The CCA-F exam assumes a shared vocabulary. If a scenario says the agent "stopped with stop_reason: max_tokens" or asks where a rule belongs — CLAUDE.md, a hook, or a permission — you need those terms to be second nature. This glossary defines the concepts the exam leans on, grouped by domain. Each definition is deliberately short; the linked domain guides go deep.
Every entry reflects how Claude, the Anthropic Messages API, Claude Code, and MCP actually behave as of mid-2026. Where a term names a specific API value or file path, it is stated exactly.
Agentic architecture & orchestration (D1)
| Term | Definition |
|---|---|
| Agentic loop | The cycle your orchestration code runs: call the model, read stop_reason, execute any requested tools, append results, and call again — until the model finishes. The loop is code you own, not a model feature. |
stop_reason | The field on every Messages API response saying why the model stopped: tool_use, end_turn, max_tokens, stop_sequence, pause_turn, or refusal. Your loop branches on it. |
tool_use (stop reason) | The model is requesting one or more tool calls. Execute them, append the tool_result blocks as a user message, and resend to continue. |
end_turn | The model finished its response naturally. The loop terminates and returns the final answer. |
pause_turn | A long-running server-side tool flow (e.g. web search) paused at an iteration limit. Resume by resending the conversation — no manual 'continue' message needed. |
| Terminal guarantee | The design invariant that every session ends resolved or escalated to a human with context — never a silent hang, whatever stop_reason fires. |
| Handoff package | A structured payload carrying findings, pending action, and — critically — authorization state when control passes to another step or a human, so nothing is re-derived or skipped. |
| Session resumption | Restoring an interrupted agent by injecting prior findings and re-analyzing only what changed — not replaying the whole history and re-running side-effecting tools. |
| Dynamic task decomposition | Generating and revising subtasks as new information surfaces, instead of executing a fixed pre-planned sequence that ignores intermediate findings. |
Tool design & MCP (D2)
| Term | Definition |
|---|---|
| MCP (Model Context Protocol) | An open, Anthropic-originated standard for connecting AI applications to external systems through a client-server architecture. Servers expose tools, resources, and prompts. |
| MCP server scope | Where a server is configured in Claude Code: local (default, this project, private), project (shared via a committed .mcp.json), or user (all your projects, private). |
| Tool (function) schema | A tool's name, description, and input_schema. Claude selects and calls tools from these; the schema shapes the arguments. |
tool_result | The block you return after executing a tool. Set is_error: true on failure and make the content structured (error type, recoverability, suggested next action). |
tool_choice | Request parameter controlling whether/which tool is called: auto, any, tool (force a named tool), or none. Forcing a tool guarantees invocation. |
| Environment-variable expansion | In .mcp.json, Claude Code expands ${VAR} and ${VAR:-default} so secrets stay in the environment and only the variable name is committed. |
| Tool discovery | Verifying Claude actually sees a server's tools — via /mcp in a session or claude mcp list / claude mcp get <name> on the CLI. |
Claude Code configuration (D3)
| Term | Definition |
|---|---|
CLAUDE.md | Auto-loaded guidance the model reads and usually follows — good for conventions and context, but discretionary. Not a place for rules that must always hold. |
| Custom slash command | A Markdown prompt invoked as /name. In .claude/commands/ (project, shared) or ~/.claude/commands/ (user, personal). Now unified with Skills. |
| Skill | A SKILL.md-defined capability whose body loads only when used (on demand or when relevant). Good for multi-step procedures that would bloat every prompt. |
.claude/rules/ | Glob-scoped guidance that applies only to matching files/paths — precise instructions without loading into every session. |
| PreToolUse hook | A command the harness runs before a tool executes; can allow, deny, or ask. Deterministic enforcement outside model discretion. |
| PostToolUse hook | A command run after a tool completes (matcher e.g. Edit|Write) — ideal for auto-format/lint/test on every file change, regardless of instruction-following. |
| Settings permissions | Allow/deny/ask rules on tools and commands in .claude/settings.json. Evaluated deny > ask > allow; a deny is absolute. |
| Plan mode | A read-only mode where Claude proposes an approach before editing. Use for large-scope, hard-to-reverse, uncertain, or review-gated work. |
Prompt engineering & structured output (D4)
| Term | Definition |
|---|---|
| Structured output | Getting the model to emit a defined shape. Reliability ladder: strict tool use / output_config.format (strongest) → non-strict tool schema → prompt-based JSON (weakest). |
| Strict tool use | "strict": true on a tool — Anthropic compiles the JSON schema into a grammar and constrains sampling, so tool_use.input is guaranteed to validate. |
output_config.format | The structured-outputs feature that constrains the model's response text itself to a JSON schema (rather than a tool call). |
| Nullable / optional field | Schema design letting the model represent missing data (leave out of required, allow null) instead of fabricating a value. |
| Enum with 'unknown' | A closed-vocabulary field that includes an explicit unknown member, giving ambiguity a home rather than a guess. |
| Few-shot examples | One or two worked input→output pairs placed in earlier turns to anchor formatting and edge cases — more effective than emphatic instructions. |
| Assistant prefill | Legacy technique of seeding the assistant turn (e.g. {) to force JSON. Now returns a 400 on current Claude models — use output_config.format or a tool schema instead. |
| Message Batches API | Asynchronous, high-volume processing: submit many requests, poll in_progress → ended, results keyed by custom_id. ~50% cheaper, up to a 24-hour window. |
Context management & reliability (D5)
| Term | Definition |
|---|---|
| Stateless API | The Messages API keeps no memory between calls. Multi-turn state exists only because the client resends the full messages history (including tool turns) each request. |
| Context window | The finite token budget shared by the system prompt, all turns, tool definitions/results, and files read. Overfilling it degrades the model's attention. |
| Context rot | The degradation in coherence as a context window fills with low-value material — the reason to control what enters context, not just raise the limit. |
| Subagent isolation | Delegating a noisy subtask to a subagent with its own context; only a distilled result returns to the parent, keeping the main thread clean. |
| Scratchpad file | Findings persisted to disk instead of held in the window — durable, effectively unbounded, and survives context resets and new sessions. |
| Escalation criteria | The rules for handing off to a human: honor an explicit escalation request immediately; escalate on irreversible, out-of-policy, or unsafe actions. |
| Confidence-based routing | Directing human review by risk — confidence scores, document characteristics, field-level ambiguity — rather than random sampling, so scarce review lands on likely errors. |
Exam & scoring
| Term | Definition |
|---|---|
| CCA-F | Claude Certified Architect – Foundations: Anthropic's entry-level architect certification. Five weighted domains, 29 objectives, scenario-based. |
| CCA-P | Claude Certified Architect – Professional: a separate, more advanced certification. Not covered by this platform. |
| Domain weighting | How the exam allocates questions: D1 27%, D2 18%, D3 20%, D4 20%, D5 15%. Study time should follow it. |
| Scaled score | The exam's 100–1000 result (720 to pass), produced by Anthropic's private item-response weighting. No public formula converts practice accuracy to it — beware any tool that predicts one. |
| Per-objective breakdown | The honest signal a practice tool can give: raw score, percentage, and accuracy per objective — the basis for targeted study. |
Frequently asked
What's the single most important term to understand for D1?
stop_reason. The whole agentic loop is your code reading that field and deciding whether to run tools and continue (tool_use) or terminate (end_turn), while handling max_tokens, pause_turn, and refusal explicitly.
What's the difference between a hook and a permission in Claude Code?
Both are deterministic. A permission is an allow/deny/ask rule on a tool or command, evaluated deny > ask > allow. A hook is a shell command the harness runs before (PreToolUse) or after (PostToolUse) a tool — useful for gating actions or auto-running format/lint/test.
Is 'prefill' still worth knowing for D4?
Know the concept, but know it's dated: seeding the last assistant turn to force JSON now returns a 400 on current Claude models. The modern tools are strict tool use and output_config.format.
Independent, unofficial study material. Not affiliated with, endorsed by, or authorized by Anthropic. Every example is original and written to teach the public exam objectives — no real exam questions are reproduced. Technical details reflect Claude, the Anthropic API, Claude Code, and MCP as of July 25, 2026; always confirm specifics against current official documentation.
Test yourself
Turn what you just read into answers you can check.