Claude Certified Architect – Foundations

The Claude Certified Architect – Foundations (CCA‑F) exam is proctored and scenario‑based, covering five domains: (1) AI system design fundamentals, (2) prompt engineering & context management, (3) tool use & MCP integration, (4) Claude Code workflows & automation, and (5) agentic architecture, evaluation, reliability & security. It rewards judgement — knowing which approach fits a situation — over memorisation.

How to use this pack: read each domain's mind map first for the mental model, drill the cheat sheets, lock in the mnemonics, skim the TL;DRs the night before, then practise with scenario questions. Always confirm details against Anthropic's official certification and Build‑with‑Claude documentation.

• LADDER — add complexity in order: Language → Augment (RAG) → Do (tools) → Delegate (agents) → Evaluate → Refine.
• QLC — choose models by Quality bar → Latency budget → Cost ceiling. SCB — Stream, Cache, Batch.
• RICE‑F — prompt anatomy: Role, Instructions, Context, Examples, Format.
• PRUNE — context window: Preserve, Rank, Use summaries, Never unbounded, Evict noise.
• NDS — a tool is Name, Description, Schema. TRP — MCP is Tools, Resources, Prompts.
• TELL — tool errors: Tag is_error, Explain, List valid input, Let it retry.
• MHPC — Claude Code levers: Memory, Hooks, Permissions, Connectors. SCAR — Scope, CLAUDE.md, Approve, Review.
• REPS — ship agents with Reliability, Evaluation, Protection, Security.
• RAIL — Retries, Approvals, Idempotency, Logs.

Every scenario names a constraint — cost, latency, accuracy, or safety. Find it first; the right answer optimises *that*, not everything. Then: prefer the simplest option that satisfies the constraint (over‑engineering is the most common wrong answer); be suspicious of absolutes (“always”, “never”, “guarantees”); for destructive or irreversible actions, the option with validation, least privilege, or a human gate usually wins; and “measure first, then change one thing” beats blind tuning. Finally, check the option actually solves the stated problem — distractors often solve a neighbouring one.