Claude from Zero to Mastery (Claude Certified Architect prep)
You do not need to know anything about AI. We start with what Claude actually is and go step by step to designing systems that work correctly in practice too. That is exactly what the Claude Certified Architect exam measures.
47 lessons, in order
- 110 min
0.1 What is Claude, really?
Before this: nothing. Start right here.
How is an ordinary program different?So what is Claude?Why is it sometimes confidently wrong?Where does the architect come in? - 210 min
0.2 How programs talk to Claude
Before this: Lesson 0.1. No programming needed.
What does API mean?The simplest requestThe point everything hangs on: there is no memoryWhat does a response look like?Why does this matter to an architect? - 310 min
0.3 Tokens and the context window: the model's short-term memory
Before this: Lesson 0.2.
Tokens: how the model measures textThe context window: the model's deskWhy more is not always betterWhen the desk is full: summarising, and its dangerMax_tokens is not the context window - 410 min
0.4 Prompts: how to ask the model for work
Before this: Lesson 0.1 and Lesson 0.2.
Why isn't "write well" enough?Four things a good prompt hasGive the reason behind the rule tooWhere does a prompt stop? - 510 min
0.5 Tools: how a model with no hands gets things done
Before this: Lesson 0.2.
The problemHow it worksDefining a toolThe model's request and your result, side by side - 610 min
0.6 What an agent is, and why a loop comes into it
Before this: Lesson 0.5.
The problem: once is not enoughWhat is a loop?So what is an agent?Why does the loop matter so much? - 710 min
0.7 A map of the Claude family: models, API, Claude Code, SDK and MCP
Before this: Lesson 0.6.
Models: which brain?Four layers on top of the modelWhy was MCP needed?Where do skills, hooks and subagents fit on the map?How do you choose? - 810 min
1.0 Introduction: why agents break in practice
Before this: Lesson 0.5 and Lesson 0.6. You need to know what a tool is and why an agent has a loop.
The story: a great demo, a bad first weekWhy a demo misleadsWhere each failure gets fixedHow the lessons connectIts weight in the exam - 910 min
1.1 The agentic loop
Before this: Lesson 0.5 and Lesson 0.6. There you saw the idea of the loop. Here we open it up step by step. No programming needed.
The story: an agent that sometimes stops halfwayThe right way: a field built for exactly thisThe loop, step by stepWhy the other ways are wrong - 1010 min
1.2 The coordinator and subagent pattern
Before this: Lesson 1.1. You met the word subagent in Lesson 0.7.
The story: a report that looked completeThe idea: coordinator and subagentThe team's shape: a starWhat the coordinator doesWhy not the other ways - 1115 min
1.3 Building a subagent and giving it context
Before this: Lesson 1.2 and Lesson 0.7 (what the Agent SDK is).
The story: "the customer we mentioned"What a subagent seesDefining a subagentA good prompt for a subagentGive the coordinator a goal, not a procedureParallel work: several calls in one response - 1215 min
1.4 Enforcing in code, and handoff
Before this: Lesson 0.1 (probabilistic behaviour) and Lesson 1.1.
The story: a rule that sometimes got brokenEnforcing in code: the ideaHow it worksThe line: when prompt and when code?One message, several problemsHanding work to a human (handoff) - 1315 min
1.5 Hooks in the Agent SDK
Before this: Lesson 1.4 and Lesson 0.7. Remember MCP from that lesson.
The story: two faults a prompt cannot fixWhat a hook isRefusing above the capMaking the output uniformWhy a hook and not the other ways - 1410 min
1.6 Breaking work down: fixed chains versus dynamic splitting
Before this: Lesson 0.6 (workflow versus agent) and Lesson 1.2.
The story: two jobs, two failuresPrompt chaining: when the steps are knownDynamic decomposition: when the steps are not knownHow to chooseThe obvious way and why it breaks - 1510 min
1.7 Sessions: continue, resume and fork
Before this: Lesson 0.2 (no memory), Lesson 0.3 and Lesson 0.7.
The story: the agent and Friday's pictureWhat a session isThree ways backCarry on, or start fresh?Why not the other ways - 1610 min
2.0 Introduction: a tool is the model's user interface
Before this: Lesson 0.5 and Lesson 0.7. Lesson 1.1 helps too, but you don't need it.
The story: one line of text makes all the differenceA tool is the model's user interfaceA tool's life, and this domain's lessonsWeight in the exam - 1710 min
2.1 Tool descriptions and boundaries
Before this: Lesson 0.5 and Lesson 2.0.
The story: two tools, one descriptionWhy the description matters so muchWhat a good description containsOverlapping toolsWhy not the other routes - 1810 min
2.2 Structured errors in MCP
Before this: Lesson 0.5 and Lesson 1.1. We introduced MCP in Lesson 0.7, and its details are in Lesson 2.4. Here it's enough to know that MCP is a...
The story: three failures, one messageFour kinds of errorHow to tell the modelRecovery inside a subagentWhy not the other routes - 1915 min
2.3 Distributing tools, and tool_choice
Before this: Lesson 0.5 and Lesson 2.1. We introduced subagents in Lesson 0.7.
The story: the agent that can do everythingWhy too many tools is badWho decides? tool_choiceToday's alternatives - 2015 min
2.4 Connecting MCP
Before this: Lesson 0.5 and the MCP section of Lesson 0.7.
The story: the same tool, different on every laptopServers, tools and resources in plain wordsWhere each server is definedA secret is never committedResources in practiceWhy not the other routes - 2110 min
2.5 Choosing the built-in tools
Before this: Lesson 0.3 and Lesson 0.5.
The story: the agent that reads all the codeFive toolsWhy Edit needs unique textUnderstanding code, step by stepWhy not the other routes - 2210 min
3.0 Introduction: how Claude Code learns what your project is like
Before this: Lesson 0.6 and Lesson 0.7. If you've never seen Claude Code or worked with git, that's fine: we explain from zero.
The story: a brilliant colleague who forgets everything every morningWhat is Claude Code?What does a session look like?A few words you'll see all the time in this domainThree places for knowledge, two ways of working, one mode without youWeight in the exam - 2310 min
3.1 The CLAUDE.md hierarchy
Before this: Lesson 3.0 and Lesson 0.3.
The story: "it works for me"What is CLAUDE.md?How do they fit together?Writing it well and keeping it shortWhy not the other routes? - 2410 min
3.2 Commands and skills
Before this: Lesson 3.1. For the context: fork setting, Lesson 1.2 helps too.
The story: a checklist that is either forgotten or always in the wayWhat are skills and commands?How do we build one?Where do we put it?Why not the other routes? - 2510 min
3.3 Path-specific rules
Before this: Lesson 3.1 and Lesson 3.2 (for what frontmatter means).
The story: a rule scattered across thirty foldersThe right way: a rule that knows when it's neededWhat is a glob pattern?Which tool for which rule? - 2610 min
3.4 Plan mode or direct execution
Before this: Lesson 3.0. Lesson 1.2 helps with the subagent section.
The story: a migration that got stuck halfwayWhat is plan mode?When to use which?Exploring without filling the contextWhy not the other routes? - 2710 min
3.5 Iterative refinement
Before this: Lesson 0.4 and Lesson 3.0.
The story: three explanations, three interpretationsFour toolsWhy not the other routes? - 2810 min
3.6 Claude Code in CI
Before this: Lesson 3.0 and Lesson 3.1. For what a hook is, Lesson 1.5, and for MCP, Lesson 0.7.
The story: four failures in a rowFixing the first three failuresTwo things the script has to respectReview in a separate sessionWhy not the other routes? - 2910 min
4.0 Introduction: from "a good answer" to "output you can rely on"
Before this: Lesson 0.4 and Lesson 0.5. Knowing Lesson 0.6 helps too.
The story: an answer that was fine and still took the program downWhat does reliable output mean?How the lessons connectWeight in the exam - 3010 min
4.1 Explicit criteria to cut false positives
Before this: Lesson 0.4.
The story: a bot nobody looked at anymoreWhy "be careful" does not workThe right way: explicit criteriaWhy the other ways are wrong - 3110 min
4.2 Few-shot
Before this: Lesson 0.4 and Lesson 4.1.
The story: a review that fit two categoriesWhat does few-shot mean?How to write good examplesIn code review: a fixed format and fewer false alarmsIt works for data extraction tooWhy the other ways are wrong - 3215 min
4.3 Tool use and JSON schema
Before this: Lesson 0.5 (tools and JSON Schema).
The story: a page with no sizeA few words you needThe trick: a tool that only takes dataNow, how do we make sure this shape comes back?What does it not catch?Why the other ways are wrong - 3310 min
4.4 Validation and retries
Before this: Lesson 4.3.
The story: an invoice whose total did not add upWhat does validation mean?The right retry: send three thingsWhen does retrying help?Design the schema so it shows contradictions - 3410 min
4.5 Message Batches
Before this: Lesson 0.2 and Lesson 0.5.
The story: a report that cost too muchWhat do batch and SLA mean?How it worksWhy the other ways are wrong - 3510 min
4.6 Multi-instance and multi-pass review
Before this: Lesson 0.3 and Lesson 4.1.
The story: two reviews that caught nothingProblem one: the author is not a good reviewerProblem two: attention spreads thinOne more pass: a confidence for each findingWhy the other ways are wrong - 3610 min
5.0 Introduction: a system that behaves well on a bad day too
Before this: Lesson 0.3 and Lesson 0.6. Wherever you reach a team of several agents, Lesson 1.2 helps.
A bad day at our bookshopWhat does reliability mean?How do the lessons connect?Weight in the exam - 3715 min
5.1 Keeping facts in a long conversation
Before this: Lesson 0.3 (tokens, the context window and compaction).
The story: an amount that faded awayThe fix: a case facts blockHow does it work?Three more helpersWhy the obvious ways are not enough - 3810 min
5.2 When to escalate to a human
Before this: Lesson 0.6. For the handoff summary, Lesson 1.4 helps, and for examples inside the prompt, Lesson 4.2.
The story: three mistakes in one morningExplicit criteria instead of guessingHow do we give the agent these criteria?Why don't sentiment analysis and the model's confidence work? - 3910 min
5.3 Error propagation in a multi-agent system
Before this: Lesson 1.2 (coordinator and subagent) and Lesson 2.2 (structured errors in tools).
The story: a report that was confidently wrongA good error says four thingsWho does what with which error?Why are the other ways wrong? - 4010 min
5.4 Exploring a large codebase
Before this: Lesson 0.3, Lesson 1.2 and Lesson 5.1.
The story: an agent that forgot the classesFour ways, each for one problemWhy don't the obvious ways work? - 4110 min
5.5 Human review and calibration
Before this: Lesson 4.3 (structured output) and Lesson 4.4 (validation).
The story: a 97% that hid bad newsA worked example (made up, for illustration only)Confidence scores and calibrationWhy do we sample high-confidence cases too?How does the review queue work?Why are the other ways wrong? - 4210 min
5.6 Provenance and uncertainty
Before this: Lesson 1.3 (what reaches a subagent) and Lesson 5.3.
The story: a number nobody ownedWhat does provenance mean?The fix: a structured mappingWhen sources disagreeWhy are the other ways wrong? - 4310 min
Wrap-up: the exam's six scenarios
The exam picks four of these six scenarios at random. Each scenario is a real system, and the questions come from the decisions inside it. For each...
1. A customer support agent2. Code generation with Claude Code3. Multi-agent research4. Developer productivity tools5. Claude Code in CI6. Structured data extraction - 4410 min
The logic of the answers: eight rules
Now that you know the topics, these eight rules settle most of the questions. The questions are not about memorising; they ask for judgement. The...
How to answer a question - 4510 min
The exam versus today
The exam guide we have seen lags behind the current documentation (November 2026). The version we compared is "Version 1.0, effective July 2026"...
Where they differ - 4610 min
Glossary
Whenever you forget a word, look here. Each term links to the lesson that explains it in full.
Glossary - 4710 min
The exam itself and its conditions
Read this page when you are thinking about booking. It comes from the official exam guide (Version 1.0, effective July 2026) and the official FAQ....
The exam at a glanceBefore you book
Learn Claude from Zero to Mastery (Claude Certified Architect prep) in your language
Every lesson, quiz, and tutor reply renders in your language. The AI tutor matches the language you ask in.
