📖 Glossary
Agentic loop
Claude's core cycle: gather context (read/grep) \u2192 take action (edit/run) \u2192 verify (tests) \u2192 repeat until the task is done.
REPL
The interactive read\u2013eval\u2013print loop you get after typing claude \u2014 you type, it responds, in one running session.
CLAUDE.md
Markdown project-memory file Claude reads at the start of every session: commands, style, architecture, gotchas.
Permission mode
How much Claude may do without asking: plan, default, acceptEdits, or bypass. Cycle with Shift+Tab.
Plan mode
Read-only mode where Claude explores and proposes a plan without editing source, gated by your approval.
Skill
A reusable packaged workflow in .claude/skills/<name>/SKILL.md, invoked with /name or auto-selected when relevant.
Subagent
A specialized assistant running in its own isolated context window for side tasks; defined in .claude/agents/<name>.md.
MCP
Model Context Protocol \u2014 an open standard connecting Claude Code to external tools and data (GitHub, databases, browsers).
Hook
A shell command that runs automatically at a lifecycle event (e.g. PostToolUse) \u2014 deterministic, configured in settings.json.
Headless
Non-interactive mode: claude -p "prompt" runs one task and exits \u2014 the basis for scripting and CI.
Checkpoint
A saved state you can rewind to with Esc Esc, undoing recent file changes.
UCC-1
A financing statement filed to perfect a lender's security interest in a debtor's collateral.
\u00a79-503
The UCC rule requiring a registered organization's exact legal name on a filing; a wrong name can be seriously misleading.