Glossary and cheat sheet
Glossary
Term Meaning
Artifact Substantial output (document, code, diagram, small app) shown in its own panel for editing, reuse and sharing
Augment vs redesign Speeding up a step in an existing process, vs changing the process because Claude exists Connector A link letting Claude read content in another app, such as Google Drive or Gmail, using your permissions
Context window Everything Claude can see at once: instructions, conversation, files and tool results, measured in tokens
Data minimization Sharing only the data a task needs
Extended thinking A mode where Claude reasons step by step before answering; better for hard problems, slower and costlier
Few-shot prompting Including examples of the desired output in the prompt Grounding Telling Claude to base its answer on supplied sources
Hallucination Fluent, confident output that is false or invented Haiku / Sonnet / Opus Claude's model tiers: fastest and cheapest / balanced default / most capable Incognito chat A chat not saved to history or memory
MCP (Model Context Pro‐
tocol)
An open standard for connecting Claude to tools and data sources Memory Claude's ability to carry context and preferences across chats; viewable and editable Project A workspace with its own instructions and knowledge, shared across its chats Project instructions Standing guidance applied to every chat in a Project, similar to a system prompt Prompt chaining Using one step's output as the next step's input Prompt injection Instructions hidden in content Claude reads, such as a document or web page, that try to redirect it
Research mode Claude runs many searches and returns a cited report Scaled score A score converted to a 100–1,000 scale; 720 passes CCAO-F Skill A packaged set of instructions, templates and scripts Claude loads when a task needs it Task decomposition Breaking a complex request into ordered, reviewable steps Token The unit Claude measures text in; roughly three-quarters of an English word
One-page cheat sheet
Feature: chat for one-offs, Project for recurring context, Research for cited multi-source answers, Artifact for reusable output, code execution for numbers and files. Model: Haiku for volume, Sonnet by default, Opus for hard reasoning. Start small; move up only when quality falls short.
Prompt: task, context, audience, constraints, format, example, sources.
Verify: confidence is not evidence. Check citations, numbers and regulated claims at the source.
Review: scale human review to stakes and reversibility. People decisions always get human judgment.
Configure: instructions say how; knowledge says what. One current version per source; an owner and a review cadence.
Govern: classify data, minimize, anonymize, use approved tools. "Don't remember this" is not a control.
Escalate: APIs, unreviewed automation and large-scale integration go to technical teams. Troubleshoot: diagnose, change one thing, then make the fix permanent in configuration.