The 10 most frequently encountered terms when using Claude, in order of importance: Token (unit of measurement), Context Window (memory limit), Hallucination (incorrect confident output), Temperature (randomness control), System Prompt (background instructions), Prompt Engineering (instruction optimization), RAG (external knowledge integration), Fine-Tuning (specialized model training), AI Agent (autonomous multi-step tasks), MCP (external tool connection protocol). Understanding these 10 terms lets you comprehend 90%+ of Claude and AI-related discussions.
These 10 terms matter because they represent the most critical "decision points" when using AI tools: when Hallucination is most likely to occur (telling you when to verify), how Context Window limitations affect your usage (telling you how to organize information), when RAG versus Fine-Tuning is appropriate (telling you how to select tools). Mastering these terms isn't just knowing word definitions — it's understanding AI tools' underlying logic, enabling more strategic use.
The most direct impact of these 10 terms is enabling you to correctly diagnose problems when Claude output goes wrong. Claude gives a confident-sounding answer that seems suspicious? That might be Hallucination — you know to verify. Conversation gets longer and Claude starts forgetting earlier content? That's a Context Window issue — you know to open a new conversation or restate key information. You're using a Claude-based tool but it behaves differently from Claude.ai? That's a System Prompt difference — you know what's happening behind the scenes.
Convert understanding of these 10 terms into practical actions: in your Claude.ai Projects Instructions, use the System Prompt concept to set your frequently used instructions; when Claude gives you an answer you're uncertain about, ask "how certain are you about this information? Are there parts I should verify myself?" (countering Hallucination); if you're using Claude via API, track your Token usage at console.Anthropic.com/settings/usage to avoid exceeding budget at month-end.
AI terminology intimidates many people. Token, Hallucination, Context Window, Temperature — these words appear in every Claude tutorial, but they're rarely explained in terms of what they actually mean for practical use.
This article covers 10 of the most essential terms, each with an "immediately applicable understanding" rather than a textbook definition.
Immediately applicable understanding: A Token is the unit Claude uses to measure text length — roughly 0.75 of an English word, or 1-2 Chinese characters. It determines how much content can fit in a conversation (Context Window) and the cost of API usage.
Why it matters: When using Claude via API, costs are calculated per Token, not per word. A 1,000-word English article is approximately 1,300-1,500 Tokens.
Immediately applicable understanding: The Context Window is the maximum total text Claude can "remember" within one conversation. Like short-term memory — beyond this limit, Claude starts "forgetting" content from the beginning of the conversation.
Claude's Context Window is 200K Tokens — large, but it still has boundaries when handling very long documents.
Immediately applicable understanding: Hallucination is when Claude states incorrect information with a confident tone. It's not lying — it's using its prediction mechanism to output "the most likely answer," which happens to be wrong.
When it's most likely to occur: asking for specific numbers, dates, or names; asking about events after the training cutoff; asking about very niche or specialized domain knowledge.
Immediately applicable understanding: Temperature controls the "randomness" of Claude's output. Low Temperature (near 0): nearly identical outputs each time — good for tasks requiring precision and consistency (code, data analysis). High Temperature (near 1): more varied, more creative outputs — good for creative writing.
Immediately applicable understanding: The System Prompt is "behind-the-scenes instructions" sent to Claude before conversation begins, defining its role, what it can do, and what it can't. The Instructions you set in Claude.ai Projects is a form of System Prompt.
Immediately applicable understanding: Designing and optimizing the instructions you give Claude so it outputs results that better match your needs. Not programming — anyone can learn it. Core: clearly tell Claude who, for whom, to achieve what, with what constraints.
Immediately applicable understanding: A technique that lets AI "look up information" before answering — placing your external documents or knowledge base somewhere searchable, AI searches relevant fragments before answering, then generates a response based on the search results. Claude Projects' document upload is essentially a simplified RAG.
Immediately applicable understanding: Re-training an existing model on your specific data to make it perform more consistently on your specific tasks. Analogy: pre-training is general university education; Fine-Tuning is on-the-job training. Much more costly than RAG, but eliminates the need to provide background context every conversation.
Immediately applicable understanding: An AI that can autonomously execute multi-step tasks — not just answering questions, but "deciding how to proceed, which tools to use, executing actions, evaluating results, and adjusting." Claude Code is a type of Agent.
Immediately applicable understanding: A standard protocol that connects Claude to external tools (GitHub, Google Drive, databases, etc.) — like AI's version of a USB interface. With MCP, you can have Claude directly read your Google Drive files, open GitHub Issues, or query your database without manually copying and pasting.