Exam guide vs current docs: what to answer
Last verified October 27, 2026. The exam is written against Exam Guide v1.0 (July 2026). The products keep moving. Where they differ, the guide decides the keyed answer; knowing the current behavior keeps you from being confused in practice.
| Topic | What the guide tests | What the current docs say | On the exam |
|---|---|---|---|
Forced tool_choice | "auto", "any" and forced {"type": "tool", "name": ...}; "any" guarantees a tool call when the document type is unknown; forcing a named tool makes it run first | Claude Opus 5.5, Sonnet 5.5 and Fable 5.1 reject "any" and forced selection with a 400; you steer with "auto" plus an instruction, strict: true on the tool, or structured outputs | Answer with "any" / forced selection as the guide describes |
| Spawning subagents | The Task tool; a coordinator's allowedTools must include "Task" | The tool is named Agent in allowedTools and tool_use blocks in current Agent SDK versions (it still appears as Task in the init tools list) | Use the guide's name, Task |
Skill frontmatter allowed-tools | Restricts the tools a skill may use during execution | Pre-approves the listed tools so they run without a permission prompt; disallowed-tools is the field that removes tools | The option that limits tools via allowed-tools is the keyed answer |
| Checking which memory files loaded | /memory verifies which memory files are loaded and diagnoses inconsistent behavior | /memory lists and opens the memory files; /context shows which CLAUDE.md and rules files loaded at launch | /memory |
Things that have not changed and are safe to rely on
- Agentic loop: continue on
stop_reason == "tool_use", finish on"end_turn", append the assistant turn and thetool_resultblocks to history. - Hooks: a pre-execution hook (
PreToolUsein the Agent SDK) can deny a tool call;PostToolUsecan rewrite a result before the model sees it. - CLAUDE.md hierarchy (user, project, directory),
@pathimports with a four-hop limit,.claude/rules/files withpaths:globs that load on demand; a rule withoutpathsloads always. - MCP scopes: project
.mcp.json(shared,${VAR}expansion) versus user~/.claude.json; tools from all configured servers are available at once. - CLI:
-p/--print,--output-format json,--json-schema,--resume <id or name>,--fork-session. - Message Batches: 50% discount, up to 24 hours,
custom_id, results in any order, no tool loop inside a request.
Every MockCerts explanation that touches one of the four discrepancies carries a one-line "current-docs note", so you learn both the exam answer and the real behavior. Start with the free mock.
Practice
Take the free 25-question diagnostic, no sign-up.Not affiliated with Anthropic. MockCerts is independent practice material. "Claude" is a trademark of Anthropic, PBC, used only to identify the exam. Questions are original and written against the public exam guide. Content-integrity statement.