A sequence of steps in which the AI works towards an outcome, using skills, connectors and schedules and making judgement calls within limits you set.

A hypothetical form of AI with broad, human-like ability to learn, reason and solve problems across many different tasks.

A common pattern where productivity may dip at first as people learn AI, before improving once new habits and workflows settle in.

A software system that carries out tasks on your behalf, taking several steps by itself within defined rules.

Low-quality AI output that sounds plausible but has not been shaped or checked by someone who knows the situation.

Getting an AI system to do what people actually want, not just what the instructions literally say.

Something Claude makes for other people to use, such as a document, dashboard, slide deck, calculator or quiz, which opens beside the chat and can be shared as a link. ChatGPT’s equivalent is called a Site.

A knowledge mode where the model improves its answer by using external information such as web search, connected apps, tools or databases.

Translating a translation back into the original language so a reader can check whether the meaning has drifted.

Systematic skew in AI outputs caused by patterns in training data, source material, system design or user prompting.

Starting a new chat from an earlier point in an existing conversation without changing the original thread.

A prompt scaffold that asks clarifying questions before turning a rough task into a stronger working brief.

The AI reads a web page and acts on it by clicking, typing and filling in forms inside your own browser.

The gap between what the latest models can do and how people are actually using them.

The AI opens an application on your computer and works through its screen one step at a time.

External services linked to a chatbot so the model can work with information from places such as email, calendars, documents or cloud storage.

A controlled link that lets the AI search and work with information held in another account or service, such as email.

The information the model has to work with in a chat: your instructions, earlier messages, uploaded files and anything else you supply.

Reusable prompt structures that help the model understand the task, build a better brief and produce stronger outputs.

The model’s working memory within a single conversation, including any uploaded files.

The amount of text, data, files and prior conversation the model can actively handle at once.

The part of the Claude desktop app where Claude works on tasks with access to your files and tools, including scheduled tasks and recorded skills.

Standing preferences that shape how ChatGPT responds, such as tone, spelling, formatting and style.

A research mode in which the AI searches many sources over a longer time and returns a structured report.

A saved set of colours, fonts, logos and layout rules that Claude applies to visual materials.

Speaking into the message box so your words are transcribed into text.

How extensively the model works through a task before responding, from a quick answer to a more considered one.

The tendency to mistake a chatbot’s conversational style for real understanding, empathy or human-like awareness.

One of the most capable models currently available, which costs more per token than a lighter model.

A prompt scaffold that identifies what is missing, unclear, assumed or weak before work proceeds.

A knowledge mode where the model answers from its built-in training and the current chat, without live search or supplied source material.

AI that creates a new response in the moment from patterns it has learned, rather than retrieving a stored answer.

A service that stores every version of a website or app’s code, so a change can be reviewed and rolled back.

Documents supplied or nominated as source material so the model’s answer stays anchored to specific evidence.

A knowledge mode where the model’s answer is anchored to specific documents, data or sources you provide.

Basing the AI’s answer on specific material you provide, so it can be checked against that source.

Rules, filters or design controls intended to reduce harmful, unsafe, biased or inappropriate AI outputs.

When an AI presents false or made-up information as if it were true.

Keeping a person responsible for reviewing, judging and owning the final output.

The process of producing an answer or conclusion from patterns, probabilities or supplied information.

The hypothesis that once AI can improve itself, it could rapidly become more capable than humans.

Using adversarial or manipulative prompts to bypass a chatbot’s safety rules or behavioural controls.

The point where a model’s training stopped, so that it does not know anything that happened afterwards.

Anchoring a model’s answer in specific, verifiable source material to improve accuracy and reduce drift.

The underlying AI model that powers chatbots by learning and generating language patterns at scale.

ChatGPT’s file storage area where uploaded documents can be stored, found and reused across work.

Context that carries forward, within a conversation or between conversations, so work can build up over several steps.

Using a prompt to improve, rewrite or generate another prompt.

A Microsoft application layer that brings AI assistance into Microsoft 365 tools, built on top of underlying AI models.

The underlying AI engine that does the work. More capable models handle harder tasks, and lighter ones are faster and cheaper.

Google’s source-grounded AI tool for working with a bounded set of documents and turning them into summaries, briefings, study aids or audio overviews.

How a model learns: it picks up statistical relationships from very large amounts of text without storing the original material.

An add-on that gives the AI an extra tool or ability.

A prompt scaffold that asks the model to assume a task has failed and identify the most likely causes.

Already trained before you use it, so a conversation does not add new facts to the underlying model.

A persistent workspace that keeps instructions, source documents and related conversations together for ongoing work.

An instruction, question or task given to an AI model.

A reusable prompt structure designed to make the model’s work clearer, stronger or more consistent.

A proposed change to a website or app, held for a person to review and approve before it goes live.

An instruction the AI runs automatically at a set time or on a recurring schedule.

Information with some organisation, such as headings or repeated fields, but no fully standard format.

Using AI to analyse information with some organisation but inconsistent formatting, such as policies, product catalogues, quotes or reports.

ChatGPT’s version of an artifact: an interactive website or lightweight app that ChatGPT builds, hosts and shares for you.

A reusable set of instructions for a task you repeat, made up of a name, a description and the steps to follow.

Analysis that stays tied to nominated documents, data or evidence instead of relying on general model knowledge.

Information organised into consistent fields, such as spreadsheets, sales data and pricing tables.

Using AI to analyse information organised into consistent fields, such as spreadsheets, dashboards, lists or financial results.

Different response modes that vary how much effort the model spends reasoning, checking and working through a task.

The small piece of text a model actually processes, and the unit AI usage is metered and priced in.

The material used to teach a model patterns before it is released for use.

The one-off process where a model learns from training data and turns those patterns into model behaviour.

The model architecture that helps an AI system read words in context rather than treating each word in isolation.

Loose information with no fixed format, such as emails, reviews, transcripts and meeting notes.

Using AI to analyse loose information with no fixed format, such as emails, reviews, transcripts, complaints or meeting notes.

Supplying an open, static web link so a chatbot can fetch and read the content.

Using AI to compare two or more versions of a document and identify meaningful changes, additions, removals or risk shifts.

Describing a change in plain language and letting the AI write the code to update a website or app.

A spoken conversation with the AI, where you talk and hear the answer back.

The AI looking things up on the internet during a chat to bring in current information.

Editable sections inside a ChatGPT response that make it easier to draft and revise text without losing the working version.

Language patterns that can suggest AI involvement, especially when several appear together in over-polished or generic text.

The model’s first plausible output before proper human review, judgement and refinement.