Binary CipherRead the five-domain map.
Notice that agentic architecture carries the most weight, but 73% of the blueprint sits across tools, Claude Code, prompting, and reliability.
Open the blueprint
Free subscriber resource · web only
Join the newsletter to open the Claude Architect Prep Lab.
Get permanent browser access to the curriculum, 60-question mock, architecture playground, official references, and progress tracking. No Gumroad account, purchase, or course download is required.
- Free Binary Cipher newsletter subscription
- Course stays on BinaryCipher.dev
- Your study progress stays in this browser
Free for newsletter subscribers · Web only · Progress stays in your browser
Independent certification prep labMaster Claude Certified Architect – Foundations.
Build the architectural judgment the certification is designed to test. Follow a domain-weighted curriculum, practice with 60 original scenario questions, make production tradeoffs in the architecture playground, and revise from official sources—not memorized dumps.
Start here in 15 minutes.
Use this sequence to turn a large exam blueprint into one specific next move.
Take the diagnostic cold.
Do not study first. A clean baseline tells you where judgment—not recall—is breaking down.
Start diagnostic modeExplain one architecture tradeoff.
Use the playground and say why your choice is safer, simpler, or more observable than the alternatives.
Open the decision lab
Portable plan
Save the recommended study order.
Copy the fast-start plan and keep the one-page domain and decision-rules reference open while you study.
Copied to clipboard.
Know what the exam is really testing.
The certification validates architecture tradeoffs across realistic production scenarios. Treat the weights as study allocation—not isolated silos.
D127%
Agentic Architecture & Orchestration
Choose single calls, deterministic workflows, agents, parallel workers, or orchestrator-worker systems from requirements—not fashion.
D218%
Tool Design & MCP Integration
Design clear, narrow, typed tool boundaries and reason about MCP servers, transport, discovery, permissions, and failures.
D320%
Claude Code Configuration & Workflows
Apply instruction and setting scopes, plan mode, subagents, skills, hooks, permissions, and CI-safe execution.
D420%
Prompt Engineering & Structured Output
Set success criteria, provide context and examples, constrain outputs with schemas, validate, retry, and evaluate.
D515%
Context Management & Reliability
Preserve critical facts, control context growth, propagate errors, design retries, calibrate confidence, and escalate safely.
RoleArchitect
LevelFoundations
Questions60
Testing time120 min
Official scaled pass mark720 / 1000
LanguageEnglish
The official score is scaled; 720/1000 should not be treated as a 72% raw-score requirement. This lab uses an 80% practice target to build margin, not to predict the official result.
A domain-weighted mastery curriculum.
Complete the seven stages in order on your first pass. On later passes, use mock results to revisit only weak objectives.
Saved locally
Your curriculum progress
0 of 7 stages complete.
Recommended pace
Four weeks, roughly 24 focused hours.
Each stage combines concepts, an artifact you must produce, and practice. Reading alone is not completion.
LearnOfficial docs
BuildOne artifact
ProveScenario questions
Baseline and exam mental model.
Take the diagnostic, read the five domain statements, and write down why plausible distractors can still be wrong for the stated constraints.
10-question baselineTradeoff languageStudy plan
Deliverable: a ranked list of your two weakest domains and three failure patterns.
Agentic architecture and orchestration.
Master the complexity ladder: single call, prompt chain, routing, parallelization, evaluator-optimizer, orchestrator-workers, then autonomous agent. Add stopping conditions, observable ground truth, and human checkpoints.
Agent loopTask decompositionSubagentsHooks
Lab: diagram a support agent that can read safely, requires approval for refunds, and terminates on a verified resolution or escalation.
Tool design and MCP integration.
Design narrow tools with unambiguous names, typed inputs, useful descriptions, structured results, actionable errors, least privilege, and clear separation between read and write actions.
Tool schemasMCP lifecycleResources vs toolsPermissions
Lab: specify three MCP tools for a ticket system, including validation, authorization, idempotency, and model-recoverable error payloads.
Claude Code configuration and workflows.
Understand CLAUDE.md and settings scope, plan mode, permissions, subagents, skills, hooks, reusable commands, headless execution, and the boundary between probabilistic guidance and deterministic enforcement.
Instruction hierarchyPlan modeSkillsCI/CD
Lab: design a repository configuration that shares standards, keeps personal preferences local, blocks unsafe commands, and validates changes in CI.
Prompt engineering and structured output.
Start with measurable success criteria. Use clear instructions, relevant context, examples, separation of data and directions, schema-constrained output, validation-retry loops, and representative evaluations.
Few-shot examplesXML structureJSON SchemaEvals
Lab: extract invoice data into a strict schema, test malformed documents, and distinguish semantic validation from syntactic validity.
Context management and reliability.
Control context growth without losing transactional facts. Separate working summaries from durable state, preserve provenance, classify failures, retry selectively, expose uncertainty, and escalate irreversible decisions.
Context budgetPinned factsBackoffEscalation
Lab: write a reliability plan covering timeouts, overloaded responses, invalid tool results, partial work, audit trails, and human handoff.
Integrate, simulate, and review.
Run all playground scenarios, sit the 60-question mock under time, review every miss by decision rule, then repeat only weak domains after a 48-hour gap.
Full mockError logTeach-backFinal check
Ready signal: at least 80% on two original practice sets, no domain below 70%, and a clear spoken rationale for each playground decision.
One end-to-end architecture example.
Use this customer-support flow to connect all five domains before testing yourself.
Classify risk first.
Answer retrieval is read-only. Refunds and account changes are write actions with stricter authorization and approval.
Use narrow tools.
Fetch verified customer and policy data through typed interfaces; return actionable errors and source identifiers.
Pin critical facts.
Keep account, consent, case state, approvals, and committed actions outside lossy conversation summaries.
Stop on evidence.
Finish only after the environment confirms the outcome; otherwise retry safely, ask for missing data, or escalate.
Original mock test and domain drills.
Every item includes the direct answer, reasoning, a practical application, and the tempting trap. Questions are original BinaryCipher practice material and are not copied from the certification exam.
Choose a practice mode
Build judgment, then review the trap.
Your answers and active attempt stay in this browser. The official exam uses scaled scoring; this practice engine reports raw accuracy and uses 80% as a mastery target.
Closed-reference practice is recommended. No API key, sign-in, or data upload is required.
DiagnosticQuestion 1 of 10
Time left20:00
Answered0 / 10
Architecture Decision Lab.
Build a solution from five decisions. The lab scores how well the choices fit the scenario constraints, then explains the tradeoffs and safer production boundary.
Request
→Workflow
→Typed tools
→Validated result
Teaching goal
Learn to choose the simplest architecture that meets the requirements, then add reliability at irreversible or high-uncertainty boundaries.
Official reference library.
Use the first resource in each domain to learn, the next to implement, and the mock explanations to decide what deserves a second pass.
Building effective agents
Anthropic’s core mental model for choosing between single calls, workflows, and autonomous agents, with routing, parallelization, orchestrator-worker, and evaluator-optimizer patterns.
Claude Agent SDK overviewThe official starting point for agent loops, tool access, permissions, sessions, and production-oriented SDK capabilities.
Claude agent patterns cookbookOfficial runnable examples for turning orchestration patterns into concrete implementations and evaluation loops.
Tool use with ClaudeThe canonical guide to tool definitions, input schemas, tool choice, tool-result blocks, parallel calls, and client-side execution.
Model Context Protocol introductionThe official MCP orientation for the host-client-server model and the difference between tools, resources, and prompts.
Build an MCP serverA practical official tutorial for designing and exposing a small MCP server with discoverable, typed capabilities.
Connect Claude Code to MCPOfficial configuration and scope guidance for local, project, and managed MCP servers in Claude Code.
Claude Code documentationThe official map for installation, interactive use, permissions, configuration, extensibility, and team workflows.
Manage Claude’s memoryHow CLAUDE.md files, scoped instructions, imports, and the instruction hierarchy shape team-consistent behavior.
Claude Code settingsThe official reference for user, project, local, and managed settings plus permission and precedence behavior.
Create custom subagentsUse specialized context, prompts, and tool permissions to delegate bounded work without contaminating the main task context.
Automate workflows with hooksA practical guide to deterministic lifecycle automation, validation, notifications, and policy enforcement.
Claude Code common workflowsOfficial examples for understanding codebases, fixing bugs, refactoring, testing, pull requests, and controlled iteration.
Prompt engineering overviewThe official sequence: define success criteria, build evaluations, establish a baseline, then improve the prompt only where evidence shows a gap.
Claude prompting best practicesDirect instructions, context, examples, XML structure, role prompting, and long-context techniques in one official guide.
Structured outputsThe official approach to schema-constrained JSON and reliable tool inputs, including compatibility and validation considerations.
Define success criteria and build evaluationsHow to turn “better output” into measurable criteria, representative test cases, automated grading, and human calibration.
Context windowsThe official model for token accounting, context growth, tool-result history, and how context affects multi-turn applications.
Prompt cachingUse stable prefixes and explicit cache breakpoints to reduce repeat processing while preserving the correct request structure.
API errorsThe official error taxonomy and request-id guidance needed for retry design, observability, support, and safe failure handling.
Rate limitsRequest and token limits, response headers, acceleration limits, and the signals an architect should use for backoff and admission control.
Anthropic Academy: Build with ClaudeAnthropic’s official learning directory for API development, agents, MCP, Claude Code, tool use, prompting, evaluations, caching, and production patterns.
Revision rules that beat memorization.
Use these as a fast decision filter when multiple options look technically possible.
Start with the simplest sufficient system.
Prefer a single call or deterministic workflow when the path is predictable. Add autonomy only when the task genuinely requires dynamic planning.
Put determinism at hard boundaries.
Use code, schemas, permissions, hooks, and tests for rules that must always hold. Prompts guide behavior; they are not an enforcement layer.
Give the model ground truth.
Tools should return enough evidence to observe progress, verify outcomes, and recover—especially after write actions.
Preserve facts outside lossy summaries.
IDs, approvals, commitments, money, consent, and action state belong in durable structured state with provenance.
Validate meaning, not only shape.
Schema-valid output can still be wrong. Check business invariants, source support, allowed values, and cross-field relationships.
Retry only recoverable failures.
Back off on transient load, repair invalid structured output with bounded attempts, and escalate authorization or ambiguous policy decisions.
High-signal glossaryTerms you should be able to explain without product-marketing language.
Common trapsWhy a polished-looking answer can still be the wrong architecture.
Maximum autonomy: more agent freedom increases cost, latency, and error propagation when a fixed path would work.
One giant tool: broad tools hide intent, weaken permission boundaries, and make errors harder for the model to correct.
Prompt-only governance: telling a model “never do X” is weaker than removing the permission or enforcing a deterministic gate.
Keep everything forever: unbounded transcripts add cost and noise without guaranteeing critical-fact retention.
JSON means correct: parseable output can still violate policy, reference the wrong entity, or invent unsupported values.
Retry every error: authorization, invalid requests, and policy ambiguity need correction or escalation—not repeated calls.
Final readiness checklistUse this before scheduling or sitting the official exam.
One-page revision artifact
Open the reference boardKeep the domain map and decision rules open while you revise.
The web reference contains the blueprint weights, architecture ladder, tool checklist, reliability loop, and final readiness signals.