🗼 Architect's Ascent
Pass the Claude Certified Architect exam by playing a game.
Climb the tower. Defeat the anti-pattern bosses. Out-architect The Proctor.
▶︎ pankajarm.github.io/cca-f-game — no install, no signup, progress saved in your browser.
A study game for the Claude Certified Architect — Foundations (CCA-F) exam,
built from the ground up and current as of June 2026. Instead of grinding a
PDF of flashcards, you climb a ten-floor tower, fight a boss made of each way
real agent systems break, and finish against The Proctor — a timed mock exam
built to the official blueprint: 60 questions, 120 minutes, 4 of the 6 official
scenarios, with about one choose-two (multiple-response) item in ten. The
Blueprint maps the question bank to the official exam guide's 30 task statements
so you can study the guide line by line. All 865 questions were written to the
exam blueprint and checked by an adversarial reviewer.
It's one folder of static files. Nothing to build, nothing to install.
| |
|---|
| |
| Your study hall — readiness, then Learn · Practice · Test | The Tower — ten floors, ten bosses, five domains |
⚔️ Fight the anti-patterns, don't memorize them
Each floor boss is a production failure mode — and it talks back. The Loop
Terminator stopped reading your agent's output three turns ago. The Tool Hoarder
gave one agent eighteen tools and wonders why it picks wrong. You beat them by
choosing the answer that fixes the root cause, not the one that bolts a patch
onto a broken design.
10 questions, 4 hearts, combo streaks that land critical hits and drop power-ups:
🗜️ /compact trims two wrong options · 💭 Extended Thinking reveals a hint · 🛡️ Checkpoint forgives a miss.
🎓 Learn it first: slides, video, labs
New to agents, MCP or Claude Code? Learn teaches the exam guide in three stages before you test
yourself. Stage 1 gives a short primer for each domain. Stage 2 has one lesson for each of the 30
task statements: the idea, how it works with a small original code sample, how it shows up in a scenario,
and the exam traps (the tempting answer next to the one the exam wants, and why). Stage 3 walks
through all six scenarios and exam day.
Every lesson comes as a slide deck (in the game, or full screen and printable at learn/deck.html), a
narrated video with captions, a hands-on lab, a two-question check and a drill on the real bank. A
placement check and a "weak spots" panel send you to the lessons you need. Each lesson links the official
docs, Anthropic Academy courses, cookbook notebooks and official Anthropic talks that go deeper.
The videos are AI-narrated with Kokoro-82M, an open-weights
text-to-speech model run locally (scripts/build_videos.py). Lessons are unofficial study material written
from the exam guide and Anthropic's documentation; see docs/SPEC-learn-track.md.
📋 Then survive The Proctor
The mock exam follows the official exam guide (v1.0, July 2026): 60 questions in
the exact domain weights (16/11/12/12/9), grouped under 4 of the 6 exam
scenarios drawn at random, a 120-minute clock with a pace readout,
flag-for-review and a question navigator. Domains and task statements stay hidden
until you submit. Close the tab and it resumes, with the clock still running, as it
would on exam day. The result gives an estimated scaled-score range next to your
raw score (72% raw maps to the 720 line; the real scaling isn't published) plus
breakdowns by domain and by scenario, and points you at the task statements to drill
next. About 6 of the 60 items are choose-two (5 options, pick 2, all or nothing); the real
exam has multiple-response items too, but doesn't publish their share. Prefer to learn as you go? The
Practice Exam uses the same build with a 2-minute timer per question and the
answer after each one.
| |
|---|
| |
| Exam-style conditions — scenarios, timer, pace, flags, navigator | Estimated score range + domain and scenario breakdown |
🎮 A study path built around the real exam
The home screen opens on Exam readiness: for each domain, how much you've covered and your first-try accuracy against the 72% line, your latest mock estimate, and one Next best step button. Below it the modes are grouped the way you should use them: Learn, Practice, Test. Every question in every mode maps to an exam-guide task statement.
| Mode | What it is |
|---|
| 🎓 Learn | 43 lessons in three stages: domain primers, one per task statement, scenario walkthroughs. Slides, narrated video, lab, quick check, drill. |
| 🗼 The Tower | The campaign. Ten floors, ten bosses with banter and grudges. Each floor covers a group of task statements and serves 10 questions you haven't seen yet. Clear at 7/10 to ascend; ★ for a perfect 10. |
| 📘 The Blueprint | The official exam guide as a map: 5 domains → 30 task statements → key ideas, your coverage, links to the official docs, and a drill of up to 10 questions for each. Also by scenario. |
| 📋 The Proctor | Mock exam (60 Qs, 120 min, 4 of 6 scenarios, resumable, estimated scaled-score range, about 6 choose-two items) or Practice exam (2 min per question, instant feedback). Each new exam prefers questions you haven't had in an earlier one: the bank holds a dozen fully separate mocks. |
| ⚡ Recall Sprint | 90 seconds over questions you've already studied (the whole bank until you've seen 20); combos multiply your score. Builds recall, doesn't count toward readiness. |
| 🔦 Anti-Pattern Hunt | Twelve monsters, each an anti-pattern the exam's wrong answers are built from. Hunt one through the task statements it haunts: 6 of 8 right captures it. |
| 🩹 Mistake Clinic | Every question you miss checks in here. Get it right twice in a row to discharge it — spaced repetition aimed at your weak spots. |
| 🏆 Trophy Hall | 14 achievements, 7 titles from Apprentice to Certified Architect, daily streaks. |
| |
|---|
| |
| The Bestiary — collect all 12 anti-pattern monsters | Proper dark mode for late-night cramming |
🧠 The question bank is the point
questions.js holds 865 scenario questions (100 of them choose-two), split by the exam's weights
(D1 229 · D2 154 · D3 168 · D4 166 · D5 148) with at least 18 on every one of the
30 task statements. Every question has a stable id, the task statement (ts) and
scenario it tests, and new ones carry the official doc page (src) that backs the
answer. scripts/check_bank.py keeps the bank honest: valid tags, no retired ids,
answer positions balanced, and the correct option is the longest one only about as
often as chance (so "pick the longest answer" doesn't work). It validates both item
types: single-answer (4 options, one key) and multiple-response (type: 'mr', 5 options,
two keys, each option index a key in 25–55% of MR items, the two keys being the two
longest options in at most 20%). scripts/check_cues.py is a separate gate: it
scores test-taking strategies (follow or avoid the hint, absolute words, the longest
or shortest option) by how far they beat chance, separately for single-answer
(max +2 points) and multiple-response (max +3 points), and --list names the items
that leak. scripts/test_tooling.py self-tests the checkers on fixtures. The first 197 were authored
against a June-2026 brief of the real exam — stop_reason loop mechanics, hooks
vs prompts, MCP config scopes, the CLAUDE.md hierarchy, skills frontmatter,
claude -p CI patterns, batch API economics, schema design, context management,
escalation policy — then every question was audited by an adversarial reviewer
agent that argued for each distractor like a grumpy test-taker filing an appeal.
Anything defensible got rewritten. The 548 added in October 2026 were written
from the exam guide and current Anthropic docs, one batch per task-statement group,
and each batch went through an independent reviewer that re-checked every answer
key against the docs before it was merged (scripts/check_batch.py,
scripts/add_questions.py). None are real or reconstructed exam items.
The principle baked into every answer: the correct option fixes the root
cause; the distractors are patches on a flawed design. (Better tool
descriptions before a routing classifier. A deterministic hook before a
sterner prompt. A nullable schema field before a "don't fabricate" warning.)
📐 The exam this trains you for
| Fact | Value |
|---|
| Exam | Claude Certified Architect — Foundations, exam code CCAR-F |
| Source | Official Exam Guide v1.0, effective July 2026 |
| Format | 60 scenario-based items, multiple-choice and multiple-response |
| Time | 120 minutes, proctored (online or Pearson VUE test center) |
| Scoring | Scaled 100–1000, pass at 720; report shows % correct by domain |
| Audience | Solution architects with 6+ months building with Claude |
| Cost | $125 USD per attempt |
| Scenarios | 6 in the bank, 4 drawn at random per sitting |
| Validity | 12 months; free online renewal assessment |
| Eligibility | Employees of Claude Partner Network organisations; the free courses on academy.claude.com are open to everyone |
| Practice | The old official practice exam was retired in June 2026; the guide's sample questions are the only official items |
Domain blueprint — the simulator draws with exactly these weights, and the
tower is laid out to match:
| Domain | Weight | Tower floors |
|---|
| D1 · Agentic Architecture & Orchestration | 27% | F1 Loop Chamber · F2 Coordination Hall · F3 Hook Forge |
| D2 · Tool Design & MCP Integration | 18% | F4 Tool Armory |
| D3 · Claude Code Configuration & Workflows | 20% | F5 Config Maze · F6 Pipeline Forge |
| D4 · Prompt Engineering & Structured Output | 20% | F7 Prompt Workshop · F8 Validation Gauntlet |
| D5 · Context Management & Reliability | 15% | F9 Memory Halls · F10 Synthesis Chamber |
▶︎ Play it
Online — nothing to install: pankajarm.github.io/cca-f-game
Locally:
python3 scripts/serve.py # → http://localhost:4173
…or just open index.html in any browser. Progress (XP, floors, bestiary, exam
history, mistake deck) saves automatically to your browser's localStorage.
🖨️ Prefer paper?
questions.js is the single source of truth; the printable study materials are
generated from it with python3 scripts/build_tests.py:
practice-tests/test-01…10-*.md — every question for each floor's task statements, answers in collapsible spoilerspractice-tests/full-exam-01…03.md — three 60-question paper exams at exact domain weights, no question shared, with answer keysGAME.md — quest log, XP rules, and how to run boss battles with Claude Code as your dungeon masterSTUDY_GUIDE.md — a 12-week, 1-hour-a-day syllabus mapped to the tower
✨ Design
Built for long study sessions, not eye strain. Inter for prose, Bricolage
Grotesque for display, and monospace (JetBrains Mono) reserved for actual code
tokens like stop_reason and .mcp.json — which the game detects and typesets
automatically. Warm ivory-and-terracotta by day, a proper dark mode by night (it
follows your system setting until you pick one), colours checked against WCAG AA in
both. Fully playable by keyboard: every card is a button, answers take 1–4,
Enter or Esc closes the explanation, Esc quits a Recall Sprint. Animations
switch off when your system asks for reduced motion, and audio can be muted.
Not affiliated with or endorsed by Anthropic. Question content is original,
written to the public shape of the exam — not dumps. If you can clear the tower
and keep beating The Proctor comfortably above 720, you're studying the right things.
Go book the real one. 🎓
