Build the skills the Claude Certified Architect exam validates — not just the answers.

How autonomous agents actually run: the loop that drives tool use, how a coordinator delegates to subagents, how context moves between them, and how to enforce critical steps deterministically instead of hoping a prompt holds. This is the largest domain and the conceptual backbone of everything else.

How to design tools an LLM can actually pick correctly, how MCP servers/tools/resources fit together, how to return structured errors, and how to distribute tools across agents without degrading selection.

CLAUDE.md hierarchy and scoping, custom slash commands and skills, path-specific rules, plan mode vs direct execution, iterative refinement, and CI/CD integration.

Explicit criteria over vague instructions, few-shot prompting for consistency, enforcing schema-valid output via tool_use, validation/retry loops, batch processing, and multi-instance review.

Preserving critical facts across long interactions, escalation/ambiguity patterns, error propagation across agents, large-codebase context strategies, human review/confidence calibration, and provenance in multi-source synthesis.