The mental model that passes the CCAR-F exam
Prompts guide. Code enforces.
Nearly every wrong answer on the CCAR-F substitutes a probabilistic mechanism for a deterministic one. The model is asked to please remember to do something, when the architecture could have made it impossible to do otherwise. Once you internalise this, whole families of distractors collapse.
| When the requirement is… | The wrong instinct (probabilistic) | The exam answer (deterministic) |
|---|---|---|
| Tool A must run before tool B | Few-shot examples showing the order; a system-prompt rule | A programmatic prerequisite check, or a PreToolUse hook that blocks B until A's result is recorded |
| Know when the agent loop ends | Parse the text for "done" / "final answer" | Branch on stop_reason: tool_use vs end_turn vs max_tokens |
| Output must be valid JSON of a fixed shape | "Respond only with JSON, no prose" | output_config.format with a JSON Schema, or a tool with strict: true |
| Escalate hard cases to a human | Ask the model for a confidence score and threshold it | Rule-based triggers on observable state: refund > $X, N failed tool calls, customer tier, policy flag |
| An agent must not touch production | "You are read-only. Do not modify anything." | Scope its tool list: tools: ["Read","Grep","Glob"], plus permissions.deny |
| Retry a failed tool call | Blanket retry with backoff on every error | Return structured errors carrying an isRetryable flag; retry only those, surface the rest |
| Compliance rule must always hold | Put it in CLAUDE.md | Managed settings (permissions.deny) or a hook — CLAUDE.md is context, not enforcement |
| Reduce context pressure | Move to a bigger context window | Isolate work in subagents; clear stale tool results; offload to memory or files |
Five corollaries worth memorising
- Least privilege is always the answer to a tool-access question. If one option gives every agent every tool "for flexibility", it is wrong.
- Silent failure is never acceptable. A subagent that returns an empty result on error is worse than one that returns an explicit error the coordinator can act on.
- Isolation is a feature, not a cost. Subagent context isolation is usually the point of the design, not a side effect to be worked around.
- Cost and latency are architectural constraints, not afterthoughts. The Batches API's 50% discount is irrelevant if the stem says "real-time".
- Metadata must survive every hop. Citations, source IDs, and confidence provenance are passed as structured fields, never flattened into prose.
← Previous: CCAR-F study guide (start here) | Next: Domain 1 — Agentic Architecture & Orchestration →
Explore Other Certification Study Guides
Switch to another exam guide to prepare for your next Claude certification:
Claude Certified Associate — Foundations
Entry-level certification for non-coders: prompting, output evaluation, model selection, workflow integration, configuration, governance & responsible use.
Claude Certified Architect — Professional
Advanced enterprise architecture: solution design, system integration, evaluation frameworks, governance, safety & developer enablement.
Claude Certified Developer — Foundations
Build software on Claude: API integration, model selection, tools & MCP servers, agents vs workflows, security, safety & debugging.