The AI glossaryfor creators and founders

Use this AI glossary to understand the words behind every AI tool you open. No jargon, no fluff. Just tap a term to get a clear definition and get back to work.71 AI terms explained, in the order you meet them rather than A to Z.

~/ai-os-glossary
$ glossary define
71 terms · 10 sections

The basics

14 terms

AI

AI is a label that moves. It points at whatever computers have just learned to do well, and right now that means large language models.

Large language model (LLM)

A large language model is a program trained on enormous amounts of text that predicts what word comes next.

Model

A model is the trained thing itself: a very large set of numbers that turns your input into its output.

Token

A token is a chunk of text, roughly three quarters of a word, and it is the unit everything gets counted in.

Prompt

A prompt is what you send. The instruction, the question, the context, all of it together.

Inference

Inference is the model running: reading your input and generating a response.

Training

Training is where a model learns, once, by reading enormous amounts of text and adjusting itself to predict better.

Parameters

Parameters are the numbers inside a model that got set during training. Everything it knows lives in them.

Next-token prediction

Next-token prediction is the only thing a model does: pick the next chunk of text, add it, and go again.

Input tokens

Input tokens are everything you send: your message, your files, your rules, and the whole conversation so far.

Output tokens

Output tokens are what the model writes back, and they are the expensive half.

Cache tokens

Cache tokens are the part of your input that has already been processed and does not have to be paid for again in full.

Model provider

The model provider is whoever runs the model for you: Anthropic, OpenAI, Google, or your own machine.

Effort

Effort is how much thinking the model does before answering.

How it works

11 terms

Context

Context is everything the model can see when it answers.

Context window

The context window is the limit on how much the model can hold at once.

System prompt

The system prompt is the standing instruction sent ahead of everything else, on every single request.

Harness

The harness is everything wrapped around the model that turns it from a text generator into something that can act.

Session

A session is one run of working with an agent, from opening it to closing it.

Turn

A turn is one message from you plus everything the agent does before it hands back.

Stateless

Stateless means nothing carries forward. The model is stateless: it remembers nothing between requests.

Memory system

A memory system is what you build so an agent knows things next session that it learned in this one.

Compaction

Compaction is when a long session gets summarised so it can keep going in a smaller space.

Clearing

Clearing is ending the session and starting fresh, with nothing carried over.

Stateful

Stateful means it carries something forward. A session is stateful across turns, and an agent can be made stateful across sessions.

What it runs

7 terms

Agent

An agent is a model that has been given tools, instructions and a job, so it can act rather than only answer.

Subagent

A subagent is a helper an agent hands a job to, which works separately and reports back one answer.

Skill

A skill is a written procedure for a job you do the same way every time. Your SOP, in a form the agent follows instead of improvising.

Hook

A hook is something that runs automatically on an event, whether or not the agent decides to.

Orchestrator

An orchestrator is what decides which agent does what, in what order, and what happens with the results.

AGENTS.md

AGENTS.md is a file the agent reads at the start of every session, whatever you are working on.

Progressive disclosure

Progressive disclosure is loading only what is needed right now, and leaving a pointer to the rest.

What it reaches

9 terms

Tool

A tool is something an agent can use: read a file, run a command, search the web, send a request.

Model Context Protocol (MCP)

MCP is a standard way to give an agent tools for a service, so the same connector works in any agent that speaks it.

CLI

A CLI is a command you run in a terminal, and it costs nothing until you run it.

Tool call

A tool call is the agent asking to do something: read this file, run this command, search for this.

Permission

A permission is the check before an agent does something you have not already approved.

Sandbox

A sandbox is a walled-off place an agent runs in, so a mistake cannot reach anything that matters.

Filesystem

The filesystem is the folders and files on your computer, and it is where an agent actually works.

Tool result

A tool result is what comes back after the agent does something: the file contents, the output, or the error.

Computer use

Computer use is an agent operating a screen the way a person does: looking, clicking, typing.

When it goes wrong

10 terms

Hallucination

A hallucination is the model stating something false with complete confidence.

Knowledge cutoff

The knowledge cutoff is the date the model stopped learning. It knows nothing after it.

Prompt injection

Prompt injection is when text the agent reads contains instructions, and it follows them.

Attention budget

Attention budget is the fixed amount of focus spread across everything in the session, however much you add.

Attention degradation

Attention degradation is a session getting worse as it gets longer, because the focus is spread thinner.

Sycophancy

Sycophancy is the model agreeing with you because agreement is what people rewarded, not because you are right.

Parametric knowledge

Parametric knowledge is what the model knows from training, frozen at the moment training stopped.

Non-determinism

Non-determinism means the same question can get a different answer twice.

Contextual knowledge

Contextual knowledge is what the agent can read right now, as opposed to what it half-remembers from training.

Smart zone

The smart zone is the early part of a session, when it is sharp, before it drifts.

Knowledge and retrieval

4 terms

Retrieval augmented generation (RAG)

RAG is looking something up first, then answering from what was found rather than from memory.

Embedding

An embedding turns a piece of text into numbers, so a computer can find other text that means something similar.

Vector database

A vector database stores embeddings so you can search by meaning across a lot of documents, fast.

Grounding

Grounding is tying an answer to a source it can point at, instead of one it produced from memory.

Models and tuning

4 terms

Fine-tuning

Fine-tuning is further training on your own examples, so the model behaves differently by default.

Reasoning model

A reasoning model works through a problem before answering, rather than answering straight away.

Temperature

Temperature is how much the model varies its choices. Low is predictable, high is loose.

Multimodal

Multimodal means it handles more than text: images, audio, video, screenshots.

Limits and cost

2 terms

Rate limit

A rate limit is the ceiling on how much you can use in a window, before you have to wait.

API key

An API key is the password that lets a program use a service on your behalf.

Getting good output

4 terms

Prompt engineering

Prompt engineering is wording the request better to get a better result.

Context engineering

Context engineering is deciding what the agent should already know, and making sure it is loaded before you ask.

Evals

Evals are a set of test cases you run to check whether a change actually made things better.

Guardrails

Guardrails are the rules that stop an agent doing certain things, whatever it has been asked.

How you work with it

6 terms

Human in the loop

Human in the loop means you stay in the session, reviewing and redirecting as it goes.

AFK

AFK is starting the work and walking away, letting the agent run unattended.

Vibe coding

Vibe coding is accepting what the agent produced without reading it.

Workflow

A workflow is a job with steps that runs the same way every time.

Automation

Automation is a workflow that runs without you starting it.

Handoff

A handoff is passing what one session knows to the next one, because nothing carries over on its own.

Was this page helpful?
Previous

API

Next

Graph