mortem before reviewing.** Write down your hypothesis for each wrong answer before checking the explanation. This builds the metacognition that scenario questions reward. 4. **Bucket misses three ways**: knowledge gap (study the topic), careless mistake (re-read the question), or trap (a scenario constraint you missed). Each bucket has a different repair. 5. **Aim for upward trajectory, not absolute score.** A 720+ scaled score on the final mock with a stable per-domain floor is the readiness signal. A single 780 on a mid-plan mock with a 50% domain still in the mix is not. 6. **Never sit a mock cold.** Each mock should follow a study block, not a rest day. You want to measure your peak. ## Final Week Tips That Move the Score From candidates who passed on the first attempt: - **Read every scenario twice.** Latency budgets, cost ceilings, and compliance constraints are buried in the third or fourth sentence. Miss them and you pick a beautiful answer to the wrong question. - **Eliminate two answers first.** Most CCA-F questions have one obvious distractor and one near-miss. Knock out the obvious one before weighing the remaining two. - **Watch for absolute language.** *Always*, *never*, *only*, *every* in an answer option are usually wrong. Production engineering rarely has absolutes. - **Flag and move.** Three minutes per question is the soft ceiling. Flag and come back. - **Trust the weighting.** When you are torn between two answers, pick the one that aligns with the principles of the largest-weighted domain implicated by the scenario. - **Logistics fail more candidates than knowledge.** Test the proctoring software the day before. Clear your desk. Unplug the second monitor. Have your ID ready in good light. - **Do not study on exam day.** Last-minute cramming raises anxiety and lowers recall. The plan ended on Day 29 for a reason. ## Frequently Asked Questions ### Is 30 days really enough for CCA-F? For candidates with prior Claude API experience and 1 to 2 hours per day, yes. For complete beginners to LLM API work, plan on 45 to 60 days using the same weekly structure but with each week stretched. ### How many hours per day do I actually need? The plan assumes 1 hour on weekdays and 2 hours on weekends -- about 9 hours per week, 36 to 45 hours total. Below 30 total hours, the pass rate drops sharply. Above 60 hours, returns diminish quickly. ### What if I can only commit to 20 days? Keep Weeks 2 and 3 intact -- those carry the heaviest domain weighting. Compress Week 1 into three days of focused reading on models, the Messages API, and prompt caching. Trim Week 4 to a single full mock plus targeted repair. Do not skip the final-week timed mock. ### Can I follow this plan without a Claude API key? No. The plan is hands-on by design. The exam rewards production reflexes that you cannot fake from reading. A standard Anthropic account with a small pay-as-you-go balance is enough -- you will spend a few dollars total across 30 days if you use Haiku 4.5 for most experiments. ### Should I use Python or TypeScript for the Agent SDK? Whichever you use day to day. The exam shows pseudo-code rather than language-specific syntax. Pick one and stick with it through Week 3 so you can build muscle memory rather than juggling syntax. ### How many mock exams should I sit? Three as a minimum (Days 14, 21, 27), four if you have time (add Day 24). More than that and you start memorising specific question wording rather than learning the underlying patterns. Quality of review beats quantity of mocks every time. ### What is the minimum mock score that means I am ready? A scaled 720 on a fresh, full-length mock under strict timing, with no single domain below 65%, taken inside the final week. Below that threshold, push the exam back -- the retake fee is higher than another study week. ### What if I fail my first attempt? There is a 14-day waiting period between attempts. Use it. Pull your domain-level score report, repeat Days 22-25 of this plan on your two weakest domains, and sit the exam again. First-attempt failures who follow this targeted repair protocol almost always pass on attempt two. ### Do I need to memorise specific API parameter names? The exam is scenario-driven, not syntax-driven. You need to know what `cache_control`, `allowedTools`, `disallowedTools`, `forkSession`, `PreToolUse`, and the hook event names do, but not every keyword argument by heart. Concepts beat trivia. ## Your Next Step If you are reading this with more than 30 days to spare, the highest-leverage move you can make today is to sit a diagnostic mock exam and find out where you actually stand. Knowing your starting point turns this plan from a generic schedule into a personalised one. When you are ready, take the CCA-F practice exam at https://www.prepmycert.com/course/ccaf-claude-certified-architect-foundations. Six full-length mocks, 360 scenario questions, and a domain-level breakdown that maps directly to the four-week plan above. Good luck. Read every scenario twice.