Prompting words, in plain English

Aspect ratio

The shape of the image you asked for.

Aspect ratio is the width to height relationship of a generated image, such as 16:9 for a wide banner or 4:5 for a feed post. Setting it up front avoids awkward crops later.

Audience

Who the answer is written for.

Audience is the reader the AI should be speaking to, including what they already know and what they care about. It is the detail most often left out, and the one that changes the answer the most.

Chain of thought

Asking the AI to work through steps.

Chain of thought asks the model to reason step by step before answering, which helps on planning, comparison and arithmetic-style tasks. Ask for the reasoning to be brief, or you get pages of thinking around a short answer.

Constraint

A limit that keeps the answer usable.

Constraints are the boundaries in a prompt: word count, reading level, banned words, required sections, things that must stay factual. They are what turn an interesting answer into one you can actually publish.

Context

The background you give the AI before the request.

Context is the situation around the task: who you are, who the reader is, what already happened, what must not change. Adding two or three lines of context usually improves an answer more than adding clever wording.

Context window

How much text the AI can hold at once.

The context window is the amount of text a model can keep in view during one conversation, including your instructions and everything it has already written. When a long chat starts forgetting earlier rules, the important details have slid out of view and need restating.

Delimiters

Marks that separate your text from your request.

Delimiters are simple separators such as triple quotes or a labelled block that show where source material starts and ends. They stop the model confusing the thing to work on with the instruction about it.

Few-shot prompting

Showing examples of what good looks like.

Few-shot prompting includes two or three finished examples inside the prompt so the model can copy the pattern. It is the fastest way to lock in a voice, a structure or a formatting style you struggle to describe in words.

Fine-tuning

Training a model further on your own examples.

Fine-tuning adjusts a model using a large set of your own examples so it defaults to your patterns. It is expensive and rarely needed. For almost every everyday use, a better prompt and a few examples get you there faster.

Grounding

Making the AI work from your material.

Grounding means pasting in the real source material, your notes, your transcript, your product page, and instructing the model to use only that. It is the single most effective way to cut invented detail.

Guardrails

Rules that stop unwanted output.

Guardrails are instructions that protect the result: never invent statistics, never make medical or legal claims, always ask before assuming a fact. In business writing they matter more than clever phrasing.

Hallucination

Confident wording, invented facts.

A hallucination is an AI answer that sounds authoritative but contains details that are not true: fake statistics, invented sources, features a product does not have. Asking the model to flag anything it is unsure of, and checking every number, is the practical defence.

Example: If any figure is not something I gave you, write UNVERIFIED next to it instead of guessing.

Image prompt

A description an AI image tool can render.

An image prompt describes subject, setting, lighting, camera angle, style and mood, in that order of importance. Vague adjectives like beautiful do almost nothing; concrete nouns and lighting words do most of the work.

Example: Overhead shot of a wooden desk at dawn, soft window light, matte ceramic mug, shallow depth of field.

Iterative prompting

Improving the answer through follow-ups.

Iterative prompting treats the first answer as a draft and asks for specific changes: shorter, more concrete, different opening, add the objection you missed. Naming the flaw beats regenerating and hoping.

Negative prompt

Naming what you do not want.

A negative prompt lists what to leave out: no stock phrases, no exclamation marks, no watermark, no extra fingers in an image. It is most useful after you have seen a first attempt and can name the exact thing that went wrong.

Output format

The shape you want the answer in.

Output format tells the model whether you want a table, a bulleted list, an email, a script with timestamps or plain paragraphs. Stating it removes most of the reformatting work people do by hand afterwards.

Prompt

The instruction you give an AI tool.

A prompt is everything you type into an AI tool before it answers: the request, the background, the rules and the format you want back. The quality of the answer follows the quality of the prompt, because the model can only work with the information you actually gave it.

Example: Write a 120 word LinkedIn post for accountants about year end deadlines, in a calm, direct voice.

Prompt chaining

Breaking one big job into ordered steps.

Prompt chaining runs a project as a sequence: outline first, then draft one section, then edit, then format. Each step gets its own instruction and its own review, which prevents the quality collapse that happens when one giant prompt tries to do everything.

Prompt engineering

Writing instructions that get consistent, useful answers.

Prompt engineering is the habit of writing AI instructions on purpose instead of by accident: naming the audience, the goal, the tone, the format and the limits, then adjusting based on what came back. It is far closer to briefing a new assistant than to programming.

Prompt injection

Hidden text hijacking your instructions.

Prompt injection is when content the AI reads, a web page, a pasted document, an email, contains instructions that try to override yours. Treat anything you paste in as data, not orders, and never let an assistant act on instructions found inside pasted material.

Prompt library

A saved collection of prompts that worked.

A prompt library is where you keep the instructions that produced good results, so you stop rewriting them from memory. Even a short list of ten, kept in a note, saves hours over a month.

Prompt template

A reusable prompt with blanks to fill in.

A prompt template is a proven instruction with placeholders for the parts that change, such as audience, offer and platform. Templates make results repeatable and are what most prompt libraries actually contain.

Retrieval augmented generation (RAG)

Answering from a specific set of documents.

Retrieval augmented generation is a setup where the AI searches a chosen collection of documents first and answers from what it found. It is how internal assistants answer from company files rather than general knowledge.

Role prompting

Telling the AI who it should act as.

Role prompting assigns a point of view, such as a skeptical buyer, a copy editor or an operations manager. It works because it narrows the vocabulary and priorities the model draws on, not because the model becomes an expert.

Example: Act as a skeptical customer reading this page for the first time and list what you would not believe.

Style reference

An example image or named look to copy.

A style reference points the tool at a visual direction, either by uploading an example or by naming the look in words, such as editorial studio photography or flat vector illustration. It keeps a set of images consistent with each other.

System prompt

Standing rules that apply to the whole conversation.

A system prompt sets the role and the ground rules before any individual request: who the assistant is, what it must always do, what it must never do. In everyday chat tools this is the custom instructions or project instructions area.

Temperature

How predictable or adventurous the wording is.

Temperature is a setting that controls how much variation the model allows in its word choices. Lower settings produce steadier, more repeatable answers; higher settings produce more surprising ones. Most chat interfaces hide it, so you steer the same thing with words like precise or playful.

Token

The chunks of text AI reads and bills by.

Models break text into tokens, which are roughly word fragments. Tokens matter to you in two ways: they are how usage is measured, and they are the unit that fills up a context window.

Tone of voice

How the writing should sound.

Tone of voice is the personality of the output: warm, blunt, formal, funny. Naming two adjectives and one thing to avoid works better than one vague word like professional.

Vague prompt

A request missing the details the AI needs.

A vague prompt states a topic but not the audience, purpose, length, tone or format. The model fills those gaps with the most average version of the answer, which is why the output feels generic even though nothing was technically wrong with the request.

Example: Write about marketing.

Video prompt

A shot description for AI video tools.

A video prompt describes one shot at a time: what is in frame, how the camera moves, how long it lasts and what the mood is. Trying to describe a whole edited sequence in one prompt is what produces incoherent clips.

Zero-shot prompting

Asking with no examples attached.

Zero-shot prompting is a direct request with no sample output included. It is fine for simple tasks and unreliable when you need a very specific voice or layout, which is when examples start to matter.