An independent study platform for Anthropic's Claude Certified Architect – Foundations exam: five domains weighted exactly like the real exam blueprint, each with real diagrams and code rather than plain definitions, plus an Architecture Decision Lab that hands you a requirements list and grades whichever architecture you pick with a verdict and a full trade-off explanation.
The part that isn't just a study guide: a requirements list, five plausible architectures, and a real trade-off explanation for whichever one you pick.
Each scenario starts with a real requirements list (e.g. an insurance claims assistant that has to read a policy, analyze an uploaded claim, query internal systems, escalate high-value claims, and produce structured JSON with an audit trail) and five architecture options: a single Claude API call, a deterministic workflow, a single tool-using agent, a multi-agent architecture, or an MCP-based architecture.
Picking one doesn't just mark it right or wrong. It returns a verdict (Best fit, Tempting, but wrong, Over-engineered, or Viable), a diagram specific to that choice, and a written explanation walking through why it does or doesn't fit each individual requirement, plus the specific judgment call the exam is actually testing for. That's the same scoping reasoning the real Foundations exam is built around, not just terminology recall.
Agentic Architecture (27%), Tool Design & MCP (18%), Claude Code (20%), Prompt Engineering (20%), Context & Reliability (15%): matching the actual exam blueprint, not an arbitrary split.
Every domain page carries the same lesson template: a comparison table of the approaches being taught, real architecture diagrams, syntax-highlighted code samples, a decision framework, and a "common mistakes" callout, not a wall of prose. Progress is tracked per domain and rolled up into a single weighted-overall percentage, matching how the real exam actually scores.
The current, real architecture (V0.1), not the aspirational one an earlier version of this write-up described.
Lessons, quiz banks, and scenarios live as typed, Zod-validated content, not hand-maintained page files, backed by a Vitest suite that checks every question against the schema.
Progress state lives in memory only: no database, no auth. That's a reversal of an earlier localStorage-based attempt (ADR-0003 → ADR-0004) once it became clear a learner's progress didn't need to survive a reload to be useful.
The scenario-grading system itself: a requirements list mapped against five architecture options, each with its own verdict, diagram, and requirement-by-requirement reasoning.
Lesson prose renders through react-markdown rather than raw HTML or MDX, so content can't execute arbitrary code: a real constraint once four of the five domains' content was AI-drafted (see below).
The original content plan was to write lessons in MDX. Once the lesson template's actual blocks (diagrams, comparison tables, code samples, callouts) were being built for real, MDX turned out to be the wrong fit, and content moved to a typed, Zod-validated schema instead (ADR-0002).
Progress first shipped via localStorage, ahead of any real database or auth provider (ADR-0003). That was then reversed to store nothing about a learner at all: progress is purely session-live, resetting on reload (ADR-0004), a deliberate simplification, not a fallback that fell short of the original plan.
Four of the five domains' lessons and practice questions were drafted by Claude, matching Domain 1's own hand-built template exactly. None of that shipped on the drafting alone: it had to pass schema validation and structural tests, then a fact-check pass against primary sources (the Claude Code docs, MCP spec, and Anthropic API docs, which caught one real inaccuracy in a Domain 5 code sample), and then my own full review, before any of it went live.
Each part maps to a real, current AI/software engineering skill, not just exam trivia.
Designed a typed, Zod-validated content schema for lessons, quiz banks, and scenarios, enforced by a Vitest suite: a real system for content, not four hand-maintained one-off pages.
Directed Claude to draft four of five domains' content, then required automated validation, a primary-source fact-check pass, and a full personal review before any of it shipped, catching a genuine inaccuracy along the way.
Built a Decision Lab that grades real architectural scoping judgment with a tailored verdict and reasoning per choice, not a multiple-choice terminology quiz.
Scoped, built, and shipped a complete 5-domain platform end to end, including a real architectural pivot (MDX to a typed schema) made after starting to build, not before.
Pick a scenario and see the Architecture Decision Lab grade your call:
↗ Open Claude Architect Lab