TaskbyTask.Practical AI guides for work
Reference · Plain-English glossary

AI terms in plain English

What LLM, model, prompt, token, context window, RAG, agent, MCP, skill and API mean, and why each matters in everyday work.

Task by Task12 min read
A tabbed reference index, open card and pencil illustrate looking up an unfamiliar AI term.

You can use AI without collecting a second vocabulary. Look up the term you need, then get back to the task. Product names and feature labels vary, and vendors sometimes use the same word differently, so each “At work” line tells you what to check in the tool you’re actually using.

A–C

Agent

An AI system that can work towards a goal over several steps, often deciding which tools to use and how to continue. The word is used loosely, so check what the product really does.

At work: define the goal, permissions and stopping point. A tool that can send messages needs more oversight than one that only drafts them. Anthropic’s distinction between workflows and agents.

Guide: Chat, Cowork and agents: hand Copilot a task with clear boundaries

AI

Artificial intelligence is a broad name for computer systems that recognise patterns, make predictions or generate content. It describes a family of technologies, not one product.

At work: ask what a particular system can do, what information it uses and how you will check it.

Guide: Start here with AI at work

API

Application programming interface. A way for software to ask another system to do something or return information.

At work: an application might send a document to a model through an API. Ordinary chat doesn’t need one, and chat subscriptions and developer API access are often billed separately. Anthropic’s product distinction.

API key

A credential that lets software use a service. Depending on its permissions, anyone holding it may be able to use the service or run up charges.

At work: keep it in your organisation’s secret storage. Never paste it into a prompt, a shared document or a screenshot.

Artifact

In Claude, a piece of work it creates that you would show someone else — a document, deck, dashboard or small interactive tool — which opens beside the conversation and can be edited, saved and shared.

At work: check who can open it before you share it. Viewers can also reach files from the conversation that created it. What are artifacts?

Guide: Share a Claude Artifact and check who can use it

Automation

A process that runs when a trigger or schedule is met. It may use fixed rules, AI or a mixture of both.

At work: test it on a small, reversible task. Decide who reviews failures and how to stop it before you make it recurring.

Guide: Build a small Zapier AI workflow with a failure rule

Citation

A reference telling you where a claim supposedly came from.

At work: open it and check that it supports the statement. A link existing is not proof that the source says what the answer claims.

Guide: Ask a notebook question and check the cited passage

Coding agent

An agent built for software work, such as OpenAI’s Codex or Claude Code. It can read a project, edit files and run commands, rather than only explaining code in a chat.

At work: decide which folder it may change and which commands it may run, and read the changes before anything is deployed. Claude Code overview.

Guide: The beginner’s guide to Codex for business users

Connector or integration

A link between the AI app and another service, such as cloud storage, email or a calendar. Products use different names, including connectors, apps and plugins.

At work: check the permissions and which account is being connected. A connection can expose far more information than a single uploaded document.

Guide: Check what an AI connector can read and change

Context

The information available to the model for the current response: instructions, conversation history, documents you’ve added and results from tools.

At work: important facts need to be in the right place. Don’t assume the model has seen a document because it exists somewhere in your business.

Guide: Know when to start a fresh chat

Context window

The limit on how much information a model can work with in one request, measured in tokens. Products may stretch longer work with summaries or retrieval.

At work: a large context window doesn’t guarantee that every relevant detail is noticed or used correctly.

Guide: Check an answer from a long PDF with Acrobat AI Assistant

Custom GPT

A version of ChatGPT set up with its own instructions, files and connected actions for one job. OpenAI has scheduled custom GPTs to retire on 11 December 2026 and is moving them to plugins.

At work: if you rely on someone else’s GPT, check its creator’s notice for a replacement. Access to the old GPT doesn’t carry over automatically. Custom GPT retirement FAQ.

Custom instructions or system prompt

Standing instructions applied to every conversation, or every conversation in a project, before your own message. A system prompt is the developer’s version of the same idea.

At work: keep them short and general. A rule that suits one task can quietly distort every other answer.

Guide: Give ChatGPT helpful instructions without overloading every answer

Back to the term list

D–L

Deep research

A mode in which the AI plans a research task, reads many sources over several minutes and returns a longer report with citations. ChatGPT, Gemini and Perplexity offer versions of it.

At work: citations make checking possible, not unnecessary. Check that each source supports the claim attached to it. Deep research in ChatGPT.

Guide: Check a research report before relying on it

Evaluation

A planned way to judge whether a tool or workflow meets your requirements, using examples and clear checks.

At work: test whether the tool keeps dates, flags missing information and produces a usable format. One attractive answer doesn’t establish reliability.

Guide: Compare Claude and ChatGPT on your own work

Example output

A sample answer, image or file shown to illustrate what a tool or method produces. It may have been written by hand rather than produced by the tool.

At work: if a guide or vendor shows you an example answer, ask whether the tool really produced it, with which settings, and on what input.

Fine-tuning

Further training an existing model on your own examples so it behaves in a particular way. It is a developer task with its own cost, and it is different from giving a model a document to read.

At work: most office jobs need better instructions or better sources, not fine-tuning. If someone proposes it, ask what problem a project or a skill couldn’t solve. OpenAI fine-tuning guide.

Generative AI

AI that produces new content, such as text, images, audio or code, from patterns learned during training and the information given to it.

At work: a fluent new paragraph may still contain a wrong fact or a promise nobody agreed to.

Grounding

Basing an answer on relevant evidence given to or retrieved for the task.

At work: ask the tool to use the current policy and name the supporting section. Grounding helps; the answer still needs checking.

Guide: Use business files and check the answer

Hallucination

A plausible-looking but incorrect or unsupported output, including made-up facts, quotes, citations or explanations.

At work: a calm tone and an exact-looking number don’t prove accuracy. Check important claims against the original source. Reliability limitations.

Guide: Choose the checks an AI task needs

Inference

In AI, inference means using a trained model to produce a response from new input. Training prepares the model; inference is the model being used. Google’s technical definition.

The word also has its everyday meaning: a conclusion drawn from evidence rather than stated in it.

At work: an “inference cost” usually means running the model. “The shipment was delayed” may be recorded; “the supplier is unreliable” is an inference that needs more evidence.

LLM

A large language model is trained on large amounts of data to work with language. It generates responses from learned patterns and the context available to it. Many can also work with images or other inputs.

At work: it can draft and organise material, but it doesn’t automatically know your current prices, policies or agreements. Background on model development.

Back to the term list

M–P

MCP

Model Context Protocol: an open standard for connecting AI apps such as Claude and ChatGPT to data, tools and workflows in other systems. Many connectors are built on it.

At work: an MCP connection is still a connection. Ask what it can read, what it can change and whose account it uses. What is MCP?

Guide: Check what an AI connector can read and change

Memory

A feature that carries selected information or preferences between conversations. It may work differently inside and outside projects.

At work: check what is remembered and where it is used. Memory settings and model-training settings control different things. ChatGPT data controls · Claude chat search and memory.

Guide: Review assistant memory and choose what to keep

MFA

Multi-factor authentication. An extra verification step when signing in, beyond a single password.

At work: turn on the methods your organisation supports and plan how you would recover access. OpenAI MFA · Microsoft Entra MFA.

Model

The trained system doing the processing. An app can offer several models with different capabilities, speed, costs and limits.

At work: ChatGPT is an app; the model selected inside it is one part of the service. Changing the model doesn’t change your account’s data policy.

Model training

Adjusting a model using data. This is different from giving an existing model a document as context for one task.

At work: check the provider’s policy for your exact account type and settings. Don’t infer the answer from whether you pay for the app.

Guide: Work out what you are allowed to put into an AI tool at work

Multimodal

Able to work with more than one kind of information, such as text and images. Which inputs and outputs are supported depends on the model and product.

At work: a tool that accepts a PDF may still handle its text differently from its charts. Check the feature rather than assuming “multimodal” means it sees everything.

Output

What the system produces: an answer, image, document, code or other result.

At work: an output is a draft until the relevant checks and sign-offs are done.

Plugin

A reusable package for a workflow. In ChatGPT and Codex a plugin can bundle skills (instructions) and apps (connections to other services), and it is replacing the custom GPT. Claude also uses the word for bundles of skills and connectors.

At work: a plugin can add both instructions and access. Check what each part can read or change before you install it. Moving custom GPT workflows to plugins.

Guide: Skills, plugins and connectors in Codex: know what adds access

Project

A feature for keeping related chats, reference files and instructions together. Memory and sharing behaviour vary by service and settings.

At work: a project helps with recurring tasks. It doesn’t give a personal account business-plan protections. ChatGPT projects · Claude projects.

Guide: Use Claude Projects so you stop re-explaining the same context every Monday

Prompt

The request and material you give the AI to guide its answer: instructions, source text, examples, images or other inputs.

At work: “Turn these checked notes into six numbered steps” is a prompt. A good one explains the task, source, constraints and intended output. Prompting basics.

Guide: Write your first useful prompt

Prompt engineering

Designing and improving prompts to get more useful, consistent results.

At work: give a clear brief, read the answer and change the instruction that caused the problem. You don’t need a grand title for doing this carefully.

Prompt injection

Instructions hidden in material the AI reads, such as a webpage or document, that try to redirect what it does.

At work: treat source content as evidence, not orders. Avoid granting action-taking permissions you don’t need, and don’t assume a warning in your prompt blocks every attack. Prompt-injection guidance.

Guide: Recognise instructions hidden inside a document or webpage

Back to the term list

R–W

RAG

Retrieval-augmented generation. The system finds relevant information and gives it to the model as context for an answer.

At work: a policy assistant might retrieve handbook passages before answering. Missing or outdated sources still produce a poor answer, and RAG doesn’t mean the model was retrained on your files. Anthropic’s explanation.

Reasoning model

A model that works through a problem in intermediate steps before answering, usually taking longer. Products label this “thinking” or offer it as a separate model choice.

At work: use it for multi-step problems such as reconciling figures or weighing options. Longer thinking is not a citation; check the result the same way.

Retention

How long a service keeps information, subject to its policies, settings and legal exceptions.

At work: “not used for training” and “not stored” are different claims. Check both, alongside access and sharing.

Retrieval

Finding relevant material in a source such as a document collection, search index or connected system.

At work: the system may retrieve selected passages rather than read your entire library for every answer.

Skill

A saved set of instructions, and sometimes scripts and reference files, that teaches the AI how to do one kind of task. Claude loads a skill only when the task needs it; in ChatGPT and Codex, skills travel inside plugins.

At work: a skill changes how the AI works, not what it can reach. Access comes from connectors and apps. What are skills?

Guide: Save a useful AI task so you can reuse it

SSO

Single sign-on. Signing in to a service through your organisation’s identity system.

At work: use the route your administrator gives you. An existing personal account and a company-managed sign-in are not the same arrangement.

Token

A unit a model uses to process information. For text, it might be a word, part of a word or a punctuation mark. Token counts are not word counts and vary by model and language.

At work: tokens affect context limits and often API costs. You rarely need to count them for an ordinary chat. Tokens and limits.

Tool

A capability the AI can call, such as a calculator, search service or file editor.

At work: some tools only read; others can change things. Ask what the tool may do before granting access.

Workspace

An area in an AI service used by a person or an organisation. A managed organisation workspace may have central membership, billing and admin policies.

At work: confirm which workspace is active before adding company material. Its name alone tells you nothing about what data belongs there.

Guide: Set up your first AI workspace

Back to the term list

Keep this: when a term affects your data, cost or permissions, check what it means in the product you are actually using. For a practical starting point, follow the beginner learning path.

Take the terms back to a useful task

Use the printable workplace AI map to connect the concepts with an action and a check. The free task-kit library supplies reusable instructions and practice sources; no signup or installation is required.

Optional analytics

With your permission, Google Analytics uses cookies to measure visits and which guides people read. It stays off until you accept. Use Analytics choices in the footer to change your choice. Rejecting after accepting refreshes this page. Privacy details.