CCAF Guide Study Guide CLI CC Settings Mock Exam Glossary SDK Config

Quick-Reference Cheat Sheet

Last-minute review — one line per concept from the exam blueprint appendix. Skim this the morning of your exam.

Claude Agent SDK

Agentic loop

Drive control flow with stop_reason: continue on tool_use, stop on end_turn.

stop_reason

tool_use = run tools & iterate; end_turn = finished. Never parse text to decide.

Task tool

Spawns subagents. allowedTools must include Task. Multiple calls in one response = parallel.

allowedTools

Per-agent tool restriction. Scope to role — too many tools degrades selection.

Hooks

PostToolUse normalizes results; interception hooks block outgoing calls. Deterministic vs prompt.

fork_session

Branch from a shared baseline to explore divergent approaches.

Model Context Protocol (MCP)

Tool descriptions

PRIMARY selection signal — include inputs, examples, edge cases, boundaries.

isError flag

Signals tool failure back to the agent.

Structured errors

errorCategory (transient/validation/permission), isRetryable, human-readable text.

MCP resources

Expose content catalogs (schemas, issue summaries) to cut exploratory calls.

.mcp.json

Project scope (shared via VCS). Use ${ENV_VAR} expansion for secrets.

~/.claude.json

User scope — personal/experimental servers, not shared.

Claude Code

CLAUDE.md hierarchy

user (~/.claude) → project (.claude or root) → directory. User-level isn't shared.

@import

Reference external files to keep CLAUDE.md modular.

.claude/rules/

YAML paths: glob frontmatter for conditional, path-scoped conventions.

.claude/commands/

Project slash commands, shared via VCS.

.claude/skills/

SKILL.md frontmatter: context: fork, allowed-tools, argument-hint.

Plan vs direct

Plan = complex/architectural/multi-file; direct = simple/well-scoped.

--resume / --continue

--resume <name> = specific session; --continue = most recent.

/memory, /compact

Inspect loaded memory; compact context during long sessions.

Explore subagent

Isolate verbose discovery; return a summary to preserve context.

Claude Code CLI

-p / --print

Non-interactive mode for CI — prevents input hangs, exits after output.

--output-format json

Machine-parseable output for pipelines.

--json-schema

Enforce a structured output shape in CI.

Claude API

tool_use + JSON schema

Most reliable structured output; eliminates syntax errors (not semantics).

tool_choice

auto (may text), any (must call a tool), forced {type:'tool',name:'...'}.

Nullable fields

Make optional when source may lack data — prevents fabrication.

Message Batches API

Trade-offs

50% cheaper, up to 24h, NO latency SLA. No multi-turn tool calling.

custom_id

Correlate request/response; resubmit only failed IDs.

Fit

Overnight/weekly jobs — never blocking pre-merge checks.

Reliability & Context

Lost in the middle

Put key findings at start/end; use section headers.

Case facts block

Persist exact numbers/dates outside summarized history.

Escalation triggers

Explicit human request, policy gaps, no progress — NOT sentiment/confidence.

Provenance

Preserve claim→source mappings + dates through synthesis; annotate conflicts.

Confidence calibration

Field-level scores + labeled sets; segment accuracy by doc type/field.

Ready? Test yourself with the practice quiz →