Study Resources, Plan & Cheat Sheet

Official study modules (Anthropic Skilljar)

Access via the certification enrollment page: https://anthropic.skilljar.com/claude-certified-architect-foundations-access-request

Foundation courses

• Claude 101
• AI Capabilities and Limitations
• Building with the Claude API

Advanced topics

• Introduction to Model Context Protocol
• MCP: Advanced Topics
• Introduction to Subagents
• Introduction to Agent Skills

Claude Code specialization

• Claude Code 101
• Claude Code in Action
• Introduction to Claude Cowork

Hands-on artifacts to build

The exam rewards people who have actually built these. Build a working version of each:

1. An end-to-end agent with tool calling (the API + tool_choice + structured output).
2. A CLAUDE.md hierarchy with path-specific rules (root + subdirectory + @imports).
3. MCP tools with structured error handling (isError: true, error categories, retry flags).
4. A data-extraction pipeline with a validation loop (schema + semantic validation + correction requests).
5. A batch-processing example for high-volume extraction.

Suggested study plan

Quick-reference cheat sheet

Responsibility split — model interprets language & chooses; code enforces permissions, compliance, state, retries, idempotency, validation, audit.

Output control — tool use / structured outputs > “please return JSON” in text. tool_choice: auto / any / named / none.

Errors — application errors = results with isError: true, not exceptions. Categories: transient infra, permanent validation, business rule, permission, uncertain write. Never retry uncertain side effects.

MCP — tools (model-controlled actions), resources (app-controlled context), prompts (reusable workflows). Annotations are untrusted hints. Claude Code scope precedence: project > local > user.

Extraction — optional/nullable fields reduce hallucination; give an absence path (null / empty array / unclear). Schema-valid ≠ correct — add semantic validation + provenance.

Context — capacity ≠ attention. Sliding window / structured summaries / structured state / retrieval. Don’t keep every RAG result. Inject external updates as system context.

Subagents — no inherited context; restate task, findings, sources, constraints, output shape. Task/Agent tool must be in parent allowedTools. Slowest subtask = total parallel latency.

Claude Code — Grep=contents, Glob=filenames. CLAUDE.md path-scoped by placement. Plan mode = workflow control; extended thinking = reasoning quality. Hooks (PreToolUse etc.) enforce hard rules.

Escalation — user asks / authority needed / unsafe-uncertain state / no progress / policy breach. Always with a structured handoff. Graceful partial failure over total abort.

System prompts — sent every request; attention decays as conversation grows. Few-shot > prose. IMPORTANT/NEVER is not an enforcement mechanism.

Community study references

• Community study guide (GitHub): https://github.com/daronyondem/claude-architect-exam-guide
• Anthropic certification enrollment: https://anthropic.skilljar.com/claude-certified-architect-foundations-access-request

Note: weightings, the 5-domain blueprint, and the 6 scenarios are from community-compiled guides cross-referenced against the Anthropic course outline. Confirm against the official exam guide on Skilljar once you have access.

Interview angle

• “How do you keep current in this area?” - primary sources over summaries: the MCP specification changelog, model provider documentation and release notes, and the framework repositories. This field’s blog layer lags the specs by months and frequently misstates them.
• “What is the trap with AI content specifically?” - it goes stale faster than any other area covered here. A note asserting a model name, a context-window figure or a protocol shape needs a date attached, or it will be quietly wrong within a quarter. See ../../STACK_BASELINE.md.