Specific syntax and settings appear as exact-recall questions. Know these cold.
CLAUDE.md hierarchy
| Location | Scope |
|---|---|
~/.claude/CLAUDE.md | Personal; not in version control |
.claude/CLAUDE.md | Team-wide; shared via version control |
CLAUDE.md in a subdirectory | Directory-specific conventions |
.claude/rules/ + YAML paths: | Conditional loading by file glob |
Slash commands and skills
.claude/commands/— project-scoped, shared.~/.claude/commands/— personal.context: fork— run in an isolated subagent.allowed-tools:— restrict tool access.argument-hint:— prompt for input.
Claude Code CLI flags (CI/CD)
-p/--print— non-interactive; required for pipelines.--output-format json— structured output.--json-schema— validate output against a schema.--resume <name>— resume a named session./compact— compress context (can lose numeric detail).
MCP configuration
.mcp.json— project-level, shared via version control; use${ENV_VAR}for secrets.~/.claude.json— user-level, experimental or personal servers.
Agent SDK
AgentDefinition— name, system prompt, allowed tools.- The spawn tool (
Task) — the coordinator's allowed tools must include it. stop_reason—tool_use,end_turn,max_tokens,stop_sequence.PreToolUsehook — intercept outgoing calls (compliance).PostToolUse— intercept results before the model sees them.
tool_choice and Batches
"auto"— model decides: tool or text."any"— must call some tool.{"type":"tool","name":"…"}— must call that tool.- Message Batches — 50% off, up to 24 hours, no latency guarantee, no multi-turn tool calling,
custom_idto correlate results. - Prompt caching — up to 90% off repeated context, immediate.
Study with others. Join the free WhatsApp study community — no cost, leave any time.
Study notes, not official material. This guide is written by Agmo Academy to help you prepare. It is not affiliated with or endorsed by Anthropic — always confirm details against Anthropic's current exam guide before you sit the exam.