Glossary
Claude Certified Architect (CCAR-F) exam glossary
The terms the Claude Certified Architect exam leans on, defined in a line each: stop_reason and tool_choice, PreToolUse hooks, prompt caching, the Batch API.
The CCAR-F expects you to use these terms precisely, because a wrong answer can depend on a term used loosely. The exam tests whether you can apply a term in a scenario, so learn what each term rules out as well as what it means.
Agentic architecture
| Term | What it means |
|---|---|
| Agentic loop | Send a request, inspect stop_reason, then run the tools or stop. |
stop_reason | The API field that decides whether the loop continues: tool_use continues, end_turn stops. |
| AgentDefinition | Declares a subagent's description, system prompt and the tools it may use. |
| Task tool | How a coordinator delegates to a subagent. Called Agent in current Claude Code. |
| Handoff | Control passing from one agent to another. The receiver inherits no conversation history. |
fork_session | An independent branch from a shared starting point, for exploring in different directions. |
| Hub-and-spoke | A coordinator delegating to focused specialists, each with a small set of tools. |
| Attention dilution | Uneven depth of analysis when too many items are handled in one pass. |
These carry Agentic Architecture & Orchestration, the heaviest domain on the exam at 27%.
Tools, MCP and Claude Code
| Term | What it means |
|---|---|
| MCP | Model Context Protocol: JSON-RPC 2.0 over stdio or streamable HTTP. |
| Host, client, server | The host contains the client, and the client connects to servers. Servers never talk to each other. |
tool_choice | auto, any, or a named tool the model must call. |
.mcp.json | Project-level MCP server configuration, committed and shared with the team. |
PreToolUse / PostToolUse | Deterministic hooks that run before and after a tool executes. |
.claude/rules/ | Rules that load only for matching paths, set with a paths field in their frontmatter. |
| Skill | A directory with a SKILL.md entry point, frontmatter and supporting files. |
| Command | A flat .md prompt template invoked with /, which can take $ARGUMENTS. |
These are the vocabulary of Tool Design & MCP Integration (18%) and Claude Code Configuration & Workflows (20%).
Prompting and reliability
| Term | What it means |
|---|---|
| Few-shot prompting | A few worked examples, with their reasoning and at least one edge case. |
| Prompt chaining | Extract, validate, format, synthesise, with each step its own focused prompt. |
| Retry with error feedback | Resending the original prompt with the failed output and the specific validation error. |
| Batch API | Asynchronous processing at roughly half the cost, with no latency guarantee. |
| Prompt caching | Reusing a cached prompt prefix across requests, for lower cost and latency. |
| Persistent fact block | Specific facts carried forward word for word rather than summarised. |
| Scratchpad file | Notes written to disk so they survive a context reset. |
| Stratified validation | Measuring accuracy per category, sampled in proportion, rather than in aggregate. |
These run through Prompt Engineering & Structured Output (20%) and Context Management & Reliability (15%).
The exam asks which term applies to a system that is failing, and Never the answer on the Claude Architect exam shows how the wrong options lean on these same terms.
