Everything You Need on Exam Day
Format, domain weights, scoring details, and a decision framework cheat sheet for the Claude Certified Architect — Foundations exam.
Exam Format
- Format
- Multiple-choice questions (MCQ), scenario-based
- Proctoring
- Remote via ProctorFree
- Duration
- 120 minutes
- Questions
- 60 questions
Domain Weights
Percentages reflect approximate weight in the scored exam. Focus extra effort on D1 Agentic Systems (27%) and the tied D3/D4 pair (20% each).
Scoring
720 / 1000
Passing Score
Scaled score required to pass
No penalty
Guessing
Answer every question — no deduction
Linear
Score Mapping
Raw correct answers map linearly to 1–1000 scale
Exam Day Checklist
- ✓Desktop or laptop with webcam
- ✓Chrome browser required
- ✓Stable internet connection
- ✓Government-issued photo ID
- ✓Clean, well-lit workspace
- ✓Close all other applications
Decision Framework Cheat Sheet
Programmatic vs Prompt Enforcement
Deterministic constraint needed?
Programmatic — use code-level guards, not instructions
Behavioral nuance or tone control?
Prompt enforcement — system prompt rules are appropriate
Critical safety / security boundary?
Always programmatic — never rely solely on prompt
Output format & structure?
Programmatic parsing + prompt hints together
Plan Mode vs Direct Execution
Task is complex or multi-step?
Plan mode — review plan before executing
Actions are irreversible (delete, deploy)?
Plan mode — always checkpoint first
Simple, well-scoped, reversible task?
Direct execution is fine
User expects real-time output?
Direct with streaming; surface errors immediately
When to Escalate to Human
Ambiguous or conflicting instructions?
Escalate — ask for clarification before acting
Action scope exceeds granted permission?
Escalate — do not self-authorize expansion
Confidence below acceptable threshold?
Escalate or abort gracefully with explanation
Repeated failures on the same subtask?
Escalate with context; avoid infinite retry loops
Context Window Strategy
Long conversation with growing history?
Summarize older turns; keep recent context verbatim
Large document retrieval needed?
RAG / external memory — do not stuff full docs inline
Persistent state across sessions?
External storage (DB / file); resume via summary prompt
Parallel sub-agents?
Each agent gets its own scoped context window