AI Vocabulary Cheat Sheet

AI / Machine Learning / LLM

AI is the umbrella. ML is software that learns patterns from data. LLMs (Large Language Models) are a specific kind of ML trained on massive text to predict the next token. ChatGPT, Claude, Gemini are LLM-powered.

Token

The unit an LLM reads and writes. Roughly 1 token = 4 characters or 0.75 words. 'Hello world' is ~2 tokens. Pricing and context windows are measured in tokens.

Context window

How much text the model can hold at once (input + output). Modern models: 128K to 1M+ tokens. If your conversation gets too long, the model 'forgets' the beginning.

Temperature

Creativity dial (0 to 1). Low = deterministic, predictable, factual. High = varied, creative, weirder. Use low for compliance/legal, high for brainstorming.

Hallucination

When the model confidently states something untrue. Always possible. Mitigation: ground the model in real source documents and require citations.

Prompt

Your instruction to the model. The single biggest lever on output quality.

System prompt

The standing instruction that sets behavior across a whole conversation or project. Where you define role, voice, constraints, and rules.

RAG (Retrieval-Augmented Generation)

Letting the model search a private knowledge base before answering. How Claude Projects, Custom GPTs, and most internal AI tools work.

Fine-tuning

Training a base model on your specific data so it permanently learns your style. Powerful, expensive, rarely the right first move. RAG usually wins.

Agent

An AI that can take actions (search the web, use tools, write files, send emails) — not just answer questions. We cover this in Module 10.

Frontier models

The current top tier: Claude (Anthropic), GPT-4/5 (OpenAI), Gemini (Google). All are general-purpose, all are getting better monthly.

Open-weights models

Models you can download and run yourself: Llama (Meta), Mistral, Qwen, DeepSeek. Useful for privacy, cost, and customization.

Search-augmented LLMs

LLMs that browse the web for fresh info: Perplexity, Google AI Mode, ChatGPT search.

Multimodal

Models that handle text + images + audio + video. All frontier models are multimodal now.

Role + Goal + Voice + Constraints + Format

The five-part recipe. Tell the model who it is, what it's doing, in whose voice, with what constraints, and in what format. Use this once and you'll never go back.

Few-shot examples

Showing 2-3 examples of input → output before asking the model to do it. The single most powerful upgrade to any prompt.

Chain-of-thought

Asking the model to think step by step. Improves reasoning on complex tasks. 'Walk through your reasoning before giving the answer.'

Reflection / self-critique

After the first answer, ask: 'Critique this. What's weak? Rewrite to fix.' Output quality jumps.

Negative constraints

What NOT to do. 'Do not use the word leverage. Do not use bullet points.' Often more useful than positive instructions.

Data privacy

Anything you paste into a consumer AI tool MAY be used to train future models (varies by provider). Use enterprise/Team plans for any client or sensitive data. Read the data-handling page of any tool you adopt.

Hallucination in outbound work

AI making up facts that go to customers, clients, or the public. Mitigation: humans review outbound for at least 90 days. Specific facts (dates, prices, names, citations) get verified.

Shadow AI

Employees using AI tools nobody at leadership knows about. Not bad — but it means risks aren't being managed. Surface it gently. Don't punish.

Bias

LLMs reflect the bias of their training data. Most visible in: hiring, performance reviews, customer-facing tone. Always have a human in the loop for people-impact decisions.

Copyright / IP

Images and text from generative AI live in murky legal territory. Don't use AI-generated assets in places where ownership matters (logos, trademarked content, contracts) without legal review.

Audit trail

For regulated work (legal, medical, financial), log what AI was used for, by whom, with what inputs. Most enterprise tools support this.

Automation platforms

Tools that connect apps and let an AI step run inside the chain. Trigger → AI does the thing → Output goes somewhere.

Trigger / Action

Trigger = the event that starts it (new email, new lead, new file). Action = what happens after (AI summarizes, then posts to Slack).

Webhook

A URL that listens for events. Your AI gets called when something happens elsewhere. The glue between systems.

Coding assistant

AI in your code editor. Suggests completions, explains code, writes tests, refactors. Multiplies developer productivity.

Agentic coding

AI that does the work end-to-end — reads files, plans, edits, runs tests, iterates. Cursor's agent mode and Claude Code are examples.

MCP (Model Context Protocol)

Open protocol that lets AI tools securely connect to your databases, APIs, and services. Why Claude Code can talk to your GitHub, Slack, Linear, etc.

AI coworker

An AI tool that acts on your behalf, in context, where you actually work — not just inside a separate chat window.

Autonomous agent

AI that takes a goal and does the multi-step work itself, checking in only when it needs to.

Project / Custom GPT / Workspace

A persistent container with a system prompt + knowledge files + chat history. Claude calls it Projects. ChatGPT calls them Custom GPTs. Same idea.

Knowledge files

Documents you upload once that the AI can reference in every conversation. Style guides, SOPs, FAQs, examples.

System prompt (project-level)

The standing instructions for the project: who the AI is, how it should behave, what to never do.

Conversation hygiene

Start a fresh conversation when the topic changes. Long conversations dilute focus and burn through context.

Free tier patterns

Most tools offer: unlimited basic use with caps, OR free trial with all features. Read which is which before adopting.

When to upgrade

Upgrade when free is genuinely blocking you, not when you 'might want more.' Most people overspend by upgrading too early.

Chatbot vs agent

Chatbot answers your question. Agent takes a goal and figures out the steps itself, taking real actions across systems.

When you need an agent

Multi-step tasks across multiple systems, where the steps can vary based on what's found. If a workflow is the same every time, you don't need an agent — you need automation (Module 5).

Agent platforms

Frameworks/products that let you build or run agents: Claude Agent SDK, OpenAI Agents, AutoGen, CrewAI, LangGraph. Plus consumer agents like Manus and Devin.

Tool use / function calling

Letting the model call APIs, search, edit files. The mechanical foundation of every agent.

Evals

How you measure whether an agent is doing the right thing reliably. Production agents without evals will silently break.