Before you open a single doc
- Pick your track and read its official exam guide end to end. Domains, weights and rules change, so trust the guide over any blog post (this one included).
- Confirm the current exam fee, format, time limit and retake policy on the official page. I won't guess at those here.
- Book a rough exam date. A deadline about four weeks out keeps most people honest.
Architect Foundations: where the points are
The official guide splits scored items across five domains. Spend your hours roughly in proportion:
- Agentic architecture and orchestration, about 27%. Agent loops driven by
stop_reason, handing work to subagents and passing context to them, task decomposition, and knowing when code should enforce a rule instead of trusting the model. - Claude Code configuration and workflows, about 20%. The CLAUDE.md hierarchy, hooks, plan mode, slash commands, permissions and running Claude Code inside CI.
- Prompt engineering and structured output, about 20%. Few-shot examples, turning vague asks into clear criteria, JSON schemas through tool use, and validate-then-retry loops.
- Tool design and MCP integration, about 18%. Tool descriptions that keep similar tools apart, error responses that say whether a retry makes sense,
tool_choice, and wiring MCP servers into agents. - Context management and reliability, about 15%. Long sessions, keeping facts and sources intact, escalation rules and when a human should review.
Build stuff, because reading won't cut it
- Write a small agent loop against the Claude API yourself and handle every
stop_reasonit can return. - Build one MCP server with two tools that look alike, then tune the descriptions until Claude picks the right one every time.
- Set up a repo with a CLAUDE.md, a hook and a custom slash command, and see what each one actually changes.
- Get Claude to return JSON that matches a schema, then break it on purpose and write the retry.
- Run a long task until the context fills up and notice what gets lost.
The last week
- Work through the sample questions in the official guide and write down why each wrong answer is wrong.
- Reread the domain you scored worst on. It's usually orchestration or context.
- Check your ID, your setup and the exam rules a day early so nothing surprises you on the day.
Official info lives at Anthropic Academy. If anything here disagrees with it, go with Anthropic.
Want a hand with it?
For $150 you get exam access, my prep pack and Q&A with me while you study. The exam fee is separate. See what's included.
Get exam access + prep, $150