Advibesor field guide

A glossary of Claude Code

17 terms that show up once you start vibe-coding for real: the tools an agent calls, the modes it can run in, what it remembers, and the rules underneath. Defined plainly, no jargon left unexplained.

17 / 17

AskUserQuestion: Stops the agent mid-task to ask you a real question, multiple choice, when a decision belongs to you and not to a guess. No silent assumptions buried three files deep.
Agent (Subagent): Spawns a fresh Claude instance with its own context window to go handle one job on its own: research a codebase, write a component, review a diff. It reports back a summary; its scratch work never clutters the main conversation.
Bash: The tool behind every git commit, npm install and file move the agent makes. A real, persistent shell, not a simulation, so state like your current directory carries between calls.
Edit: Changes a file by exact string replacement instead of rewriting it top to bottom. Smaller diffs, fewer accidents. A file has to be read before it can be edited.
Read / Grep / Glob: Three tools that split file work between them: Glob finds files by pattern, Grep searches what is inside them, Read opens the exact lines. Together they are how the agent explores a codebase without guessing.
Skill: A packaged playbook of instructions loaded into the conversation for one specific job: deploying, reviewing, drawing a diagram. Some trigger automatically when a request matches; others are called by name.
Artifact: A hosted, shareable web page the agent can publish straight out of a conversation. This glossary is one: written, styled and put online without ever leaving the chat.
ReportFindings: A structured, typed way to file code review results (file, line, severity) instead of dumping paragraphs of prose. Built for the moment output needs to become UI, not just text.
CLAUDE.md: A plain markdown file, one per project and one per user, read automatically at the start of every session. The closest thing Claude Code has to long-term memory: instructions here persist across every future conversation until someone edits the file.
System Prompt: The instructions the agent receives before your first message: what tools exist, how to behave, what today's date is. Invisible by default, but the reason the agent already knows the rules before you have said a word.
Context Compaction: When a long conversation grows close to the model's context window, older turns are automatically summarized to make room. The conversation does not stop, it just gets a shorter memory of its own early chapters.
Plan Mode: A mode where the agent is only allowed to look, not touch. It researches, then proposes a step by step plan for approval before a single file changes.
Extended Thinking: A private budget of reasoning the model can spend deliberating before it answers, sometimes called 'ultrathink' at its highest setting. You see the conclusion; the scratch work stays out of the way.
Permission Mode: The dial between "ask before every single action" and "just handle it." It decides whether a tool call runs immediately, waits for a yes, or requires a plan first.
MCP: Model Context Protocol: an open standard for connecting Claude Code to outside tools and data, a database, a design file, a ticket tracker, through one common interface instead of a custom integration for each one.
Hook: A shell command wired to a moment in the agent's lifecycle: before a tool runs, after a session starts, when the agent stops. It lets you intercept and enforce rules deterministically, not just by asking nicely in a prompt.
Sandbox: An isolated execution boundary, filesystem and network, that commands can run inside without asking permission for every step. The fence that makes "just handle it" safe to say.

Learning to vibe-code is mostly learning the vocabulary. Advibesor turns it into quizzes you can swipe through in a minute a day.

Start learning →