Beginner’s guide · Step 3 of 5
AI vocabulary the words you need
Prompt, token, hallucination, agent… AI has its own jargon. Here are the words you will come across most often, explained plainly and ordered from simplest to most advanced.
If you only remember 3 words
Prompt
what you ask
Hallucination
when the AI makes things up
Context
what it has in front of it
34 terms
- Artificial intelligence (AI) The basics
- A set of techniques that let a machine perform tasks that used to require human intelligence: understanding text, recognising images, translating, reasoning.
- Generative AI The basics
- AI that creates new content (text, images, sound, video, code) from a request. Claude, Gemini and ChatGPT are generative AIs.
- Large Language Model (LLM) The basics
- The “engine” of an assistant like Claude or Gemini. A program trained on huge amounts of text to predict the most likely continuation of a sentence.
- Model The basics
- A specific version of an AI, with its own level of power and speed. One assistant often offers several models. Claude: Haiku, Sonnet, Opus. Gemini: Flash, Pro.
- Training The basics
- The phase in which the model learns by analysing billions of texts. It happens before you use it: your conversations do not change it live.
- Token The basics
- The small piece of text the AI reads and writes: a short word, part of a word or a punctuation mark. Usage limits and prices are counted in tokens. In English, 100 tokens are roughly 75 words.
- Context window The basics
- How much text the AI can keep “in view” at once: your conversation, attached documents, its instructions. Beyond that, it forgets the beginning. Today’s large models can read the equivalent of several books at once.
- Hallucination The basics
- A wrong answer given with confidence: an invented figure, a quote that does not exist, an imaginary source. It is the main reason to always check. Ask the AI for its sources, then open them.
- Multimodal The basics
- Describes an AI that understands several kinds of content: text, images, PDFs, audio, video. Take a photo of a chart and ask it to analyse it.
- Reasoning (thinking) The basics
- A mode in which the AI “thinks” step by step before answering. Slower, but much better for complex problems, calculations and logic.
- Knowledge cutoff The basics
- The date at which the model’s training data stops. For current events, the AI has to search the web.
- Prompt Everyday use
- The request you write to the AI. Its quality drives the quality of the answer: context, goal, expected format, example. “Summarise this meeting report in 5 points for my team, neutral tone.”
- System instructions Everyday use
- Standing instructions given to the AI before the conversation (role, tone, rules). You do not see them in the exchange, but they shape every answer. Also called a “system prompt”.
- Conversation (thread) Everyday use
- An ongoing exchange with the AI. It remembers everything said in the thread, but not, by default, your other conversations. New topic? Start a new conversation.
- Artifact Everyday use
- In Claude, content created next to the conversation (document, web page, spreadsheet, diagram, mini-app) that you can edit, download or share.
- Web search Everyday use
- The AI’s ability to look things up online while answering, to find recent information and cite its sources.
- Deep research Everyday use
- A mode where the AI runs dozens of searches, reads the pages and writes a structured, sourced report. Takes a few minutes. Called Deep Research in Gemini, Research in Claude.
- Memory Everyday use
- A feature that lets the AI remember things about you across conversations (your job, your preferences). You can view and erase it in settings.
- Usage limit Everyday use
- The number of messages or amount of work allowed over a period. It is lower on free plans and increases with a subscription.
- Project Customising
- A space grouping conversations, reference documents and shared instructions on one topic. A “Website” project with your brand guidelines and copy.
- Gem Customising
- Gemini’s former custom assistant (a name, instructions, documents). Google is replacing it with skills from late 2026, with automatic conversion. Remember to convert your Gems into skills.
- Custom GPT Customising
- In ChatGPT, an assistant preconfigured for a specific task.
- Skill Customising
- A folder of instructions, templates and sometimes scripts that teaches the AI a specific method. The AI loads it automatically when the task calls for it. Same format in Claude and Gemini. A “quote” skill that always applies your layout.
- Style Customising
- A setting that adapts how the AI writes (concise, explanatory, formal) or imitates your own style from examples.
- Agent (agentic AI) Going further
- An AI that does not just answer: it acts. It breaks a mission into steps, uses tools, checks its work and keeps going until the result is reached. Claude Code, which edits and tests code on its own.
- Tool use Going further
- A capability given to the AI to act on the world: search the web, run a calculation, read a calendar, send an email.
- Connector Going further
- A link between the AI and an app (Gmail, Drive, Calendar, Notion, Slack…) so it can read or create information there.
- MCP (Model Context Protocol) Going further
- An open standard, launched by Anthropic, that acts as a “universal plug” to connect an AI to apps and data. Most connectors are built on it.
- API Going further
- A technical entry point to use a model from a program, without the chat interface. Billed per use, in tokens. An Apps Script that calls the Gemini API to sort your emails.
- RAG Going further
- A technique that retrieves relevant passages from your documents before answering, so the AI relies on your sources rather than its memory. RAG stands for Retrieval-Augmented Generation.
- Fine-tuning Going further
- Re-training an existing model on your own examples to specialise it. Rarely needed today: good instructions or a skill are often enough.
- Temperature Going further
- A technical setting (via the API) that controls creativity: low makes the AI predictable; high makes it more inventive but less reliable.
- Open-weight model Going further
- A model whose “weights” are published: you can download it and run it on your own servers. Mistral, Llama or Gemma are examples.
- Prompt injection Going further
- An attack where booby-trapped text (in a web page, email or document) tries to give hidden orders to the AI. A real security issue for agents. Never let an agent act on sensitive data without oversight.
No term matches your search.
Frequently asked questions
What is the difference between AI, generative AI and an LLM?
AI is the overall field. Generative AI is a branch of it that creates content. The LLM (large language model) is the technology that powers text-generating AIs like Claude or Gemini.
Why talk about tokens rather than words?
Because the model does not read words but pieces of words. Limits (document size, conversation length) and API prices are therefore expressed in tokens.
How can I avoid hallucinations?
Give the AI your own documents as a source, turn on web search, ask it to cite its sources and to say when it is unsure, then check the important facts.
Skill, Gem, project: what is the difference?
A project groups conversations and documents on a topic. A Gem was Gemini’s custom assistant, which Google is now replacing with skills. A skill is a method the AI loads automatically when the task calls for it, in any conversation.
FF
Développeur freelance spécialisé dans l'optimisation des flux de travail professionnels, le développement d'add-ons sur-mesure (Google Sheets, Docs, Gmail) et l'intégration de modèles IA (Gemini).
Updated on