Claude Certified Architect – Foundations: A Simple Prep Guide 🇮🇳

A no-nonsense, beginner-friendly guide to help you clear the Claude Certified Architect – Foundations exam by Anthropic — written for people preparing in India, with notes also available in Hindi and Tamil.

I cleared this certification myself — view my verified Credly badge ↗. This repo is my attempt to make the path a little shorter for the next person. It's not official Anthropic material — just a simple map of what to study and in what order.

Languages: English · हिन्दी · தமிழ்

---

What is this certification?

The Claude Certified Architect – Foundations credential is for people who design and build applications with Claude — using the Claude API, Claude Code, the Agent SDK, and MCP (Model Context Protocol). It's not a coding trivia test — it checks whether you can make sound architecture decisions: when to use an agent vs. a simple API call, how to design tools, how to keep systems reliable.

Who is this guide for?

• Engineers who have already built something with Claude (even a small side project) and want to formalize that knowledge.
• Folks who prefer a checklist over a 6-hour video course.
• Anyone who wants a quick, honest view of where the exam actually focuses.

If you haven't built anything with Claude yet, do that first — a tiny project teaches more than passively reading this guide.

---

Step 1 — Take the free Anthropic Academy courses

You don't need a paid bootcamp. Anthropic's own courses cover the syllabus. Go through these in order:

#	Course	Focus
1	AI Fluency: Framework & Foundations	How to think about working with AI, not just prompting it
2	Claude 101	Product basics — quick one
3	Building with the Claude API	The core of the exam: tool use, structured output, agent design
4	Claude Code in Action	Running real, trustworthy Claude Code sessions
5	Introduction to Model Context Protocol	MCP building blocks: tools, resources, prompts
6	Introduction to Agent Skills	Skills — often under-studied, comes up in the exam

Optional but useful if your org uses a specific cloud: Claude with Amazon Bedrock, Claude with Google Cloud's Vertex AI. Skip these if you don't touch that cloud day-to-day.

Access request for the course/exam portal: anthropic.skilljar.com — Claude Certified Architect Foundations access request (needs a work/partner email).

---

Step 2 — Know exactly where the exam puts its weight

This is the official breakdown from Anthropic's own exam guide (v0.1, Feb 2025) — always cross-check against the current version when you register, since these numbers can change:

Domain	Weight
1. Agentic Architecture & Orchestration (agentic loops, coordinator/subagent patterns, hooks)	27%
2. Tool Design & MCP Integration (tool interfaces, error responses, tool_choice)	18%
3. Claude Code Configuration & Workflows (CLAUDE.md, rules, skills, CI/CD)	20%
4. Prompt Engineering & Structured Output (few-shot, JSON schemas, batch processing)	20%
5. Context Management & Reliability (long-context handling, escalation, error propagation)	15%

The exam draws 4 scenarios (picked at random) out of a published set of 6: customer support agent, Claude Code for dev productivity, multi-agent research system, developer productivity tooling, Claude Code in CI/CD, and structured data extraction. All questions are single-answer multiple choice (4 options); there's no penalty for guessing.

Rule of thumb the exam rewards again and again: pick the simplest pattern that solves the problem. Augmented single call → Workflow → Agent, in that order of preference. Don't reach for a multi-agent system when one tool call would do.

Explicitly out of scope (per the official guide) — don't burn prep time here: fine-tuning, API billing/auth/rate limits, MCP server hosting/infra, model internals/RLHF, vision, computer use, streaming/SSE, and cloud-provider-specific configuration.

Read these to back up the table above:

• Official exam guide (PDF) — the actual syllabus/domain breakdown, always cross-check against this
• Building Effective Agents — Anthropic's own agent vs. workflow reasoning, the backbone of the "largest" row above
• Claude Code docs · CLAUDE.md / memory · hooks · skills
• Tool use & tool_choice
• MCP introduction · MCP tools · MCP resources
• Context windows

---

Step 2.5 — The one mental model that answers most scenario questions

Nearly every scenario question is a variation of one loop. Learn it once instead of memorizing 60 separate cases.

Five actors: User/entry point → the harness (your deterministic code) → the model (Claude, reasoning only) → Claude API (the boundary between them) → tool backends (the real database/payment rail/API a tool call reaches).

The loop: assemble context → call the model → model returns a decision (end_turn or tool_use) → harness checks the requested tool call against hard rules (spend limits, permission scopes) before running it → tool executes → result feeds back → repeat until done or escalated.

The one line worth memorizing: the model decides, the harness executes. Claude's job ends at returning a structured decision — whether that decision is allowed to become a real action is the harness's call. This is why a limit ("never refund over ₹15,000 without review") has to be a code-level if, not a stronger sentence in the system prompt — no prompt wording is a guarantee, and a differently-phrased request can talk a model past an instruction that a hard-coded check would have blocked every time.

When a question asks "why did the agent do X" or "what would have prevented X," first ask: is this a missing-context problem, a model-reasoning problem, or a missing/weak gate? Most "design a safer system" questions are really asking you to add a gate.

---

Step 3 — Build one small thing before you sit the exam

You don't need anything fancy. A weekend project is enough:

• A coordinator that delegates to 2–3 worker sub-agents (e.g., a search agent + a summarizer agent).
• One tool exposed via MCP, so you feel the difference between a tool (model calls it) and a resource (you hand content to the model directly).
• One place where you deliberately force structured output using tool_choice, so you see why "asking nicely" in a prompt isn't the same as a guarantee.

Building this once will make several exam questions feel familiar instead of theoretical.

Docs to build from: Agent SDK overview · Tool use implementation guide · MCP tools & MCP resources

---

Step 4 — Where most people actually lose marks

Based on shared experience across people who've taken this exam, these are the classic traps:

1. Required fields that shouldn't be required. If a field might genuinely be absent from the input, mark it optional / nullable. A required field with no matching data pressures the model to invent something. (Tool use docs)
2. Relying on prompt wording for structured output. A well-written instruction is a request. tool_choice (forcing a specific tool call) is a guarantee. If a broken format would break your downstream system, force it — don't ask politely. (Tool use docs)
3. Reaching for an agent when a workflow (or a single call) would do. More autonomy = less predictability. The exam rewards restraint. (Building Effective Agents)
4. Confusing MCP tools and resources. Not everything the model needs should be a "tool call" — sometimes just handing over the content as a resource is cheaper and more reliable. (MCP tools vs. MCP resources)
5. Ignoring resumability. If your multi-step pipeline can't pick up from where it failed, every failure costs you the whole run — and that's a real architecture flaw, not a minor detail. (Context windows)
6. Enforcing a hard limit in the system prompt instead of in code. "Never approve refunds over ₹15,000" as prompt wording is a request the model can be talked past by an unusual enough phrasing. The same rule as a harness-level check before the tool executes is a guarantee. If a question describes a limit being violated, the fix is almost always a code-level gate, not a stronger sentence.
7. Assuming subagents share context automatically. A coordinator's subagents are isolated by default — if the coordinator doesn't explicitly pass one subagent's output into the next one's prompt, the second subagent genuinely has no idea the first one ran. "Its context window was too small" is a common wrong answer when the real cause is "nobody passed the data."
8. Treating schema-valid as correct. A JSON response can pass schema validation (right types, right shape) and still be semantically wrong — line items that don't sum to the stated total, a date that doesn't match the source. Validation-retry loops fix format errors; they don't fix meaning errors, which need their own check.

---

Step 5 — Exam day tips

• Scoring: results are a scaled score from 100–1,000; you need 720 to pass (roughly 69% correct). No penalty for guessing an unanswered question, so never leave one blank.
• Expect two plausible-looking answers per question. You're usually picking the better trade-off, not spotting an obviously wrong option. Read for the constraint stated in the question (compliance, latency, cost, team size) — that constraint decides the answer.
• Don't skip the "boring" topics you think you already know well from real work — that's often where blind spots hide.
• Keep a mental checklist: Could this be a single call? Could this be a workflow? Do I really need an agent?
• The exam itself is proctored via Pearson VUE — book/manage your session at pearsonvue.com/anthropic.

---

Practice before the real exam

I built a small free app to drill practice questions in the same style as the exam — useful for a final sanity check before you book the real thing.

• Try it: claude-architect-cert-app.streamlit.app
• Source: github.com/mukilankarthik/claude-architect-cert-app

---

FAQ

Do I need a paid course? No — the free Anthropic Academy courses above cover the syllabus.

Is this an official Anthropic resource? No. This is a community guide written after personally clearing the exam. Always verify current exam scope, pricing, and access on Anthropic's official pages.

I don't read English comfortably — is there help? Yes — see the Hindi and Tamil versions linked at the top. More Indian languages are welcome via PR (Telugu, Kannada, Malayalam, Marathi, Bengali, etc.).

---

Contributing

Found an error, or want to add a regional language guide? PRs welcome:

• Add guides/guide_<language-code>.md following the structure of the existing guides.
• Keep the tone simple — short sentences, no jargon left unexplained.
• Don't copy content from other prep guides or blogs — write it in your own words, based on your own understanding/experience.

License

MIT — use, adapt, and share freely. See LICENSE.