Recommended material:

• https://www.certsafari.com/anthropic/claude-certified-architect-foundations
• https://claudecertificationguide.com/mock-exam
• https://github.com/paullarionov/claude-certified-architect/blob/main/guide_en.md

Tips

• Constrain, don't add. When something's failing, the fix is usually fewer tools, a narrower scope, or a smaller task, not more tools, more agents, or a bigger context window.
• Prompts suggest; systems enforce. A rule stated in a prompt is a request Claude can still miss under pressure. If it must always hold (money, compliance, ordering), enforce it in code or a hook, not by asking nicely.
• Fix the earliest layer. A bad final output is often the right fix applied one step too late. Trace back to where the error actually entered (bad chunk, bad retrieval, bad tool schema) instead of patching the output.
• Match the fix to the failure.
  • Vague rule -> explicit criteria.
  • Inconsistent output -> few-shot examples.
  • Invalid JSON -> schema.
  • Valid-but-wrong -> semantic/human validation, not a stricter format.
• Don't collapse states; escalate on policy, not vibes. "Tool errored" and "tool found nothing" are different signals needing different handling. Escalate on concrete triggers (a specific error, a retry ceiling), never on model confidence or tone.
• Auto mode's classifier checks safety, not correctness. It blocks dangerous-looking actions; it has no opinion on whether the code actually works. Pair auto mode with a Stop hook running your tests, that's what closes the correctness gap.
• Loops need a hard ceiling, not just a stop condition. A tool-call loop that keeps retrying the same failed approach rarely self-corrects, it just burns tokens until something else breaks. Cap retries explicitly.

Scenarios

1. Customer Support Resolution Agent (Agentic Architecture, Tool Design, Context/Reliability)

process_refund is the highest-stakes tool here: expect it to need a cap without human sign-off, idempotency, and validated customer/order context before it's even reachable. escalate_to_human needs clear triggers (ambiguity, anger signals, repeated tool failure, amount over threshold). 80%+ resolution still requires a real escalation path — an option that maximizes resolution by removing refund caps or confirmation is the trap, every time.

2. Code Generation with Claude Code (Claude Code Config, Context/Reliability)

CLAUDE.md scoped by directory in a monorepo, not one giant root file. Plan mode for "refactor our auth system," direct execution for "fix this null pointer." Slash commands for repeated workflows. Long coding sessions need compaction/fresh sessions rather than infinite context growth.

3. Multi-Agent Research System (Agentic Architecture, Tool Design, Context/Reliability)

This is the case where multi-agent is actually justified — search and document-analysis are independent and parallelizable. Subagents must return synthesized findings with citations preserved, not raw dumps, or the coordinator's context balloons for no benefit. Citation metadata needs to be passed structurally between synthesis and report-generation agents. If a subagent fails, don't silently halt the whole pipeline — retry, proceed with a flagged gap, or escalate depending on criticality.