Prompt glossary: the words you need to know

A colleague tells you: "Your problem is the context window, you need a bit of few-shot and a lower temperature." You nod, open a tab, and land on three definitions that contradict each other. The vocabulary of AI is a mix of technical terms, borrowed jargon, and everyday words that take on a particular meaning, and it discourages more people than the tools themselves.

This glossary gathers thirty-five words you actually run into, arranged by theme rather than alphabetically, because they shed light on one another. Each definition fits in a few lines, with an example when it helps, and points to the article that goes further. Use your browser's search if you're looking for a specific term.

A prompt, word by word

Before the vocabulary, here is a prompt in which every part has a name. The terms in bold are defined below.

You are a management advisor for small nonprofits.                   (role)

The pétanque club of [CITY] has 60 members, dues of €40.             (context)
Here is the statement for the year: [PASTE THE STATEMENT].           (supplied document)

Write a summary for the annual meeting.                              (task)

Answer in 10 lines maximum, with three sections:
revenue, expenses, outlook.                                          (output format)

Example of tone: "We ended the year with €312 more than we started with."  (example)
Do not cite any figure that is not in the statement.                 (constraint)

This breakdown matches what the anatomy of a good prompt details.

The basics

Prompt. The message you send to the AI: a question, an instruction, a text to process, or all of these together. For image generators, it's the description of the image. Also called a query or an input. The beginner's guide covers it.

Language model. The program that produces the answers. It was trained on enormous amounts of text and writes its answer piece by piece, calculating what is most likely to come next. The abbreviation is LLM, for large language model. How it works, without jargon.

Generative AI. A family of tools that produce new content (text, image, sound, code) instead of sorting or searching for what already exists. ChatGPT, Claude, Gemini, Le Chat, and Midjourney are part of it.

Assistant, chatbot. The application in which you talk to a model. The same model can sit behind several applications, with different settings and features.

Conversation. The series of exchanges within which the model sees everything that came before. A new conversation starts from zero, unless the tool offers a memory, whose behavior varies from tool to tool.

Training. The phase during which the model learns from texts before being made available to you. What happened after it ended is unknown to it, unless it has access to a search.

Tokens. The pieces of words that the model reads and writes: a short word counts as one, a long word may count as several. Tools' limits and prices are often measured in this unit. See tokens and the context window.

Context window. The amount of text, in tokens, that the model can keep in view at once: your request, the conversation, the attached documents, and its own answer. Beyond that, it loses elements. For big documents, read working with long documents.

What you write in a prompt

Role. The instruction that gives the AI a point of view: "you are a management advisor." Useful for setting a language level and an angle, without making the model more competent than it is. See giving the AI a role.

Context. The information about your situation that the model can't guess: who you are, who you're writing for, what has been decided. You sort it with the test in this article.

Instruction. What you ask it to do. The more precise and verifiable it is, the better it is followed.

Constraint. A limit placed on the answer: length, banned words, scope ("only from the text provided").

Variable. A placeholder in brackets, like CITY], that you replace each time you use the prompt. A variable turns a prompt into a reusable template; see [your prompt library.

Output format. The form you want for the answer: list, table, three paragraphs. It can be imposed in one sentence, as in the article on format.

Structured output, JSON. An answer in a format a program can read, like JSON, with fixed fields. Useful when the result feeds a spreadsheet or an application. See JSON output.

System prompt, system instructions. A block of instructions placed before the conversation, which sets the AI's general behavior (tone, rules, allowed topics). Depending on the tool, it sometimes takes the form of custom settings. See system instructions.

Negative prompt. In image generators that allow it, the list of what you don't want to appear (blur, text, distorted hands). See the negative prompt.

Techniques

Zero-shot, one-shot, few-shot. The number of examples supplied in the prompt: none, one, a few. Examples fix the format and tone better than adjectives. Few-shot prompting details how to choose them.

Step-by-step reasoning. Asking the model to lay out its reasoning before concluding, which helps with multi-step problems. Often called chain of thought. Depending on the tools, some models do it on their own initiative. See reasoning step by step.

Prompt chain. Splitting a task into several successive prompts, the result of one becoming the input of the next. More reliable than a single big prompt. See chaining prompts.

Iteration. Improving a result through small successive corrections. The method is in improving a prompt through iteration.

Meta-prompting. Asking the AI to write or improve a prompt for you. Handy for getting started, and worth reviewing before use. See meta-prompting.

Prompt injection. Instructions hidden in a text the AI reads (web page, email, document) that it may take for orders. This risk mostly concerns tools connected to your files or to the internet.

Settings

Temperature. A setting that controls the amount of randomness in choosing words. Low, the answers are more predictable; high, more varied. Not all tools let you change it. See temperature and parameters.

Top-p. Another randomness setting: the model chooses only among the most probable words, up to a cumulative threshold. It's mostly set in developer tools.

Maximum length. A cap on the number of tokens in the answer. An answer cut off midway sometimes comes from this.

What goes wrong

Hallucination. A false answer, delivered with confidence: an invented source, figure, function, or quotation. It follows from how the model works, even with a flawless prompt. See reducing hallucinations and checking an answer.

Bias. A systematic tendency in the answers: stereotypes, majority viewpoint, a country's habits. The swap test helps you spot them.

Sycophancy. A tendency to agree with you or to find your text good, even when it isn't. Counter it by explicitly asking for criticism.

Conversation drift. After many exchanges, the model may mix old instructions with new ones. In that case, start from a fresh conversation.

Tools and their words

Multimodal. A model that handles several types of content: text, image, sometimes sound or video. You can show it a photo and ask a question about it.

RAG. Short for retrieval-augmented generation: the tool first searches for passages in a document base, then gives them to the model along with your question. That's how some assistants answer from your files or from the web.

Agent. A program that lets a model chain actions (search, open a file, write) to reach a goal, instead of answering once. Instructions for agents have their own rules.

Fine-tuning. Retraining an existing model on chosen examples to specialize it. Developer work, which has nothing to do with writing a prompt.

Memory. A feature of some tools that keeps information from one conversation to the next. Its existence, settings, and duration vary; check them in the tool, especially before entrusting it with personal data. See privacy.

Putting it to use

Don't try to memorize everything. Keep the eight words that come up in your exchanges: prompt, context, format, example, iteration, hallucination, token, context window. The day someone drops a term you don't know, come back here or, better, ask the AI to explain it with an example from your line of work. If you'd rather start with practice, the prompt catalog shows an excerpt and a real result before any purchase.