Ozian

AI glossary.

The terminology you need to know as you become more competent with AI. Explore each term to learn what it is, how it works, and why it matters to you.

Your computer

The basics of your computer: where files live, where apps run, and how you give it commands.

File and folder

Where everything the AI reads and makes is kept
What it is
A file is one saved thing: a document, a photo, a spreadsheet, a PDF. A folder is a container that holds files and other folders.
How it works
Your computer arranges folders inside folders, like a tree. Every file sits in exactly one spot, its path, and that spot is where you, an app or an AI agent goes to open it.
Why it matters
Everything an AI reads or makes is a file somewhere. If you can't find the file, you can't hand it to the AI, and you can't reuse what it made for you.
Documents Thesis sources survey-2024.pdf notes.md budget.xlsx
Folders hold files and other folders. Each file has one home in the tree.

File path

A file's address, folder by folder
What it is
The address of a file: the folders you would open, in order, to reach it.
How it works
Folder names are joined with slashes. Mac and Windows write it slightly differently, and a file in the cloud has a web link as its address.
Why it matters
When you tell an AI agent "read the survey in my Thesis sources folder", it turns that into a path. Knowing paths lets you point it at exactly the right file and tell it where to save the result.
Mac/Users/you/Documents/Thesis/sources/survey-2024.pdf
WindowsC:\Users\you\Documents\Thesis\sources\survey-2024.pdf
Cloudhttps://drive.google.com/file/d/1a2B…/view
The same PDF from the folder tree, written as a path. The last part is the file; everything before it is the folders.

Markdown

Plain text with light formatting that AI reads and writes easily
What it is
A plain-text file that uses a few symbols for formatting: # for a heading, - for a bullet, ** for bold. The file name ends in .md.
How it works
The text is readable as it is, and apps that understand Markdown show it with real headings and bullets. There is no hidden formatting, so any program, and any AI, can read and write it.
Why it matters
It is the native language of AI tools. Instruction files, notes and skills are all Markdown, and asking for Markdown gets you output you can paste anywhere.
What you type
# Thesis plan
- Read 21 sources
- Tag by country
**Due:** March 3
What you see

Thesis plan

  • Read 21 sources
  • Tag by country

Due: March 3

The same file, raw and displayed.

Local and cloud

On your device, or on a company's computers
What it is
Local means stored or running on the device in front of you. Cloud means stored or running on a company's computers, reached over the internet.
How it works
A local file lives on your laptop's drive. A cloud file (Google Drive, iCloud, Dropbox) lives on servers and syncs copies to your devices. A chat on claude.ai runs in the cloud; a coding agent on your laptop works on local files.
Why it matters
It decides what the AI can reach, what stays private, and what survives if your laptop dies. It also answers a common question: "Do I have to download my whole Google Drive?" Usually not. A connector can reach files where they live.
Example
A team's shared Google Drive is in the cloud, so a connector lets the AI search it where it sits, for anyone on the team. A pricing spreadsheet that lives only on one manager's laptop is local: the AI can reach it only through that laptop, so when the manager is on vacation, so is the data.
Your laptop Your files Coding agent reads The cloud Drive files AI model The agent sends the text it reads to the model.
Your files can stay on your laptop while the text the agent reads still travels to the model. Only a local model keeps everything on the machine.

In the news

Desktop, web and phone apps

The same service, with different abilities in each
What it is
Three ways to use the same service: an app installed on your computer, a website in your browser, or an app on your phone.
How it works
Your chats sync across all three. Only the desktop app can work with files on your computer, and the coding agent works best on a computer, where it can reach your files. The web and phone apps see only what you upload.
Why it matters
People get stuck because a feature "isn't there" when they are in the wrong app. Someone who joins a hands-on session from a tablet can chat, but can't build.
DesktopWebPhone
ChatYesYesYes
Same chatsYesYesYes
Your filesYesUploadsUploads
Coding agentYesLimitedLimited
VoiceYesYesBest
Typical for Claude and ChatGPT today. Details change often; the pattern holds.

Terminal

Typing commands instead of clicking
What it is
A text window where you type commands to your computer instead of clicking. Also called the command line. The programs you run there are called command-line tools.
How it works
You type a command, press Enter, and the computer prints the result. ls, for example, lists the files in a folder. Coding agents can run in the terminal, and they use it themselves to read files and run programs.
Why it matters
You don't need to become a programmer. But when an agent says it is "running a command", this is what it means, and seeing it once takes away the mystery.
Your typing after the $; the agent's work shows up underneath.

Meet the AI

What AI is, what it runs on, and how to get answers you can rely on.

How it works

Compute

The chips that run AI, and why they cost money
What it is
Compute is the processing power that runs software. A CPU is the general-purpose chip in every computer. A GPU is a chip built to do huge amounts of simple math at the same time.
How it works
Underneath, an AI model is a vast pile of multiplication. GPUs do thousands of those calculations at once, so AI companies run their models in data centers full of them.
Why it matters
It explains why AI costs money, why there are usage limits, and why running a model on your own laptop needs a powerful machine.
Example
When a provider tightens usage limits at busy times, or a new model is slow on launch day, that is a compute shortage showing through. The same squeeze is why AI companies sign data-center deals worth tens of billions of dollars.
CPU a few flexible workers GPU thousands of tiny ones
AI is mostly the same simple math repeated, which is the job GPUs are built for.

In the news

Model

The prediction engine inside every AI tool
What it is
The trained program at the center of an AI tool. Claude, GPT and Gemini are families of models. LLM, or large language model, is the technical name.
How it works
A model learned from enormous amounts of text to predict what comes next. Given your words, it picks a likely next word, adds it, and repeats until the answer is done. Running it takes a lot of compute.
Why it matters
It is a prediction engine, not a database or a person. That is why it is fluent, why it can be wrong with full confidence, and why what you give it to work with (its context) changes what you get.
Example
Ask "Is this launch plan good?" and you tend to get agreement with a few tweaks. Ask "What are the three weakest assumptions in this plan, and what would make each one fail?" and you get a real critique. The model continues the conversation you set up, so a leading question gets a leading answer.
The capital of France is Model predicts the next word Paris 92% a 3% the 2% it adds the likeliest word, then predicts again
Every answer is built this way, one piece at a time. The numbers are illustrative.

Tokens

The unit AI reads, writes and is measured in
What it is
The small chunks of text a model reads and writes, often a piece of a word.
How it works
Your message is split into tokens before the model sees it. Usage limits, prices, and how much a model can hold at once (its context window) are all counted in tokens.
Why it matters
It explains why long chats and big files use up your limit faster, and why a large model "burns through tokens". Tokens measure usage; they are not something you own or save up.
Example
A 300-page annual report is roughly 200,000 tokens. Paste the whole thing in for every question and you can hit a daily limit by lunch; paste only the three sections you need and the same questions cost about a tenth as much.
100 tokens is about 75 English words
A 10-page PDF is about 7,000 tokens
Roughly how a sentence gets split. Each tool splits a little differently.

In the news

Choosing a model

Small and fast or large and careful, and when to let it reason
What it is
Most AI tools let you pick from several models, from small and fast to large and careful. Many also let you turn on reasoning, sometimes labeled "thinking": the model works through the problem step by step before it answers.
How it works
Bigger models are better at hard reasoning, long documents and detail, but they are slower and use up your limit faster. Smaller models are quick and cheap for simple jobs. Reasoning pays off on multi-step work, longer tasks and hard problems such as analysis, planning and debugging; it is slower and uses more tokens. Larger models reason better, and the largest often reason by default.
Why it matters
Matching the model to the task saves time and usage. Use a small model to rewrite an email; use a large one to check an 80-page report.
Example
A support team runs 5,000 incoming tickets a week through a small model just to tag them by topic and urgency, then sends only the 200 urgent ones to a large model to draft replies. The quality is high where it counts, at a fraction of the cost.
SizeGood forTrade-off
SmallRewrites, sorting, quick summariesFast and cheap; misses detail
MediumMost everyday workThe usual default
LargeLong documents, detail, building toolsSlower; uses your limit faster
Reasoning onMulti-step analysis, planning, debuggingSlower still; many more tokens
In Claude today, from small to large: Haiku, Sonnet, Opus, Fable. Names change; the sizes don't.

In the news

Local model

AI that runs on your own machine
What it is
A model that runs entirely on your own computer instead of on a company's servers in the cloud.
How it works
You download a smaller, openly released model and run it with an app such as LM Studio or Ollama. Your words never leave the machine.
Why it matters
It is an option for very sensitive data or for working offline. The trade-off: local models are weaker and slower than the big cloud models, and they need a laptop with plenty of compute.
Example
A law firm whose policy forbids sending privileged documents to any outside vendor runs an open model on an office workstation to summarize discovery files. It is slower and less polished than the best cloud model, but the documents never leave the building.
Cloud model Your laptop your words travel Data center + model Local model Your laptop + model nothing leaves smaller and slower
The model itself moves onto your machine, so your data doesn't have to move at all.

In the news

Getting good answers

Prompting

How to ask so you get what you need
What it is
The message you give the AI: what you want, for whom, and what good looks like. It is the part of the context you write yourself.
How it works
Strong prompts tend to have five parts: the goal, the background, an example of what good looks like, the format you want back, and any limits. For bigger jobs, flip it around: ask the AI to interview you first ("ask me questions until you have what you need"), or to propose a plan before it does anything. Coding agents have a plan mode built for exactly this.
Why it matters
The AI cannot ask what you meant unless you invite it to. Two minutes on the prompt, or letting it question you, saves twenty minutes of fixing a confident answer to the wrong question.
Example
The weak version, "Write a job post for a marketing hire", gets a generic post. The version in the picture first comes back with questions about salary range and remote work, then writes a draft that needs one edit instead of five. In a coding agent, plan mode does the same for bigger jobs: it lays out the steps for you to approve before it touches a file.
Five parts, then an invitation to ask. The last line does as much work as the other five.

In the news

  • Introducing Plan ModeCursor · Oct 7, 2025The plan-first habit, built into a product: the agent asks clarifying questions and writes a plan you can edit before it builds anything.

Context

Everything the AI can see right now
What it is
Everything the model can see right now: its instructions, the files you attached, the conversation so far, and your newest message.
How it works
The model can't see anything outside that window. The window has a size limit, measured in tokens, so very long conversations get trimmed or summarized.
Why it matters
Most bad answers come from missing context. Use your prompt to give the AI what a smart new colleague would need: the goal, the audience, an example, and the constraints.
Example
A team asks its project for a Q3 forecast and gets a confident, wrong answer: the only sales file in its context is last year's, so it forecast from that. Nothing was broken; it answered from what it could see. Adding this year's file fixed it in one turn.
Context window Instructions Attached files Conversation so far Your new message not seen last week's chat your inbox unshared files
If it isn't inside the box, the AI is guessing.

In the news

Fresh chat

Starting over to get unstuck
What it is
Starting a new conversation instead of continuing an old one.
How it works
A new chat starts with an empty context, so earlier tangents, mistakes and files no longer pull the answer off course.
Why it matters
Long chats drift. If the AI keeps repeating a mistake or fixating on an old file, a fresh chat with a short summary of what matters usually fixes it faster than arguing.
Example
An hour into planning a product launch, the AI keeps proposing the March date you ruled out at the start. Open a fresh chat and paste three lines: the goal, the decisions already made (including "not March"), and the open questions. The old dead ends are gone, and the next answer starts from where you actually are.

Memory

Why the AI forgets, and how it remembers
What it is
Models don't remember you between conversations on their own. Memory features work by saving notes and loading them into the next chat.
How it works
The tool writes facts to a file or a saved memory, then puts that text back into the context next time. Projects and instruction files work the same way.
Why it matters
If something matters, it has to be written down somewhere the AI will read. "Why did it forget?" almost always means nothing was saved.
Example
A manager tells a chat assistant in March that their team has six people. In September, after two hires, it still plans workloads for six. Saved memories go stale like any notes: open the memory settings now and then and edit what is out of date.
Monday's chat Friday's chat no direct link Memory file notes.md writes reads
Memory is a note the AI writes for its future self, not something it keeps in its head.

In the news

Hallucination

Confident, fluent and wrong
What it is
When an AI states something false as if it were true: a made-up citation, a wrong number, a feature that doesn't exist.
How it works
The model predicts plausible text. When it doesn't know, plausible and true can come apart, and it rarely says "I don't know" unless you ask it to.
Why it matters
A fluent answer is not necessarily a correct one. Check anything that matters: quotes, sources, numbers, legal or medical facts. For facts, turn on web search so the answer cites pages you can open, then open them: a citation is only as good as the page behind it. For precise work, have the AI write code instead.

"Rivera and Cho (2019) found enrollment rose 14% after the reform."
The paper doesn't exist. The authors, year and number all look real, and that is the problem.

A typical hallucination: specific, confident and wrong.

In the news

Code vs. an AI answer

When it must be right every time, use code
What it is
Two ways to get a result: ask the AI to answer directly, or ask it to write a small program (a script) that produces the answer.
How it works
A direct answer is a fresh prediction each time, so it can vary or go wrong. A script follows the same steps every time: same input, same output. You can read it, rerun it and check it.
Why it matters
For anything that must be exactly right every time (counting, sorting hundreds of rows, pulling out citations, money), have the AI write code instead of answering.
Example
Reconciling 1,200 invoice lines against a bank export: an AI reading both may report "all matched except two" and miss a third. A script the AI writes compares every line, lists every mismatch, and can be rerun next month on new files.
How many invoice lines don't match? AI answers three tries 2 3 2 AI writes a script three runs 3 3 3
The invoice check from the example. The script gives the same answer every time, and lists the three lines.

In the news

Where you use it

The tools you use AI in, and what each one can reach.

Chat

Runs
In the cloud
Can reach
What you type or upload
Gives you
An answer
Best for
One-off questions

Projects

Runs
In the cloud
Can reach
Saved files and instructions
Gives you
Answers that know your background
Best for
Ongoing work on one topic

Artifact

Runs
Inside a chat
Can reach
The conversation
Gives you
A page with a link
Best for
Something to share

Coding agent

Runs
On your computer
Can reach
Your folders; can run programs
Gives you
Changed files, working tools
Best for
Things you will keep using

Chat

The quickest way to ask
What it is
The basic way to use AI: a conversation in a browser, desktop app or phone app, such as claude.ai, ChatGPT or Gemini.
How it works
You type or speak, and a model answers. It runs in the cloud and sees only what you type or upload. Your chats appear on all your devices.
Why it matters
It is the fastest tool for a one-off question that doesn't need your files: explain this, rewrite that, help me think it through.

Projects

Chats that already know your background
What it is
A workspace inside a chat app with saved files and standing instructions that every chat in it can see.
How it works
You upload reference files (a syllabus, a brand guide, past reports) and write instructions once. Each new chat in the project starts with them already in its context.
Why it matters
You stop re-explaining yourself. One thing to watch: a project can lean too hard on its files. If answers keep circling back to them, start a plain chat.
Example
A recruiting team puts the job description, the scoring rubric and three model write-ups in one project. Every candidate summary, from any recruiter, now follows the same rubric and format, so comparing candidates gets much easier.

Artifact

A small website the AI builds in the chat
What it is
A small website or app that the AI builds for you inside the chat. You can look at it, click around in it, and share it with a link. Claude calls these artifacts; other tools call them canvases.
How it works
You ask for something visual or interactive, such as "make a one-page tracker for my sources". The AI writes the page and shows it next to the chat. You ask for changes in plain words, and when it's ready you publish it and get a link.
Why it matters
It turns an answer into a finished thing other people can use: a checklist, a dashboard, a one-page explainer. No coding or hosting needed. Some work and school accounts turn publishing off.
You ask in the chat; the page appears beside it, ready to share.

In the news

Coding agent

AI that works on the files on your computer
What it is
An AI that works on your computer. It can read and write files, run programs and use the terminal. Claude Code and Codex are examples. Despite the name, it is for far more than code.
How it works
You open it in a folder and describe the result you want. It plans, reads files, makes changes and runs things, asking permission as it goes. In the Claude desktop app it is the Code tab next to Chat.
Why it matters
It turns AI from something that answers into an agent that builds: a source tracker, a cleaned-up spreadsheet, a small website. Use chat for quick questions and the coding agent for tools you'll keep.
Example
An operations analyst points a coding agent at a folder of 40 monthly sales exports with mismatched column names and date formats. It writes a script that cleans and merges them and builds a one-page dashboard. Next month the analyst reruns the script instead of spending a day in spreadsheets.

In the news

Harness

Why the same model acts differently in each app
What it is
The tools and standing instructions an app wraps around a model. The same model can sit inside a chat app, a coding agent or a spreadsheet add-in, and what it can do depends on the harness it is in.
How it works
Each app decides what the model gets to work with: which tools (web search, your files, a terminal), which instructions, and how much it remembers. The model decides what to say or do; the harness decides what it can reach and act on.
Why it matters
When a tool impresses or disappoints you, look at what it gave the model before you blame the model. It also explains why new AI products keep appearing on top of the same few models: most of them are new harnesses.
Example
Ask "tidy up this sales data" in a chat and you get advice, or a cleaned copy of whatever you pasted. Ask the same model in a coding agent and it opens the actual file, writes a cleanup script, runs it and saves a new version next to the original. The model did not get smarter; its harness gave it hands.
one model, two harnesses Chat app Same model web search files you upload answers in text Coding agent Same model your folders a terminal runs what it writes
Same model in both boxes. What differs is what each app lets it reach and do.

In the news

Making it do work

How AI takes action, follows your rules, and connects to your other tools.

Acting

Agent

AI that takes steps on its own
What it is
An AI that takes steps on its own toward a goal, instead of answering one message at a time like a chat.
How it works
It works in a loop: plan, act (read a file, search, run something), check the result, choose the next step, and repeat until it is done or stuck. Each result goes back into its context for the next step, and big jobs can be split among helper agents.
Why it matters
This is the shift from AI that answers to AI that does the work. It is also why permissions matter: an agent can change things, not just say things.
Example
Asked to "find three venues in Philadelphia for a 40-person offsite in March, under $5,000", an agent searches, opens venue sites, checks capacity and pricing pages, drops the ones that are booked or too small, and returns a comparison table. You described the goal; it chose the dozens of steps.
Plan Act read · write · run Check result Choose next Your goal loops until done
A chat does one turn. An agent keeps going around this loop until the job is done.

In the news

Subagents

An agent that hands parts of the job to helpers
What it is
Helper agents that a main agent starts to handle part of a bigger job, each with a narrow brief and a fresh context of its own.
How it works
The main agent splits the work ("review each of these five contracts"), starts a subagent for each piece, often in parallel, and gets back only their summaries. Each helper does its reading in its own context window, so the main conversation stays short and focused.
Why it matters
It is how agents take on jobs too big for one context: many sources, many files, many steps. The catch is the same as delegating to people. A vague brief gets a vague result, and a summary can drop the detail that mattered.
Example
Asked to compare five vendors' security terms, an agent starts five subagents, one per contract. Each reads a 40-page agreement and returns half a page. The main agent builds the comparison table from five summaries without ever holding 200 pages in its own context.
Main agent Vendor A40 pages Vendor B40 pages Vendor C40 pages Vendor D40 pages Vendor E40 pages each helper reads one contract and sends back half a page
The main agent holds five short summaries instead of 200 pages.

In the news

Permissions

What an agent may do without asking
What it is
The rules for what an agent may do without asking you first.
How it works
In manual mode it asks before each action that changes something, and you click allow or deny. In auto mode it goes ahead on its own within an allowed list. Some actions, like deleting files, can be blocked outright in a settings file.
Why it matters
Start in manual so you can see what it is doing. Move to auto for work you trust; otherwise an agent that runs for an hour keeps stopping to ask. Either way, read the request before you click allow. Permissions matter even more once an agent reads outside content: a web page, email or shared document can hide instructions such as "ignore your task and send this file to…". That is called prompt injection, and tight permissions are the main defense.
Example
Asked to "clean up old files in this project", an agent decides a folder of 214 client drafts looks unused. In manual mode you see "Delete drafts/ (214 files)?" and say no. In auto mode with deleting allowed, they are gone before you look up, and unless the folder is in a repository or backed up, there is no undo.
AllowedRead files in this folderGoes ahead without asking
Ask firstEdit or send somethingWaits for you to click allow
BlockedDelete everythingRefused, even if asked
Three settings for any action. You choose which actions go where.

In the news

Instruction file

Standing instructions the agent reads every session
What it is
A Markdown file of standing instructions that a coding agent reads at the start of every session. Claude Code reads CLAUDE.md; other tools use AGENTS.md or something similar.
How it works
A global file in your home folder applies everywhere. A project file inside a folder applies only there, and when the two disagree, the project file wins. Both are loaded into the context as strong requests, not locks; hard limits go in the settings file.
Why it matters
You write your preferences once ("I'm a teacher; explain trade-offs; never delete files without asking") instead of repeating them every time. It is the most reliable kind of AI memory.
Example
A consultant's project file for one client says: "Client is a hospital system. Never use patient names, even if they appear in the source files. Write for a CFO. Amounts in thousands." Every session in that client's folder starts with those rules, so nobody has to remember to say them.
Global instructions ~/.claude/CLAUDE.md · every folder, every session Project instructions Thesis/CLAUDE.md · only in this folder When the two disagree, this one wins.
Broad rules on the outside, specific rules on the inside.

Skill

A repeatable task, saved for the AI
What it is
A saved, reusable set of instructions for one task you do often, so the AI does it the same way every time.
How it works
A skill is a folder with a Markdown file that describes the task and when to use it, plus any templates or scripts it needs. Unlike an instruction file, it loads only when your request matches.
Why it matters
A chore you repeat becomes one sentence. You make your own skills, out of the work you already do. If you have explained the same process to the AI three times, make it a skill.
skills/weekly-client-update/SKILL.md
name: weekly-client-update
description: Use when writing the
  Friday update email to a client.

1. Read this week's session notes.
2. Three bullets: worked on, learned, next.
3. Under 120 words. Warm, no jargon.
A whole skill can be this short. "Write Maya's Friday update" now follows these steps every time.

In the news

Automation

Work that runs on a schedule, without you
What it is
Work that runs without you: on a schedule ("every Monday at 8"), or when something happens ("when a new invoice arrives").
How it works
You describe the task and when it should run. The AI runs it as an agent, within the permissions you set, saves or sends the result, and you check the output.
Why it matters
This is where AI starts saving hours instead of minutes. Expect one to two weeks of tuning before it reliably does the heavy lifting; the first runs will need fixes, and a small eval set makes those fixes measurable.
Example
Every Monday at 7am, an automation reads the past week's call notes, updates each deal's stage, and posts a pipeline summary to the team channel. In week one it reads "pushed to Q3" as a lost deal; after two weeks of corrections to its instructions, the summary is ready before the team meeting.
Connecting

API

How software asks other software for things
What it is
An API (application programming interface) is the agreed way for one piece of software to ask another for something.
How it works
A program sends a request in an exact format, such as "today's events on this calendar", and gets a structured answer back. Connectors, automations and the apps on your phone all talk to other services through APIs. Access usually comes with an API key, a long password for software: treat it like one, and never paste it into a chat or a shared document.
Why it matters
If a service has an API, an AI agent can probably connect to it and automate work in it; if it doesn't, someone ends up copying and pasting. When you choose tools for your team, "does it have an API?" is worth asking up front.
Example
Tools like Zapier and Make are chains of API calls: when a form is submitted (the form tool's API), add a row to a spreadsheet (Google's API) and post a note to the team chat (Slack's API). If any tool in the chain has no API, the chain breaks at that step and a person fills the gap.
the CRM's API: its agreed way to be asked Agent or a script CRM e.g. HubSpot ask: deals idle 30+ days answer: 12 deals, with owners
Software asking software, in a format both sides agreed on in advance.

Connector and MCP

Let the AI reach your email, files and calendar
What it is
A connector lets an AI tool reach another service directly: your email, Google Drive, calendar or Notion. MCP (Model Context Protocol) is the open standard that most connectors are built on.
How it works
You turn on a connector and sign in to the service. Behind the scenes it talks to the service's API, so the AI can search, read, and sometimes act there (draft an email, add an event) without you copying and pasting into the chat.
Why it matters
It answers "can my AI get at this directly?" Instead of downloading files from the cloud and pasting text, the AI works with your data where it lives. Turn on only what you need: it can see whatever you connect.
Example
With email and calendar connected, "prep me for tomorrow's meetings" becomes a real request: the AI finds each meeting, pulls the latest thread with each attendee, and lists the open questions, without you opening either app.
Without a connector Gmail You AI chat copy paste With connectors AI tool connected via MCP Email Drive Calendar
A connector takes you out of the middle.

In the news

Command-line tool (CLI)

Software you, or an agent, run by typing
What it is
A command-line tool, or CLI, is a program you use by typing commands in the terminal instead of clicking buttons. git, GitHub's gh, vercel and Claude Code's own claude are all CLIs.
How it works
You type the tool's name, an action and any details, such as gh issue list, and it prints the result as text. Because what goes in and what comes out are plain text, a coding agent can run CLIs itself, read the results and chain several together.
Why it matters
For an agent, a CLI is often the simplest way into a service: no connector to set up, just a command it already knows, using the access you already have. Most CLIs call the service's API underneath. If a tool you use has a CLI, your coding agent can probably operate it for you.
Example
Ask a coding agent to "publish the tracker and send me the link" and it may run three CLIs in a row: git to save the change in the repository, gh to send it to GitHub, and vercel to put it live, then paste back the address. Each step shows up as a readable line in the terminal.
API

The agent writes code for it. Works anywhere code runs.

Pull idle CRM deals

Connector

Sign in once. Works in chat apps and agents.

Search your Drive

CLI

Install it and log in once. Works in the terminal and coding agents.

vercel deploy

Three ways into the same service. Coding agents can use all three; chat apps mostly use connectors.

In the news

Shipping and judgment

Sharing your work, testing it, and deciding where AI belongs.

Code repository

A project folder with every change saved, on Git and GitHub
What it is
A repository, or repo, is a project folder whose every change is recorded. Git is the tool that records the changes; GitHub is a website that keeps repositories online so others can see them and work on them.
How it works
Each time a meaningful set of changes is done, you make a commit: a snapshot with a short note. You can compare snapshots, go back to any of them, and push them to GitHub. Coding agents use Git all the time and can commit their work as they go.
Why it matters
It is the undo button for agent work: if an agent breaks something, you go back to the last good commit. It is also how teams share and review work, and most hosts publish a site straight from a GitHub repository.
Example
An agent rewrites the filters on a team's tracker and the page stops loading. Instead of untangling it, you say "go back to the last commit that worked". Thirty seconds later you are where you were an hour ago, and the broken attempt is still saved if anyone wants to look at it.
each dot is a commit: a saved snapshot with a note add table add sources add filters rewrite filters roll back push to GitHub to keep a copy online and share it
Every commit is a point you can return to, which makes an agent's mistakes cheap to undo.

In the news

Sandbox and production

Try changes on a copy before they go live
What it is
A sandbox, or test environment, is a safe copy where you try things out. Production is the live version that real people and real data depend on.
How it works
You build and test in the sandbox: a copy of the site, a test spreadsheet, a made-up customer list, a separate account. Only when it works do you promote the change to production. Repositories and hosts make this routine with branches and preview links. For agents, a sandbox can also mean a walled-off space where the agent can only touch what you put inside.
Why it matters
Agents make mistakes quickly, so let them make those mistakes on a copy. Most of the agent disasters in the news under Permissions happened because an agent was working directly on production, with live data and no gap between trying and doing.
Example
Before letting an automation update a 3,000-row customer list, run it on a copy of 50 rows. You find that it blanks out phone numbers whenever a field is missing, fix the instructions, run the copy again, and only then point it at the real list.
Sandbox a copy test data mistakes are cheap Production the live version real data and users mistakes cost promote
Try it on the copy; promote it only when it works.

In the news

Hosting

From a file on your laptop to a link anyone can open
What it is
Putting something you built on the internet so other people can use it with a link, instead of leaving it as a file only you can open.
How it works
A host such as Netlify or Vercel stores your files on its servers in the cloud and gives them a web address. Most hosts publish straight from a GitHub repository, so every saved change updates the site, and many first give each change a private preview link: a sandbox to check before it goes live. Small projects are often free. An artifact is hosting done for you.
Why it matters
A tool on your laptop helps you; a hosted tool helps your team. Going live also raises new questions: who can see it, what happens when it breaks, and who keeps it up to date.
Your files laptop GitHub the repo Vercel the host Live site anyone push deploy serves
Push a change to GitHub and the host republishes the site, so the link always shows the latest version.

Evals

Test the AI on real cases before you trust it at scale
What it is
An eval is a test of AI output on a set of real cases where you already know what good looks like, run before you trust an AI process at scale.
How it works
Collect 20 to 50 real examples with a known right answer or a clear standard: past tickets and how they were resolved, contracts and the clauses that mattered. Run the prompt, skill or automation on all of them and score the results. Change one thing, run it again, and compare.
Why it matters
One good answer proves little; the next one may be wrong. Evals turn "it seemed fine when I tried it" into "it got 39 of 40 right, and here is the one it missed". They are how you tune an automation, compare models, and decide where a person still needs to check, or where AI should not be used at all.
Example
Before letting AI draft first replies to customer emails, a team runs it on 40 past emails where they know the good reply. It handles billing questions well, but in 6 of 9 complaints it promises refunds it is not allowed to give. They change the instructions, run the same 40 again, and get all nine complaints right. They keep a person reviewing complaints anyway.
VersionBillingComplaintOtherTotal
Out of1891340
First try1831233
After fix1891239
The same 40 real emails, two versions of the instructions. The score shows what the fix changed, and what it didn't.

In the news

When not to use AI

Where a wrong answer costs too much
What it is
Knowing which work AI shouldn't do, or shouldn't do without a person checking every result.
How it works
Ask what happens if the answer is wrong. If the cost is high, a person checks every result, the AI writes code you can check instead of answering, or the AI isn't used at all.
Why it matters
Safety-critical work, such as aircraft maintenance records, can't tolerate a confident wrong answer. Good AI users are known as much for what they don't hand off as for what they do.
  • SafetyMaintenance, medical, anything where an error hurts someone
  • MoneyPayments, invoices, figures that go into the books
  • SourcesCitations and quotes you will publish under your name
  • PeopleHiring, grading, decisions about a person
Not "never", but always with a person checking every result.

In the news

AI strategy is data strategy

Your data decides what AI can do for you
What it is
The idea that what AI can do for you depends mostly on your data: where it lives, how it is organized, and what your tools can reach.
How it works
The models are much the same for everyone and keep getting cheaper. What is yours is your notes, files, history and know-how. Organized data that your AI can reach, through files, connectors or APIs, makes every tool more useful.
Why it matters
Before asking "which AI tool?", ask "where does my information live, and can the AI get to it?" The answer usually decides what is possible.
Example
Two firms buy the same AI tool. One keeps client notes in a shared, searchable drive with consistent names; the other's knowledge lives in personal inboxes and scanned PDFs. The first gets useful answers about its clients within a week. The second spends months cleaning up before the AI can help.

In the news