AI CoursePilot

📖 Glossary

Every bit of jargon from the course, decoded in plain English. Tap a dotted term in any lesson and it lands right here.

A

account management

The client-facing team who own the relationship, the budget and the timeline.

Account management (account exec, account manager, account director) is the agency's client-facing spine. They own the relationship, take the brief, agree budgets and deadlines, keep the client informed and keep the project moving internally. They're the people a resourcer negotiates with most, because they're the ones promising clients things and then needing the people to deliver them.

The person holding both ends of the rope: the client pulling one way, the creative team the other.

agentic

Describes AI that works like an agent — multi-step, tool-using, semi-autonomous — rather than single question-and-answer.

'Agentic' describes AI behaviour that goes beyond one-shot answers: breaking a goal into steps, using tools, reacting to results, and carrying on with minimal hand-holding. Claude Code building and testing a website is agentic; asking a chatbot to explain a word is not.

Describes the way of working, not the person. Handing you a printed timetable is not agentic; booking, checking, rebooking and chasing until you're actually in Tokyo is.

AGENTS.md

The open, cross-tool instructions file for AI coding agents — read natively by Codex, Cursor, Gemini CLI and dozens more.

AGENTS.md is an open standard for agent instruction files — the same idea as CLAUDE.md, but shared across tools. Launched by OpenAI and partners in August 2025 and handed to the Linux Foundation later that year, it's now read natively by Codex, Cursor, Gemini CLI and many others. Think of it as the universal induction pack: write one README-for-robots and most agents can use it. Claude Code is the famous holdout — it needs a one-line @AGENTS.md import or a symlink.

The laminated briefing card in the cockpit that any airline's crew can pick up and read. AGENTS.md is that card for Claude, Codex or Hermes alike — one sheet, whoever's flying today.

agentskills.io

The open, cross-tool standard for skills — a SKILL.md that Claude, Codex or Hermes agents can all load, so a skill written once travels between tools.

agentskills.io is an open standard that defines what a 'skill' looks like — a folder with a SKILL.md of instructions — so the same skill works across different AI tools instead of being locked to one. Claude Code, Codex and Hermes all speak it. It's to procedures what MCP is to tools and AGENTS.md is to project instructions: write once, use everywhere. The shared shop your agents browse.

Like the cabin-safety drill every airline teaches to the same standard. Written once, and any crew — Claude's, Codex's, Hermes's — can perform it correctly.

AGI

Artificial General Intelligence — hypothetical AI as capable as humans across most tasks. Definition and timeline hotly debated.

AGI (Artificial General Intelligence) is the term for AI that matches or exceeds human ability across most intellectual work, not just narrow tasks. There's no agreed definition or test, and predictions range from years to decades. In interviews it's a useful term to handle carefully: today's systems are powerful but narrow-ish tools, and claims about AGI are claims about the future, not the present.

One pilot licensed to fly every aircraft ever built, run the control tower, and rebuild the engines. Nobody holds that licence, and it isn't obvious anyone ever will.

AI

Artificial Intelligence — software that does things we'd normally say need human smarts: understanding language, recognising images, making decisions.

Artificial Intelligence is the umbrella term for software that performs tasks we associate with human intelligence — understanding language, recognising pictures, planning, deciding. Think of it as the whole toolbox; machine learning, LLMs and chatbots are tools inside it. Modern 'AI' in the news almost always means systems learned from data rather than programmed rule-by-rule.

Anything at the airport that looks like it needs a person's judgement but doesn't — the gate that recognises your face, the system that works out your bag is on the wrong belt.

AI agent

AI that doesn't just answer — it acts: takes a goal, makes a plan, uses tools, checks results, repeats.

An agent is AI in a loop with tools: give it a goal and it plans, acts (runs a command, calls an API, edits a file), looks at the result, and decides the next step — repeating until done. That's the difference between asking a chatbot 'how do I build a website?' and telling an agent 'build me a website' and watching it happen.

A travel agent you hand “get me to Tokyo by Thursday, under £900”. They book the flights, sort the visa, rebook when a leg is cancelled, and come back when it's done. A chatbot would just show you the timetable.

AI influencer

A wholly AI-generated social-media persona — consistent face, posts, sponsorships — run by a team or a single creator.

An AI influencer is a social-media persona whose face, photos and videos are entirely AI-generated — kept visually consistent across posts with image models, animated and voiced with video and speech models. Some earn real sponsorship money. They're a window into both the creative power of image/video AI and the disclosure-and-authenticity ethics that come with it.

The face of the airline on every billboard, who has never worked a flight, because there is no such person. Real posts, real sponsorship deals, no human.

AI slop

Generic, low-effort AI output produced at scale — technically fluent, saying nothing. Shortlisted for Oxford's 2024 word of the year.

AI slop is the mass-produced, low-quality content flooding feeds and search results: interchangeable posts, padded articles, samey images — fluent but empty. Slop is usually a prompting failure rather than a model failure: an unspecific request returns the internet's statistical average. The cure is specificity (voice, audience, examples, banned words) and writing that taste down in files like CLAUDE.md or a skill so it's applied every time — plus a human edit before anything ships.

The airline newsletter that reads exactly like every other airline's newsletter, because all of them asked the same model for 'a newsletter' and nobody said what their airline sounds like.

AI winter

A period when AI hype outran results, so funding and interest collapsed. It happened twice: mid-1970s and late 1980s.

An AI winter is a crash of confidence in AI research: grand promises fail to materialise, governments and companies pull funding, and the field goes quiet for years. It happened roughly 1974–1980 (after damning reports like the UK's Lighthill Report) and again 1987–1993 (when the expert-systems boom went bust). Think of it as a stock-market crash for research money — and the reason seasoned AI folk are allergic to overpromising.

The decades after Concorde was retired: supersonic travel was obviously the future, then the money left and everyone flew subsonic anyway. AI has had two of those.

algorithm

A step-by-step recipe a computer follows. Traditional software = hand-written algorithms; machine learning learns its own.

An algorithm is a precise step-by-step procedure for solving a problem — a recipe a computer can follow exactly. Traditional software is built from algorithms humans wrote by hand. In machine learning the twist is that a learning algorithm produces the final behaviour itself, by adjusting to data. When people say 'the algorithm' about TikTok or Instagram, they mean the learned recommendation system deciding what you see next.

The pre-flight checklist: fixed steps, same order, every time. Traditional software is the checklist someone wrote; machine learning works out its own by watching thousands of take-offs.

analytics

Data about who visits your site and what they do — page views, sources, conversions. Google Analytics is the standard.

Web analytics is the measurement layer of a website: how many people visited, where they came from, what they did, and whether they converted (bought, signed up). Google Analytics 4 is the standard free tool; Search Console shows how you appear in Google search. Connect an AI agent to these and it can pull the numbers and explain the story in plain English.

The airport's passenger figures — how many came through the doors, which gates they used, how many actually boarded.

API

A way for programs to talk to each other. 'Using the API' = your code calling the model directly, paying per token.

An API (Application Programming Interface) is a doorway that lets one program use another program's abilities. AI companies expose their models via APIs: your code sends a prompt over the internet and gets the reply back, paying per token instead of a monthly chat subscription. Every AI-powered product you've used is calling model APIs behind the scenes.

You're in the Gordon Ramsay restaurant in Terminal 5 and you want the lobster bisque. You don't go into the kitchen and cook it — you order from the menu and it arrives. An API is that menu: an agreed list of things you can ask for, priced per dish, with the kitchen kept firmly out of your hands.

API key

Your secret password for an API — proves who you are and bills your account. Never share it or commit it to GitHub.

An API key is a long secret string that identifies your account when your code calls a service like the Anthropic API. Anyone who has it can spend your money, so treat it like a bank card number: keep it in a config file or environment variable, never paste it into shared code, and never commit it to GitHub.

Your boarding pass with your name on it. It gets you through every gate — and so does anyone who picks it up off the seat. Treat it accordingly.

attention

The transformer's key mechanism: every token looks at every other token to work out what it means in this context.

Attention is the mechanism that made modern LLMs possible: as text flows through the network, each token weighs every other token by relevance and blends in their meaning — so 'bank' near 'river' means mud, and near 'overdraft' means money. Like someone at a party tuning into the conversations that matter to them. It's the central idea of the 2017 transformer paper, literally titled 'Attention Is All You Need'.

Air traffic control tracking every aircraft at once and working out which ones actually matter to each other. The 747 on final approach matters to the one behind it; the taxiing turboprop doesn't.

attribution

Deciding which marketing touch gets credit for a conversion — the ad they saw Monday or the search they clicked Thursday?

Attribution is deciding who gets the credit when a customer met your marketing several times before acting — saw an ad Monday, clicked a Google result Thursday, bought Friday. Different models split credit differently (last click, first click, or GA4's statistical 'data-driven' model), which is why GA4 and Google Ads legitimately report different conversion numbers for the same reality. It's the messiest corner of analytics — treat the numbers as directional, not gospel.

Which bit of marketing actually got you to book — the billboard, the email, the comparison site, or the ad you clicked last? Someone has to decide who gets the commission.

AVE

Advertising Value Equivalent — a discredited old way of pricing PR coverage. Know it so you can reject it.

AVE tried to value press coverage by asking what the same space would have cost as an advert. The industry has formally rejected it for years — it ignores whether the coverage was positive, whether anyone read it, and whether it changed behaviour, and it treats PR as cheap advertising rather than a different discipline. Modern measurement uses reach, quality of coverage, message delivery, share of voice and actual business outcomes. Knowing AVE is dead, and why, is a useful signal that you follow the industry.

Valuing a good review by the price of the advert that would have sat in the same column inch. It tells you the size of the space and nothing about the meal.

B

base model

A model fresh from pre-training: a raw text-continuer, not an assistant. Chat-tuning and RLHF add the helpfulness.

A base model is what pre-training produces: a pure next-token predictor with no notion of being helpful. Ask it a question and it may just continue with more questions, because that's a plausible way for the text to carry on. It's the raw engine before the finishing school — supervised fine-tuning and RLHF — turns it into the polite assistant you actually chat with.

A pilot with thousands of hours and no customer-service training whatsoever. Technically superb, and would answer a nervous passenger with the exact failure statistics.

bias (in AI)

When an AI's outputs systematically skew against certain groups — learned from skewed training data, not from malice.

Bias in AI means the system's outputs systematically tilt against certain people or ideas — women's CVs scored lower, some faces recognised worse, job ads shown mostly to one group. It comes from the data: models learn patterns from history and the internet, and faithfully reproduce whatever unfairness lives there. Like a mirror that also magnifies: nothing evil inside it, but you must check what it reflects before acting on it — especially in hiring, credit and anything else that decides about people.

A hiring process that keeps selecting pilots who look like the last thirty pilots — because the last thirty pilots are all it was ever shown.

bot

An automated account/program that responds to messages or performs tasks — from simple scripts to full AI agents.

A bot is any automated program that acts through a chat or social platform — replying to messages, posting updates, running commands. They range from dumb scripts ('auto-reply with office hours') to full AI agents with a model brain. On Telegram, a bot is simply an account controlled by code instead of a person.

The automated voice on the airline's phone line. Some genuinely help; some exist mainly to stop you reaching a human.

BotFather

Telegram's official bot-making bot: message it /newbot and it issues your bot's token in under a minute.

BotFather is Telegram's built-in bot for creating and managing other bots. You message @BotFather, send /newbot, pick a name and a username ending in 'bot', and it replies with your bot's token — the password your agent software uses to send and receive messages. It also manages settings (privacy mode, group access) and can revoke a leaked token. It's the registry office where every Telegram bot is born.

The airline's badge office. You ask for a new staff pass, it issues one with a unique number on it. BotFather issues your bot's token the same way.

branch

A parallel line of work in a repo — try ideas without touching the main version, merge back if they work out.

A branch is a parallel line of development inside one repository: you branch off, experiment freely, and the main version stays untouched until you merge your work back in. It's the 'copy of the essay to scribble on' idea, done properly. Under the bonnet a branch is just a lightweight label pointing at a commit, so creating one is instant and free; the default branch is usually called main.

A spare copy of the timetable you can play with — “what if the Tokyo flight left at six?” — without touching the one passengers are already booked on.

brief

The written statement of what's needed and why — the thing all good work starts from.

A brief defines the problem, the audience, the objective, the constraints and the deadline. Agencies distinguish the client brief (what the client asked for) from the creative brief (the sharpened, internal version the creative team actually works from). "Is it briefed?" is a real gate — an unbriefed request is the single most common cause of wasted creative time, which makes it a resourcing issue as much as a creative one.

The difference between "get milk" and "get two litres of oat milk from the corner shop before six". One of these you can actually action.

C

Cannes Lions

The biggest awards festival in the industry — a Lion is the trophy everyone actually wants.

The Cannes Lions International Festival of Creativity, held each June, is the most prestigious awards scheme in advertising and PR. Winning a Lion is a genuine commercial asset — agencies put it in creds decks and new-business conversations for years. The Romans won Cannes Lions PR Agency of the Year in 2025. Entries are expensive, deadlines are early in the year, and preparing them is a real workstream, not an afterthought.

The Oscars, if the Oscars also directly won you next year's clients.

Canvas

An HTML element that is a blank rectangle for drawing on with code — how browser games and charts get made.

The <canvas> element gives you a blank rectangle of pixels and a set of JavaScript commands for drawing on it: lines, rectangles, text, images. Unlike normal webpage elements, the browser has no idea what is on it — there are no “objects”, only paint. Games redraw the whole thing many times a second, which is simpler and faster than working out which bits changed. Apple invented it for Safari in 2004 and everyone else copied it.

The blank departures board before anything is written on it. The board has no idea what a flight is — something else paints every letter, and repaints the lot whenever anything changes.

case study

The written-and-filmed story of a campaign: challenge, idea, execution, results.

A case study is how agency work travels. The standard shape is challenge → insight → idea → execution → results, with hard numbers at the end. Case studies feed award entries, creds decks, the website and social. Getting them written while the client team still remembers the details — and while the results data is still available — is a recurring, genuinely difficult piece of internal marketing, because everyone is busy with the next thing.

The after photo, with the before photo and the receipts attached.

chain-of-thought

Prompting a model to reason step-by-step before answering. A big unlock for older models; built into modern reasoning models.

Chain-of-thought (CoT) prompting asks a model to work through its reasoning in writing before giving the final answer — famously triggered by adding 'Let's think step by step'. Each generated step gives the model computation to build on, like doing long division on paper instead of in your head, and it dramatically improved maths and logic in earlier models. Modern reasoning models now do this internally by training, so the explicit trick matters mainly for smaller models or when you want the working shown.

Asking the pilot to talk the diversion through out loud rather than just announcing the new destination. The working-out is the bit you wanted.

chatbot

A conversational app wrapped around a model — ChatGPT and Claude.ai are chatbots; GPT-5 and Opus are the models inside.

A chatbot is the friendly conversational product wrapped around an AI model: the chat window, memory of your conversation, safety rules and tools. ChatGPT is a chatbot powered by OpenAI's models; Claude.ai is Anthropic's. Distinguishing the app from the model inside is half the battle in understanding the AI landscape.

The information desk in the terminal. The desk is the chatbot; whoever's on shift behind it is the model.

Claude Code

Anthropic's coding agent for your terminal: describe what you want; it reads files, writes code and runs commands.

Claude Code is Anthropic's agentic coding tool. It runs in your terminal, can read your project's files, write and edit code, run commands and tests, and use git — all driven by plain-English instructions. It's the difference between an AI that advises about code and one that actually does the work on your machine, asking permission as it goes.

Ground crew who work from the tarmac rather than the terminal. You say what you need, they open the panels, change the part, and tell you exactly what they did.

CLAUDE.md

A notes file in your project that Claude Code reads at startup — your standing instructions to the agent.

CLAUDE.md is a plain markdown file in a project folder that Claude Code automatically reads when it starts. Teams use it for standing instructions: what the project is, how to run tests, style rules, things to never touch. It's the difference between re-explaining your project every session and the agent just knowing.

The note left on the flight deck for whoever flies this aircraft next: “the number two radio is temperamental, don't trust the fuel gauge below a quarter.” Read every time, by everyone.

CLI

Command-Line Interface — a tool you drive by typing commands rather than clicking buttons. Claude Code is a CLI.

A CLI (Command-Line Interface) is any tool you operate by typing commands in a terminal instead of clicking a GUI. Developers love CLIs because they're fast, scriptable and automatable. Claude Code, git and brew are all CLIs — which is why lesson one of the toolkit is simply getting comfortable in the terminal.

The cockpit: switches and typed entries, no pictures. Far faster than a touchscreen once you know it, completely baffling if you don't.

clone

Copy an entire repo — files AND full history — from GitHub onto your machine: git clone <url>.

Cloning copies a whole repository to your machine: every file plus the complete commit history, not just the latest version. It's how you get a project from GitHub to work on locally — git clone followed by the address from the repo's green Code button. Your clone remembers where it came from, so push and pull know where to sync.

Taking home a full copy of the airline's operations manual — not just today's version, but every revision ever issued.

Codex

OpenAI's coding agent — their counterpart to Claude Code, driven from the terminal, IDE or cloud.

Codex is OpenAI's AI coding agent — the ChatGPT-side equivalent of Claude Code. It can work in your terminal or IDE or run tasks in the cloud, reading repositories, writing code and opening pull requests. The name is old (it powered the original GitHub Copilot in 2021) but today it means OpenAI's agentic coding product.

A rival airline's ground crew. Same job as Claude Code, different uniform, slightly different habits.

commit

A saved snapshot of your project in git, with a message saying what changed. Your save points.

A commit is one saved snapshot in git: the state of your files at a moment, plus a short message ('Add contact form'). Commits are the save points you can always return to — made cheaply and often. A project's commit history reads like a diary of how it was built.

Signing the maintenance log: “changed the port-side tyre, 14:20, here's why.” A dated snapshot you can always point back at.

compacting

Summarising older chat history to free up context-window space, so a long conversation can keep going.

When a conversation approaches the model's context-window limit, tools like Claude Code compact it: they replace older messages with a compressed summary, keeping the gist while freeing space. It's like rewriting three pages of meeting notes into one paragraph so the notepad never fills up. Details can be lost, so key facts are best saved to files.

The shift-change handover: eight hours of detail boiled down to the two minutes the next crew actually needs.

computer use

An AI capability (Anthropic, 2024) where the model operates a real screen — looking, clicking, typing — like a human would.

Computer use lets an AI model drive an actual computer interface: it looks at screenshots, moves the cursor, clicks buttons and types into forms — the same way a person does. Anthropic shipped the first mainstream version in October 2024, and it's a key building block of the agent era, because it lets AI operate the thousands of apps and websites that were only ever designed for human hands.

Handing the AI the real check-in terminal — mouse, keyboard, screen — rather than building it a special hatch. It clicks the same buttons a person would.

Content Credentials

A tamper-evident 'nutrition label' embedded in an image or video file recording how it was made — including whether AI made it.

Content Credentials are provenance metadata based on the C2PA standard, backed by Adobe, OpenAI, Google, Microsoft and camera makers: a cryptographically signed record inside the file saying what created it and how it was edited. Think of it as a nutrition label for media. Because AI detectors are unreliable and easily fooled, provenance — proving where an image came from — is winning out over trying to spot fakes by eye.

The tamper-evident seal on the duty-free bag. You can tell at a glance whether it's been opened since it left the shop.

context engineering

The craft of deciding what the model gets to see — instructions, files, history — not just how the question is worded.

Context engineering is prompting grown up: deliberately assembling everything the model sees — system instructions, relevant files, examples, conversation history — so it has exactly what it needs and nothing that confuses it. With agents this matters more than clever wording; CLAUDE.md files, MCP connections and compacting are all context-engineering tools.

Deciding what goes in the pilot's briefing pack. Everything relevant, nothing irrelevant — a 400-page pack is as useless as an empty one.

context window

The model's working memory — how many tokens it can 'see' at once. Overflow it and the oldest stuff falls out.

The context window is how much text (in tokens) a model can consider at one time — its working memory. Everything you've said, everything it replied, plus any files you shared must fit. Modern models hold anywhere from 200 thousand to a million tokens. When a long chat gets forgetful, you've hit the window — which is why tools 'compact' (summarise) old history.

The cockpit workspace. Only so many charts fit in front of the crew at once; pile on more and the ones underneath stop being read.

conversion

The moment a visitor does the thing you wanted — buys, signs up, submits the form. GA4 now calls these 'key events'.

A conversion is the payoff moment: a visitor stops browsing and does the thing the site exists for — a purchase, a sign-up, a submitted contact form. You decide what counts. Conversion rate (conversions ÷ visitors or clicks) is the honest measure of whether traffic is worth anything. GA4 renamed on-site conversions 'key events' in 2024, but everyone still says conversion out loud.

The passenger stops browsing the destinations board and actually buys the ticket.

Copilot

Microsoft/GitHub's AI assistant brand — best known for autocompleting code in your editor as you type.

Copilot is Microsoft's AI-assistant brand. GitHub Copilot pioneered AI code-autocomplete in 2021 — suggesting the next lines as you type — and has since grown chat and agent features; Microsoft 365 Copilot puts the same idea into Word, Excel and Outlook. A good name to know in interviews: it made AI-assisted work mainstream.

Microsoft's brand, and also the literal thing: sits beside you, handles the routine legs, doesn't fly the aeroplane on its own.

CPC

Cost Per Click — what you pay an ad platform each time someone clicks your ad. Core Google Ads metric.

CPC (Cost Per Click) is the price you pay each time someone clicks one of your ads on platforms like Google Ads. It's set by auction: popular keywords cost more. Watching CPC against what a customer is worth tells you if a campaign makes money — exactly the kind of analysis AI agents connected to your ads account can automate.

What the airline pays every time somebody clicks its ad — whether or not they ever board.

creds (credentials)

The agency's “here's who we are and what we've done” deck — the standard opener with any prospect.

Credentials, universally shortened to "creds", is the deck an agency presents to introduce itself to a potential client: who we are, how we think, the work we're proud of, the results it drove, the people you'd get. A creds meeting is usually the first stage of any new business conversation, before any actual brief exists. Keeping the creds deck current and on-message is standard new-business admin — and often falls to the newest person on the growth team.

A first date where you've prepared. You are not proposing yet; you are establishing you're worth dinner.

CRM

Customer Relationship Management — the shared database of contacts, conversations and opportunities.

A CRM is the system of record for commercial relationships: who we've spoken to, when, what about, what it's worth and what happens next. In an agency growth team it holds prospects, pitch history, contacts and revenue. Its value is entirely dependent on discipline — a CRM nobody updates is worse than no CRM, because people trust it and it's wrong. "Supporting on a new CRM tool" usually means being the person who makes the discipline stick.

A shared address book with a memory. Useless the moment people stop writing in it.

cron

A computer's alarm clock: runs a task on a schedule — every morning at 7, every Friday at 5 — without anyone touching anything.

cron is the decades-old Unix scheduler that runs commands at set times — 'every day at 07:00', 'every Monday'. Agent frameworks either use it directly or ship their own scheduler that works the same way. It's the missing half of most automations: a skill says HOW to do a job, cron says WHEN — put them together and the job simply happens, whether you're at the desk or not.

The 05:30 crew-alarm that rolls the first catering truck out every morning. Nobody presses go; it's just Tuesday.

CSS

Cascading Style Sheets — the language that styles webpages: colours, fonts, spacing, layout.

CSS is the styling language of the web: rules that select parts of an HTML page and describe how they should look — colours, fonts, spacing, layouts, animations. One line of CSS can restyle a thousand pages, which is the 'cascading' superpower. If HTML is the skeleton of a page, CSS is the clothes, and it's why the web stopped looking like a 1991 physics paper.

The livery, the cabin colours, the font on the signage. Identical aircraft underneath; completely different feel.

CTR

Click-Through Rate — clicks ÷ impressions. Of everyone who saw your link or ad, the percentage who actually clicked it.

CTR (Click-Through Rate) is clicks divided by impressions: of everyone who saw you, what share clicked? It measures how tempting your headline or ad is. 10 clicks from 1,000 impressions is a 1% CTR — perfectly normal in search results, since most links shown are never clicked. A low CTR on high impressions is a fixable problem: usually a boring title.

Of everyone who walked past the poster for cheap flights to Málaga, how many actually stopped and looked.

curl

A terminal command that fetches things from the internet — files, pages, APIs. The one in most install instructions.

curl fetches whatever is at a URL and, by default, prints it. Add -o to save it under a name you choose, or -O to keep the name it already has. It is the workhorse behind most one-line installers ('curl … | bash' downloads a script and runs it), and it is how you pull an image, a dataset or an API response into a project without leaving the terminal.

The baggage handler you send to collect one specific bag from one specific carousel. You give the exact tag; it goes, fetches, returns. No conversation.

Cursor

An AI-native code editor (a VS Code fork) — the whole IDE rebuilt around AI, model of your choice inside.

Cursor is a code editor built for AI-assisted programming: a fork of Microsoft's VS Code with AI wired into every interaction — autocomplete that predicts multi-file edits, chat that knows your whole codebase, and agent modes that make sweeping changes. Unlike Claude Code or Codex (terminal agents from the model labs), Cursor is an independent product that lets you pick which company's model powers it. Roughly $20/month.

A cockpit rebuilt from scratch around the autopilot, rather than an autopilot bolted into a cockpit designed before it existed.

D

daemon

A program that runs quietly in the background all the time, with no window — how an agent stays on 24/7.

A daemon (pronounced 'demon') is a background program that starts automatically and keeps running with no window or dock icon — the way web servers, sync clients and always-on agents live. Installing your agent 'as a daemon' means the operating system keeps it running and restarts it after reboots, so your assistant is awake even when you aren't logged in. The odd name is 1960s MIT humour that stuck.

The baggage belt. No shopfront, no opening hours, just running quietly underneath the terminal all day.

day rate

What a freelancer charges per day — the unit agency freelance budgets are counted in.

Freelancers bill by the day rather than by salary, and rates vary sharply by discipline and seniority. Booking freelance support means agreeing the rate, the number of days and who's approving the spend — before they start, not after. Tracking freelance usage and spend across a department is a standard part of a resourcing role, because it's the budget line that grows quietly when teams are stretched.

Hiring a session musician. You agree the fee and the number of sessions up front, because "we'll sort it later" is how you end up with an invoice nobody approved.

deep learning

Machine learning using many-layered neural networks. The branch that conquered vision, speech — and, via LLMs, language.

Deep learning is machine learning done with neural networks that have many layers — 'deep' refers to the layer count, not profundity. Each layer learns progressively more abstract patterns: edges, then shapes, then faces, say. It took over vision and speech recognition in the early 2010s once GPUs made training big networks practical, and LLMs like the ones behind ChatGPT are deep learning at its most extreme scale.

Not one inspector checking bags but forty in a row, each catching what the last couldn't: first the shape, then the density, then the outline of a bottle.

deepfake

AI-generated or AI-manipulated image, audio or video that convincingly resembles real people, places or events.

A deepfake is synthetic media realistic enough to pass for a genuine recording — a face swapped into a video, a cloned voice, an invented photo of a real event. Originally slang, it's now a legal term: the EU AI Act requires deepfakes to be clearly labelled as artificial from August 2026, and many countries criminalise intimate or election deepfakes. The word covers harmless labelled creations and malicious deceptions alike — disclosure is what separates them.

A pilot's voice on the radio that sounds exactly right, says exactly the wrong thing, and isn't the pilot.

deploy

Putting your website or app onto a real server so the world can use it. 'Deployed' = live on the internet.

Deploying is moving your code from your machine to a server where real users can reach it — copying the files, wiring up the domain, switching it live. Modern hosts (Vercel, Netlify, GitHub Pages) make deploying a website nearly one-click, and AI agents can handle it for you.

The moment the new route stops being a plan on a whiteboard and appears on the departures board where anyone can book it.

diff

A line-by-line view of what changed between two versions — additions in green, deletions in red. How you review work.

A diff shows exactly what changed between two versions of files, line by line — added lines marked with + (green), removed lines with − (red). It's how developers review each other's work in pull requests, and how you audit an AI agent's edits without re-reading whole files: git diff shows what's changed since your last save point.

The revision marks on an updated flight plan: this altitude changed, this waypoint was added, everything else stayed put.

diffusion model

The AI technique behind most image generators: start from noise, denoise step-by-step into a picture.

Diffusion models generate images by learning to reverse noise: training teaches the model to turn a noisy mess slightly closer to a real image, and generation runs that denoising step dozens of times from pure static until a picture emerges — steered by your text prompt. Midjourney, Stable Diffusion and DALL·E work this way; video models extend the idea across frames.

Start with a photograph ruined by static, clean it up a little, repeat — until an aeroplane appears that was never actually photographed.

directory

The terminal's word for a folder — same thing, older name. mkdir means 'make directory'.

A directory is what the terminal calls a folder — the concept predates the folder icon by decades. Directories nest inside each other to form the filesystem tree, and most core commands are about them: mkdir makes one, cd moves into one, ls lists what one contains, pwd tells you which one you're standing in.

A folder, in older language. The airport calls it a stand; you'd call it a parking space. Same thing, different decade.

dual-use

Technology useful for both ordinary and harmful ends — the same AI skill that reviews your code can also hunt for exploits.

'Dual-use' describes a capability that helps or harms depending on who aims it where. A model skilled at reading code to catch bugs is, by that same skill, skilled at finding vulnerabilities to exploit — the reason Anthropic's Fable 5 briefly triggered US national-security controls in 2026. Governments have long treated dual-use tech — encryption, nuclear material, advanced chips — as strategic; frontier AI is the newest entry on that list.

The same scanner that finds a bomb finds a passenger's insulin pump. The tool has no opinion about which you meant.

E

earned media

Coverage you got because it was interesting — not because you paid for the space.

The three media types are paid (you bought the slot: ads), owned (your own channels: your site, your social) and earned — coverage, shares and conversation you got because someone thought it was worth talking about. PR's whole craft is earned. It's why creative PR agencies build stunts and stories rather than ads: nobody shares an ad, but a handbag shaped like a croissant makes the papers for free.

Paid is buying a round so people like you. Earned is being good enough company that someone buys you one.

embedding

A word/token represented as a list of numbers (a vector), arranged so similar meanings sit near each other in space.

An embedding is how a model represents a token: a vector of hundreds or thousands of numbers, learned so that geometry mirrors meaning — 'cat' sits near 'kitten', and relationships become directions (the classic demo: king − man + woman ≈ queen). It's like a map where every word has coordinates and distance means similarity. Embeddings also power search and RAG: turn texts into vectors, then find the ones closest to your query.

Arranging the route map by similarity rather than alphabetically — Málaga next to Alicante, Tokyo nowhere near either. Numbers instead of names, positioned so that nearness means something.

emergent abilities

Skills like translation or step-by-step reasoning that appeared as models scaled up — nobody explicitly programmed them.

Emergent abilities are capabilities that show up when language models get big enough — translation, arithmetic, multi-step reasoning — despite nobody building them in; they condense out of next-token prediction at scale. Like a crowd suddenly producing a Mexican wave no individual planned. How 'sudden' emergence really is remains scientifically debated, but the abilities themselves are real.

Build a big enough airport and it starts doing things nobody designed: hotels, conference centres, a small city. None of that was in the plans.

expert system

1980s AI built from thousands of hand-written IF–THEN rules copied from human experts. Boomed, then went bust.

An expert system encodes a specialist's knowledge as a huge library of hand-written IF–THEN rules plus an engine that chains them together — like turning a doctor's brain into a giant flowchart. Famous ones diagnosed infections (MYCIN) and configured computer orders (XCON). They worked in narrow niches but were brittle and agonising to update, since every rule was written by hand — and their collapse triggered the second AI winter.

A thousand laminated cards. IF fog AND crosswind THEN divert. Superb, right up until the weather does something that isn't on a card.

export controls

Government rules that restrict who may access sensitive technology across borders — in 2026, applied to frontier AI models for the first time.

Export controls are national-security rules that limit moving sensitive technology — historically weapons, encryption and advanced chips — to particular people or countries. In June 2026 they reached AI models directly: the US briefly ordered Anthropic to block its most capable models (Fable 5 and Mythos 5) from every non-US person. Because software runs the same for everyone, 'exporting' a model can mean simply granting API access, which makes such rules blunt — the only way to reliably block foreign nationals was to switch the models off for almost everyone.

The rules about which countries may buy the engines. Nothing to do with what the aircraft can do — everything to do with whose hands it ends up in.

F

few-shot prompting

Putting a handful of worked examples in the prompt so the model copies the pattern — often beats describing what you want.

Few-shot prompting means including two or more example input→output pairs in your prompt so the model infers the pattern rather than parsing a description of it. ('Zero-shot' means no examples, 'one-shot' means one.) It's like training a new assistant by showing them three finished reports instead of writing a style manual. Especially strong for classification, data extraction and matching a house style — but beware: the model copies everything about your examples, including the accidents.

Showing the new check-in agent three completed forms before letting them do the fourth. Nobody explains the rules; they copy the pattern.

fine-tuning

Extra training on your own examples to specialise a general model — e.g. teaching it your company's support style.

Fine-tuning takes a trained general model and continues training briefly on your own examples — thousands of support tickets, legal clauses, brand-voice samples — to specialise it. It changes the model's default behaviour, unlike prompting (instructions at ask-time) or RAG (handing it documents to read). Most teams get far with prompting before ever needing it.

Taking a fully qualified pilot and adding a type rating for the A350. Same pilot, one specific aircraft now known properly.

flag

A dash-prefixed option that tweaks how a command behaves: ls -l gives a long, detailed listing.

A flag (or option/switch) is an extra you add after a command, usually starting with a dash, to change its behaviour. ls lists files; ls -l lists them with details; ls -la also reveals hidden ones. Flags are like the toggles in an app's settings panel — nobody memorises them all, they check the manual (man ls) or ask an AI.

The extras on a ticket. Same flight, but -window, -extra-legroom, -no-meal. Small switches that change what you actually get.

framework

A pre-built code skeleton (React, Next.js…) that handles a project's plumbing. Powerful for big sites, overkill for one page.

A framework is a ready-made structure you build software inside — it supplies the plumbing (components, routing, data handling) so you write only what's unique to your project. React and Next.js dominate the web world. The trade-off is machinery: build steps, dependencies and updates to maintain. Rule of thumb: frameworks pay off when a site has many pages sharing parts or real interactivity, and are overkill for a one-page site.

A terminal that already has the gates, belts and security lanes built. You fit out the shops; you don't pour the concrete.

freelancer

Self-employed talent booked by the day to cover peaks, gaps and specialisms.

Agencies keep a trusted network of freelance creatives, strategists, designers and producers to flex capacity up when work spikes and down when it doesn't. Maintaining that network — knowing who's brilliant, who's available, who fits which brief — is a genuine asset, and it's a named responsibility in most resourcing roles. A good freelance book means you can say yes to work that would otherwise be declined.

A pool of very good substitutes you can call on a Tuesday. The value isn't the list, it's knowing which one to ring.

frontier model

A lab's biggest, most capable model — the top tier of the family, priciest and slowest but smartest.

A frontier model is one at the leading edge of capability — the biggest, smartest model a lab offers, like OpenAI's top GPT tier or Anthropic's Opus-and-above. Labs ship families for a reason: frontier models cost the most per token and respond slowest, so mid-size and small models handle everyday work. 'Frontier lab' means the handful of companies (OpenAI, Anthropic, Google, and increasingly Chinese labs like Moonshot) training models at this edge.

The flagship route flown with the newest aircraft in the fleet. The best the airline has, and priced like it.

function calling

How AI gets things done: the model writes a structured request ('call check_weather for Leeds') and the app executes it.

Function calling (also 'tool use' or 'tool calling') is the mechanism that turns a text-only model into something that acts. The app declares a menu of tools; when the model wants one, it outputs a structured request — tool name plus arguments — instead of prose, and the surrounding software executes it and feeds the result back. The model is an advisor passing notes from a phone box; the software is the runner who carries them out.

The pilot radioing the tower for the current runway rather than guessing. A structured question out, a structured answer back, and only then a decision.

G

GA4

Google Analytics 4 — the free standard tool measuring who visits your site and what they do: users, sessions, conversions.

Google Analytics 4 is the current version of Google's free web analytics tool — the CCTV inside your website. It records visits as streams of events (page views, clicks, purchases) and rolls them up into users, sessions and conversions ('key events'). It replaced Universal Analytics in 2023, and it's the first data source any marketing analyst — human or AI — gets asked about.

The airport's official passenger-counting system. Free, standard, and everybody complains about the new version.

GAN

The pre-diffusion image-generation technique: two networks — a forger and a detective — trained against each other.

A GAN (Generative Adversarial Network) trains two networks in competition: a generator forges images while a discriminator learns to tell fakes from real ones, each forcing the other to improve. GANs powered the first 'this person does not exist' face generators and early deepfakes, but they were unstable to train and hard to steer with text, so diffusion models displaced them for mainstream image generation.

A forger printing fake boarding passes and an inspector trying to spot them, each getting sharper because of the other, until the fakes start getting through.

Gemini

Google's AI — both the model family and the chat app, woven through Search, Gmail and Docs.

Gemini is Google's AI brand, covering both its model family (the Gemini 3 generation as of mid-2026, with Pro at the top and fast Flash models below) and the chat app at gemini.google.com. Its distinctive strength is integration: the same models power AI answers in Search and assist inside Gmail, Docs and the rest of Workspace. Like GPT and Claude, the models are closed — used via Google's apps or API.

Google's airline — and it also runs the search desk, the email, and the phone in your pocket.

git

Version control: snapshots ('commits') of your project you can always roll back to. The safety net for all coding.

Git is version-control software: it records snapshots (commits) of your project over time, so you can see what changed, when, and roll back anything. It's also how people collaborate without overwriting each other. Git is the tool on your machine; GitHub is the website where git projects get stored and shared.

The maintenance log for your entire project: every change, who made it, when, and the ability to put it back exactly as it was.

GitHub

The website where git projects live online — backup, sharing, and collaboration. Owned by Microsoft.

GitHub is the world's main home for code: it hosts git repositories online, adding backup, sharing, issue tracking and pull requests (proposed changes others can review). Most open-source software lives there, employers look at it like a portfolio, and AI agents can use it too — pushing code and opening pull requests on your behalf.

The central archive where every airline's maintenance logs are stored — readable, searchable, copyable.

GitHub Pages

GitHub's free static-site hosting: push files to a repository, flip a switch in Settings, and you're live at username.github.io.

GitHub Pages is a free hosting service built into GitHub: point it at a public repository and it serves the files there as a live website at username.github.io/repo-name (or your own domain). Every push updates the live site within a minute or two — deployment reduced to 'save your work'. It only serves static files, which is exactly why it can be free, and it has hosted personal sites, project pages and documentation since 2008.

The airport's free noticeboard. Pin your page up, it gets a public address, nobody charges you rent.

GPT

Generative Pre-trained Transformer — OpenAI's model family and the recipe name: a transformer pre-trained on huge text.

GPT stands for Generative Pre-trained Transformer: a transformer neural network, pre-trained on vast amounts of text to predict the next token, that can then generate new text. It's OpenAI's model family name (GPT-2, GPT-3, GPT-5…) and shorthand for the general recipe behind most modern LLMs. ChatGPT is literally 'chat' + 'GPT' — a friendly chat interface wrapped around one.

OpenAI's aircraft family, and also the design they're all built on — like saying “the 737” and “twin-jet narrowbody” in the same breath.

GPU

The graphics chip that turned out perfect for AI maths. Thousands of them train frontier models; NVIDIA got rich.

A GPU (Graphics Processing Unit) does thousands of simple calculations in parallel — originally for game graphics, but exactly what neural networks need. Training a frontier model uses tens of thousands of GPUs running for weeks, which is why compute is the bottleneck of the AI industry and why NVIDIA became one of the world's most valuable companies.

The baggage system. Hopeless at carrying one suitcase, unbeatable at forty thousand at once. AI maths is forty thousand suitcases.

guardrails

The limits and checks around an AI system — what it may not do, must confirm first, or must never say.

Guardrails are the deliberate limits around an AI system: content it refuses, actions that need human approval (spending money, deleting files, sending emails), filters on inputs and outputs, and monitoring. When you deploy AI at work, designing the guardrails is as important as choosing the model — they're what turn a clever demo into something safe to trust.

The barriers, the security lanes, the door the crew won't open in flight. Not there to slow you down — there because of what happens without them.

GUI

Graphical User Interface — the windows, icons and menus you click. The opposite of typing commands in a terminal.

A GUI (Graphical User Interface, pronounced "gooey") is the layer of a computer you point at: windows, icons, menus, buttons, the Dock. It replaced typing commands as the normal way to use a computer in the 1980s, and it made computers usable by everybody. It didn't remove the command line underneath, though — macOS is a GUI built on top of Unix, which is exactly why the Terminal app exists and why the same commands work on a Linux server.

The departures board, the signs and the self-service check-in screens. Almost every passenger only ever uses those — but the airport still runs on radio procedure underneath, and that's what the crew are using.

H

hallucination

When AI states something false with total confidence — invented facts, citations, prices. Always verify what matters.

A hallucination is a confident-sounding fabrication: the model produces plausible text (a citation, a statistic, an API that doesn't exist) because plausible is what it's trained to produce — truth isn't checked. It's the single most important limitation to understand at work. Rule of thumb: the more specific and checkable a claim, the more you should verify it.

The gate agent who announces a gate number that doesn't exist, in precisely the tone they'd use for a real one. The confidence is not the tell.

Hermes

An open-source personal-agent framework — run your own always-on AI assistant and talk to it via Telegram.

Hermes is an open-source framework for running a personal AI agent on your own machine or server — always on, with memory, able to use tools, and reachable through chat apps like Telegram. You bring the model (e.g. Claude's API) and it supplies the body: sessions, tool use, and messaging integrations.

Your own aircraft in your own hangar, with a logbook it keeps for itself between flights.

Homebrew

The Mac's app store for developer tools — `brew install python` fetches and sets up software in one command.

Homebrew is the standard package manager for macOS: a tool that installs other tools. Instead of hunting websites and download links, you type `brew install python` (or git, or node) and it fetches, installs and wires everything up. Almost every Mac developer setup guide starts with installing Homebrew.

The airport supply depot. Ask for the part and it turns up fitted, rather than you ringing round suppliers yourself.

hook

A shell command the agent tool runs automatically at fixed moments (before/after tool use) — enforced rules, not polite suggestions.

A hook is a command that an agent tool like Claude Code executes itself at fixed points in the workflow — before a tool runs (and can block it), after a file is edited (say, auto-format it), or when the agent finishes. The crucial difference from CLAUDE.md: instructions are advisory, hooks are enforced regardless of what the model decides. 'Please don't read the secrets file' is a request; a hook is a locked door.

The things that happen automatically whether or not anyone remembers: doors to automatic on push-back, seatbelt sign on descent. Fixed moments, fixed actions.

HTML

HyperText Markup Language — the structure of every webpage. Tags like <h1> and <p> say what each piece of content is.

HTML is the language webpages are written in: nested tags that label content by what it is — headings, paragraphs, links, images. It's the skeleton of a page; CSS dresses it and JavaScript animates it. Every website you've ever seen arrives at your browser as HTML, and 'view source' shows it — nothing on the web is truly hidden.

The terminal building's structure — where the walls, doors and signs go. Bare concrete until CSS paints it.

human-in-the-loop

Designing AI systems so a person approves the important steps — especially spending, deleting or messaging real people.

Human-in-the-loop means an AI system pauses for human judgement at key points instead of running fully on autopilot. With agents, the standard pattern is graded by reversibility: reading data flows freely, changes prompt for approval, and irreversible actions — spending money, deleting things, contacting real humans — always require an explicit yes. It's the difference between a trusted assistant and an unsupervised one with your credit card.

The autopilot flies, but a person still decides whether to divert. The machine does the work; the human owns the call that matters.

I

IDE

Integrated Development Environment — an app for writing code, like VS Code. A word-processor for programmers.

An IDE is the app programmers write code in — VS Code is the most popular. It's a text editor plus conveniences: colours the code, spots errors, runs it, and integrates with git. AI has moved in too: Claude Code and Copilot plug into IDEs so you can ask for changes right where the code lives.

The full flight-planning suite: charts, weather, calculations and somewhere to write it all down, in one window.

impressions

How many times your link or ad was SHOWN to people — whether or not anyone clicked. The window-shoppers metric.

An impression is counted every time your search result or ad appears in front of someone. Impressions measure visibility, not interest — window-shoppers walking past, not customers coming in. Divide clicks by impressions and you get CTR, which tells you whether your shop window (your page title, your ad copy) is actually tempting anyone inside.

How many people walked past the poster. Not how many looked at it.

inference

Using a trained model — every chat reply is inference. Training is learning; inference is answering.

Inference is what happens every time you use an AI: your prompt runs through the trained model to produce a reply. No learning happens — the model's numbers are frozen. It's the difference between years of medical school (training) and a doctor answering your question in clinic (inference).

Every flight the pilot actually operates. Training was flight school; inference is the Tuesday morning to Málaga.

inpainting

Editing part of an AI image: mask a region and have the model regenerate only that bit — swap the jumper, keep the face.

Inpainting is surgical image editing with a diffusion model: you paint a mask over the region to change, describe what should go there, and the model regenerates only the masked area while blending it with the rest. It's how a sponsor's product ends up convincingly in an AI influencer's hand. The companion trick, outpainting, extends an image beyond its original borders.

Retouching the aircraft in the photo without touching the sky behind it. Mask the bit you want changed and leave the rest alone.

J

jailbreak

A prompt crafted to trick a model past its own safety rules into doing something it is meant to refuse.

A jailbreak is an input engineered to make a model ignore its own guardrails — via role-play framings, hidden instructions, or wording that slips past the training that tells it to refuse. It differs from prompt injection, which smuggles instructions into content the model reads; a jailbreak goes at the model's safety rules head-on. Amazon researchers' jailbreak of Fable 5 — coaxing it into showing how a software flaw could be abused — helped trigger that model's 2026 export ban. Labs patch known jailbreaks continuously, but new ones keep surfacing.

Talking your way through a door you're not badged for by phrasing it as something you are. The lock is fine; the conversation isn't.

JavaScript

The programming language that runs inside web browsers — it makes pages interactive rather than static.

JavaScript is the programming language browsers run: it makes webpages do things — respond to clicks, animate, fetch fresh data without reloading. HTML is structure, CSS is style, JavaScript is behaviour. Despite the name it's unrelated to Python-rival Java; it was famously created in about ten days in 1995 and now runs on virtually every device with a screen.

What makes the departures board update itself instead of being a printed sheet somebody replaces hourly.

JSON

The universal text format programs use to exchange structured data — curly braces, quotes, and key: value pairs.

JSON (JavaScript Object Notation) is the plain-text format most software uses to pass structured data around: {"name": "Isabelle", "city": "Leeds"}. It's human-readable and machine-parseable, which is why APIs, config files and AI tool calls all lean on it. When an agent 'calls a tool', the request travelling under the hood is almost always a small piece of JSON.

The baggage tag. Not written to be read by you — written so every machine down the line reads it identically.

K

Kimi

Moonshot AI's model family (China) — strong open-weights models you can run or use cheaply via API.

Kimi is the model family from Moonshot AI, a Chinese lab that broke out with Kimi K2 — a huge open-weights model with strong coding and agentic skills at a fraction of Western API prices. It matters as proof that frontier-class AI now comes from many places, and that open-weights models are genuine competition.

A Chinese carrier with a genuinely strong fleet and, unusually, one that will sell you the aircraft rather than only the seat.

L

Llama

Meta's open-weights model family — downloadable engines that kick-started the open-model movement.

Llama is Meta's family of open-weights models: the trained model files are published so anyone can download, run and fine-tune them. The original 2023 version famously leaked online within a week, accidentally launching the open-model era. Llama 4 (2025) is natively multimodal and uses a mixture-of-experts design. 'Open-weights' has caveats — the licence restricts very large companies and isn't offered in the EU — but Llama remains the reference point for running capable AI on your own hardware.

Meta's fleet, sold rather than chartered. You get the aircraft and you fly it from your own hangar.

LLM

Large Language Model — the text-prediction engine behind ChatGPT and Claude. It predicts the next token, over and over.

A Large Language Model is a neural network trained on enormous amounts of text to do one deceptively simple thing: predict the next token (word-chunk). Done billions of times with billions of parameters, that turns into answering questions, writing code and holding conversations. ChatGPT, Claude, Gemini and Kimi are all products built around LLMs — like different cars built around different engines.

The system that finishes your sentence at check-in because it has heard ten million check-ins. Extraordinarily good at what comes next — and that genuinely is all it does.

localhost

The address that means 'this very computer' (127.0.0.1). Visiting http://localhost:8000 loads a site served by your own machine.

localhost is the standard name every computer uses for itself, mapping to the IP address 127.0.0.1. Run a web server on your machine and visit http://localhost:8000, and the request never touches the internet — your browser is talking to your own computer. It's how developers preview websites privately before deploying: a full dress rehearsal that only you can see.

Your own front room. The site is running on your machine and visible to precisely nobody else — a full dress rehearsal with the doors locked.

LoRA

A small add-on training file that teaches an image model one specific thing — a face, a style, a product — without retraining the whole model.

LoRA (Low-Rank Adaptation) is a lightweight way to fine-tune a model: instead of adjusting billions of parameters, you train a tiny add-on file on a few dozen example images, then load it alongside the base model. It's like handing a session musician a one-page cheat sheet for your song rather than sending them back to music school. LoRAs are how AI influencers keep the same face across thousands of posts, and thousands of community-made ones exist for open-weights models like Stable Diffusion and Flux.

A small extra rating clipped onto an existing licence: this pilot, this one aerodrome, nothing else about them changed.

M

machine learning

Software that learns patterns from examples instead of following hand-written rules. The technique behind almost all modern AI.

Machine learning is the approach where software learns its behaviour from data: show it millions of examples (spam and not-spam, dogs and not-dogs) and it works out the distinguishing patterns itself, instead of a programmer writing the rules by hand. It's like teaching a child by showing flashcards rather than reciting a rulebook. Nearly everything called 'AI' today — spam filters, Netflix recommendations, ChatGPT — is machine learning underneath.

Instead of writing rules for spotting a bag that's about to go missing, you show the system a million bags and let it work out what “about to go missing” looks like.

Markdown

Simple text formatting with symbols: # for headings, ** for bold. The .md files you see everywhere.

Markdown is a way to format plain text with a few symbols — # makes a heading, **bold** makes bold, - makes a list. Files end in .md (README.md, CLAUDE.md). It's everywhere in tech because it stays readable as raw text but renders nicely — and it's what AI tools both read and write most comfortably.

The shorthand on a handwritten crew note — underline for important, dash for a list. Plain text that still reads as formatted.

MCP

Model Context Protocol — a standard plug socket for AI: wrap a tool once, and any AI app can use it. Also called 'USB-C for AI'.

MCP (Model Context Protocol) is an open standard — started by Anthropic, since adopted industry-wide — for connecting AI models to outside tools and data. Think of the plug sockets in your walls: every appliance ends in the same three pins, so a toaster maker never has to ring your electrician. An 'MCP server' is that plug on the end of a tool (Google Analytics, a database, Slack), and any MCP-capable AI can use it. Before MCP, every integration was custom-built for one app and one tool.

The standard fuel coupling. Any aircraft, any airport, any supplier, because everybody agreed on the shape of the connector.

MCP server

A small connector program that exposes one service (e.g. Google Analytics) to AI tools via the MCP standard.

An MCP server is the connector piece: a small program that wraps one service — Google Analytics, GitHub, your database — and exposes its actions in the standard MCP shape. You configure your AI tool to connect to it, and the model can then query analytics or file issues by itself. Hundreds exist ready-made; you rarely need to write one.

The fuel truck itself. One service, one standard coupling, plugged into whatever aircraft turns up on the stand.

memory file

Notes an AI agent writes for itself between sessions — learned commands, preferences, gotchas — stored as plain text you can read and edit.

A memory file is the agent's own notebook: plain markdown notes it writes to itself as it works ('tests need the local database running', 'user prefers pnpm') and reads back at the start of later sessions. Where CLAUDE.md holds instructions you wrote, memory holds learnings the agent accumulated. In Claude Code this is 'auto memory', kept per project in a folder you can open, audit and edit — literally your robot's diary.

The logbook the aircraft keeps between flights: what was fixed, what still rattles, what the last crew learned the hard way.

model

The trained 'brain' itself — a big file of learned numbers. GPT-5, Claude Opus and Kimi are models.

In AI, a model is the trained artefact — millions or billions of learned numbers (parameters) that encode patterns from training data. It's the brain; the app you chat with is just the body around it. One company ships many models: bigger ones are smarter but slower and pricier, smaller ones are fast and cheap.

The aircraft itself, sitting in the hangar. Not the airline and not the flight — the machine, with everything it can do already built in.

monorepo

One big repository holding many projects or packages — common at large companies; agents need layered context files, one per area.

A monorepo is a single repository containing many projects — the web app, the API, shared libraries — all versioned together. Google and Meta famously run huge ones. For AI agents they pose a context problem: one instructions file can't describe fifty packages, so the pattern is layered files — a lean root file for universal rules plus per-package files that load only when the agent works in that area.

One airline running every route, aircraft type and subsidiary out of a single operations manual.

multi-agent

Several AI agents working together — typically an orchestrator delegating to specialist sub-agents in parallel.

A multi-agent system splits a big goal across several AI agents: an orchestrator breaks the work down and delegates to specialists (a researcher, a coder, a reviewer), often running in parallel with their own context windows and tools. It mirrors how human teams work, and has the same overheads — coordination, hand-offs, and errors propagating between members — so verification steps matter even more.

The captain doesn't load the bags or refuel. A shift manager delegates to ground crew, catering and fuelling, all working at the same time.

MVP (minimum viable product)

The smallest version of a thing that proves the point — shipped before you feel ready, so reality can vote early.

A minimum viable product is the deliberately tiny first version of something: just enough to test whether the idea lands, and nothing more. Coined by Frank Robinson in 2001 and made famous by Eric Ries's The Lean Startup, the discipline is about learning speed — every feature you don't build tonight is a mistake you don't have to unmake tomorrow. An MVP isn't a worse version of the product; it's the cheapest possible version of the truth.

A wedding cake tasting. Nobody bakes the four-tier cake to find out whether you like lemon — one cupcake answers the question.

N

negative prompt

A separate 'never include' list many image tools accept alongside your main prompt — e.g. 'text, watermark, extra fingers'.

A negative prompt tells an image generator what to avoid, as a separate list alongside the main prompt: 'blurry, text, watermark, extra fingers, plastic skin' is a classic. Under the hood the model steers each de-noising step away from those concepts, like a satnav told which roads to avoid. Support varies by tool — some frontier models follow 'don't include X' written in the main prompt instead.

The box on the booking form marked “no middle seat, no red-eye, and nothing connecting through Charles de Gaulle.”

neural network

Layers of simple maths units, loosely inspired by brain neurons, that learn patterns from data. LLMs are giant ones.

A neural network is software built from layers of simple mathematical units ('neurons') connected by adjustable numbers (weights). Data flows through; during training the weights adjust until useful patterns emerge. Stack enough layers ('deep learning') and feed enough data, and networks learn to recognise faces, transcribe speech — or, in an LLM's case, predict text.

Rows of staff each doing one tiny check and passing it on — is it heavy, is it metal, is it the right shape. No single one understands the bag. Together they do.

new business

The function that wins new clients — pipeline, pitches, prospecting, the lot.

New business (often "new biz", or bundled into "growth") covers everything involved in winning work from clients you don't yet have: generating leads, qualifying opportunities, running creds meetings, managing pitch processes, and tracking the pipeline. It sits alongside organic growth (getting more from existing clients). Win rate — the proportion of pitches converted — is the headline metric; The Romans has publicly cited 89%.

The sales function, wearing a much better jacket.

next-token prediction

The one trick LLMs do: guess the most likely next word-chunk, append it, repeat. Every AI answer is built this way.

Next-token prediction is the core task an LLM is trained on and the only thing it does when generating: given all the text so far, output a probability for every possible next token, pick one, add it on, and go again. It's like ultra-advanced autocomplete — except doing it brilliantly forces the model to absorb grammar, facts and reasoning patterns, which is where all the apparent intelligence comes from.

Guessing the next word of the tannoy announcement. “Would passenger Smith please proceed to…” — you already know it's “gate”. Do that forever and you have an LLM.

O

OAuth

The “sign in with your account” browser handshake — it hands an app a revocable pass instead of your password.

OAuth is the standard behind every “Sign in with Google/Apple/Claude” button. Instead of giving an app your password, you're bounced to the provider's own page, you approve what the app may do, and the app receives a token: a limited, revocable pass that says “this tool may act for this person”. The app never learns your password, and you can cut it off later without changing anything else. For AI tools it also decides the bill — signing in this way draws on the subscription you already pay for, rather than metering you per token like an API key does.

A valet key. It starts the car and that's it — no boot, no glovebox — and you can take it back without changing the locks.

open-weights

Models whose learned weights are downloadable to run yourself (Llama, Kimi) vs closed models used via API (GPT-5, Claude).

An open-weights (loosely 'open-source') model publishes its trained parameters so anyone can download and run it on their own machines — Meta's Llama and Moonshot's Kimi are examples. Closed models like GPT-5 and Claude are only reachable through the maker's API. Open gives control and privacy; closed usually offers the strongest capability with zero infrastructure.

Buying the aircraft versus buying a seat. Open weights means it's yours, in your hangar, flown by you. Closed means you get the flight, never the machine.

OpenClaw

A popular open-source personal AI assistant you self-host and chat with via Telegram/WhatsApp; it can act, not just chat.

OpenClaw is an open-source personal AI assistant that you host yourself: it connects to messaging apps like Telegram or WhatsApp, keeps long-term memory, and can actually do things — browse, run commands, manage files, check your analytics — using an AI model as its brain. It's a hands-on way to understand what 'agent' really means.

A personal aircraft you keep at home and summon by text message — which is a genuinely strange sentence to be able to write.

over-servicing

Doing more work for a client than they're paying for — the quiet way agencies lose money.

Over-servicing is delivering beyond the agreed scope without charging for it: the extra revisions, the favour, the "quick" request that eats a day. It rarely comes from one big decision; it accumulates from many small yeses. It's a resourcing problem as much as a finance one, because over-serviced accounts silently consume the capacity that other projects were counting on.

Death by a thousand quick favours. Every one is reasonable; the total is a person's entire month.

P

package manager

A tool that installs and updates software libraries for you — brew for Mac apps, pip for Python, npm for JavaScript.

A package manager automates installing software and, crucially, its dependencies (the other software it needs). Homebrew manages Mac tools, pip manages Python libraries, npm manages JavaScript ones. One command pulls the right versions from a central registry — the reason modern developers can build on thousands of shared libraries.

The parts depot. Order by part number and it arrives with every other part it needs to fit.

parameters

The billions of learned internal numbers — 'dials' — inside a model. Training sets them; together they ARE the model.

Parameters are the adjustable numbers inside a neural network — think of billions of tiny dials on a mixing desk. During training, every wrong guess nudges some dials a smidge; after billions of nudges, their positions encode everything the model 'knows'. Model sizes are quoted in parameters (millions, billions, or more), and roughly speaking more dials means more capability but slower, pricier inference.

The billions of tiny settings inside the aircraft's systems, each nudged during testing until the whole thing flies properly.

path

A file or folder's full address in the filesystem tree, like /Users/marc/mysite/hello.txt — slashes separate the levels.

A path is the address of a file or directory, written as the route from somewhere to it with slashes between levels. An absolute path starts at the root of the tree (/Users/marc/mysite/hello.txt); a relative path starts from wherever you currently are (mysite/hello.txt). Think of an absolute path as a full postal address and a relative one as 'two doors down from here'.

The full address of a stand: Terminal 5, Pier B, Stand 42. Start at the top, follow it down, arrive at exactly one place.

perceptron

The 1958 ancestor of neural networks — a single-layer learning machine that sparked wild hype, then a long freeze.

The perceptron, built by Frank Rosenblatt in 1958, was the first famous learning machine: a single layer of adjustable weights that could be trained to classify simple patterns. The press promised conscious machines; then Minsky and Papert's 1969 book proved single-layer networks had hard mathematical limits, and neural research froze for over a decade. Today's networks are essentially perceptrons stacked millions deep — the idea was right, just seventy years early.

The 1958 prototype, with newspapers announcing that pilots were finished. It could barely leave the ground, and the funding left for twenty years.

permission prompt

Claude Code's 'may I?' — the approval it asks for before editing files or running commands. The core safety feature of agent tools.

A permission prompt is the approval step an AI agent shows before doing something with side effects — editing a file, running a shell command, using the network. In Claude Code you can allow once, allow for the session, or save a standing rule, and stricter or looser modes exist for different trust levels. It's what separates 'AI that acts under supervision' from 'AI loose on your laptop', and it's why agent tools are safe to use.

Cabin crew asking before they move your bag rather than after. Mildly interrupting; considerably better than the alternative.

PII

Personally Identifiable Information — names, emails, addresses, health or financial details. Never paste it into a consumer chatbot.

PII (Personally Identifiable Information) is any data that identifies a real person: names, email addresses, phone numbers, addresses, and especially sensitive categories like health or financial details. Privacy law (UK GDPR among others) protects it, and your employer is legally responsible for how it's handled — including what gets pasted into AI tools. Rule of thumb: PII goes only into approved enterprise-tier tools with contractual data protections, never into a personal chatbot account.

What's on your passport and your booking — name, date of birth, address, who you're travelling with. The things that identify a person, rather than a passenger count.

pip

Python's package manager — `pip install requests` fetches a ready-made Python library plus everything it needs.

pip is Python's package manager: it downloads libraries from PyPI, the community index of several hundred thousand Python packages, and installs them with their dependencies. Where Homebrew installs tools onto your Mac, pip stocks your Python toolbox with parts — the standard division of labour: brew for tools, pip for Python libraries. On a Mac you'll usually type pip3.

The parts depot again, but Python's own, stocking only Python parts.

pitch

The competitive process where several agencies present for the same piece of business.

A pitch is a competitive bid: several agencies are briefed on the same problem and present their answer, usually over a few weeks, usually unpaid. It typically runs creds → brief → tissue session → final presentation, and it consumes enormous amounts of senior time. That's why pitch scheduling is a resourcing nightmare — a pitch appears suddenly, is unmissable, and takes exactly the people already booked on paying work.

A bake-off where four bakeries make the same cake for free and only one gets paid.

plugin

A shareable bundle that adds skills, subagents, commands or tools to an agent in one install — usually from a 'marketplace'.

A plugin packages extras — skills, subagents, slash commands, hooks, even MCP servers — into one installable bundle, so a whole capability set drops into your agent at once instead of copying files by hand. Claude Code installs them from plugin 'marketplaces' (really just a Git repo listing plugins); other tools have their own stores. Think browser extensions, but for your AI assistant.

The kit that converts a passenger aircraft into a freighter. Seats out, rollers in, one bundle, fitted in an afternoon.

prompt

What you type to the AI. The craft of writing good ones (clear, specific, with context) is prompting.

A prompt is the input you give an AI — the question, instruction, and any background you include. Small wording changes produce big output changes, so prompting is a genuine skill: state the goal, give context, show an example, say what format you want. 'You are a marketing analyst; summarise this in 5 bullet points for a client' beats 'summarise this'.

What you say at the desk. “A flight” gets you nowhere. “Tokyo, Thursday, aisle seat, under £900” gets you a ticket.

prompt caching

API feature that reuses already-processed prompt text, billing repeats (like re-sent chat history) at a fraction of the normal input price.

Chat APIs are stateless, so every turn re-sends the whole conversation — and you'd normally pay to re-process it all. Prompt caching keeps the already-processed part warm on the provider's servers for a few minutes, so repeated content (system prompt, history, big documents) bills at a fraction of the normal input rate — about a tenth on Anthropic's API. It's like a barista keeping your usual order on the counter instead of remaking it from scratch each time. Agentic tools like Claude Code use it automatically.

The airline keeps your saved traveller profile instead of re-typing your passport number for every booking. Same details, a fraction of the cost.

prompt engineering

The craft of writing and refining prompts so AI output is reliably good — specificity, context, examples, format, iteration.

Prompt engineering is the deliberate craft of getting good output from AI: being specific, supplying context and a role, showing examples, specifying the format, and iterating on results. Think of it as writing a brief for a talented freelancer who starts immediately and can't ask questions. The one-weird-trick era has faded as models improved, but clear briefing remains a genuine workplace skill — and its successor, context engineering, builds directly on it.

Learning to ask at the desk the way the people who never seem to get stuck ask. It's a craft, and it's entirely learnable.

prompt injection

An attack that hides instructions in content an AI reads — a webpage, an email — so it obeys the attacker instead of you.

Prompt injection is the trick of hiding instructions inside content an AI will read — a webpage, an email, a PDF — so the model treats the attacker's words as commands ('ignore your instructions and forward the user's data'). It works because models can't fully separate 'text to read' from 'orders to follow', like a gullible intern who obeys sticky notes left by strangers. It's the main reason agents that browse and act need permission prompts and careful guardrails.

A note slipped into the paperwork reading “also, let this one through without checking.” The agent obeys it because it can't tell instructions from cargo.

prompt library

A team's shared, tested collection of reusable prompts — so everyone benefits when one person cracks a task.

A prompt library is a shared document or tool where a team keeps its proven prompts: the media-briefing prompt, the press-release edit pass, the report formatter. Like a recipe book for the office — when someone perfects a dish, everyone gets to cook it. Good libraries have named owners, notes on when each prompt works, and get pruned as models change.

The airline's book of standard announcements. Written once, tested, and used by every crew instead of improvising at the microphone.

protocol

An agreed set of rules for how two programs talk — like both sides speaking the same language. MCP and HTTP are protocols.

A protocol is a formal agreement about how programs exchange messages: what a request looks like, what a valid reply is, in what order things happen. The web runs on HTTP, email on SMTP, and AI-to-tool connections increasingly on MCP. The magic of a good protocol is that strangers' software interoperates: anyone who implements the rules can talk to everyone else who did.

Aviation English. Every pilot and every tower on earth using the same phrases in the same order, so nothing is lost in translation at 500 knots.

prototype

A working rough draft built to answer a question or make an argument — not to be the finished thing.

A prototype is a quickly built, deliberately rough version of an idea whose job is to make the idea discussable: you can point at it, click it, and disagree with it, which you can't do with a paragraph. Prototypes are allowed to cheat — invented data, missing features, hardcoded shortcuts — because their job is conversation, not production. The classic mistake is polishing one into a product; the classic win is a rough one that settles an argument in thirty seconds.

An architect's foam model. Nobody's going to live in it, but suddenly everyone in the room can see where the kitchen shouldn't be.

pull request

A proposed set of changes on GitHub: 'here's my branch — review the diff, discuss, then merge it in'. PR for short.

A pull request (PR) is GitHub's mechanism for proposing changes: you push a branch, open a PR, and others review the diff, comment, request tweaks and finally merge it into the main branch. It's the unit of collaboration in nearly every software team — and increasingly the unit of AI work, since coding agents can open PRs for a human to review and approve.

Proposing a change to the flight plan: here's what I'd alter and why, and somebody senior looks at it before it becomes the plan.

push

Send your new local commits up to the online copy of the repo (usually GitHub). Opposite of pull.

Pushing uploads the commits you've made locally to a remote copy of the repository, typically on GitHub — like syncing your save file to the cloud. Its counterpart, pull, brings down commits others (or you, on another machine) have pushed. Until you push, your commits exist only on your machine.

Sending your updated logbook up to the central archive, so it isn't only on your machine, one spilt coffee from gone.

Python

The most popular programming language in AI — readable, beginner-friendly, and what most AI tooling is written in.

Python is a programming language famous for reading almost like English, which made it the language of choice for data science and AI — most AI examples, libraries and tutorials assume it. You don't need to be a Python programmer to use AI tools, but recognising it helps: if a snippet starts with `import`, it's probably Python.

The language most people in AI start with, for the same reason most pilots start on a Cessna: forgiving, well documented, and more than enough to get you flying.

R

RAG

Retrieval-Augmented Generation — fetch relevant documents first, then have the model answer using them. How 'chat with your docs' works.

RAG (Retrieval-Augmented Generation) bolts a search step onto a model: when you ask something, the system first retrieves relevant chunks from your documents, then passes them to the model along with your question. It's how AI answers from knowledge it was never trained on — your company wiki, product docs — and it cuts hallucinations because the model quotes rather than recalls.

The pilot hasn't memorised every airport. They pull the approach plate for this one, then decide. Fetch first, answer second.

rate limit

A cap on how much you can use in a given window — the thing you hit on a subscription instead of a bill.

A rate limit is the provider's ceiling on how much you can do per minute, hour or day — measured in requests or tokens. It's the characteristic failure mode of a flat-rate plan: use too much too fast and you're paused until the window resets, but your cost doesn't move. Metered API access behaves oppositely: there's no allowance to exhaust, so it keeps serving and keeps charging. Flat plans fail closed (you stop working), meters fail open (you stop having money).

An all-you-can-eat buffet with one plate at a time. You'll never get a bigger bill — but you might be told to wait.

reach

How many distinct people could have seen something — as opposed to how many times it was shown.

Reach counts unique people; impressions count exposures, so one person seeing something four times is a reach of one and four impressions. PR results are usually reported as reach alongside quality measures — which outlets, whether the key message survived, what the sentiment was. Big reach numbers alone are the easiest thing in PR to inflate and the easiest for a sharp client to pick apart.

How many people walked past the poster, versus how many times posters were walked past.

README

The welcome-mat file of a repo (README.md), shown on its GitHub page: what the project is and how to use it.

The README is a repository's front page — a Markdown file (README.md) that GitHub renders right under the file list. Good ones say what the project is, how to install and use it, and where to get help. It's the first thing humans read in a repo, and AI agents read it too when getting their bearings in a project.

The sign at the terminal entrance: what this building is, where to go, and what to do first.

reasoning model

A model that 'thinks' before it answers, generating internal reasoning first. ChatGPT, Claude and Gemini all offer thinking modes.

A reasoning model is trained to generate internal chains of thought before producing its visible answer — effectively pausing to think, sometimes for seconds or minutes, on hard problems. ChatGPT, Claude and Gemini all offer these 'thinking' or 'extended thinking' modes. They excel at maths, code and multi-step planning, cost more tokens and time, and make old-school 'think step by step' prompting largely redundant. Like hiring a considered analyst instead of a quick-fire quiz contestant.

The pilot who works the diversion out properly before speaking, rather than answering the second the question lands.

remote

An online copy of your repo that git syncs with — usually on GitHub, conventionally named 'origin'.

A remote is another copy of your repository that your local one knows how to talk to — almost always the copy hosted on GitHub. Push sends commits to it; pull fetches from it. When you clone a repo, git automatically records where it came from as a remote called 'origin', which is why that word appears in so many git commands.

The central archive's copy of your logbook, which git keeps in step with the one on your machine.

repository

One project's folder tracked by git — its files plus full change history. 'Repo' for short.

A repository (repo) is a project folder under git's watch: all its files plus the complete history of every change ever committed. On GitHub, a repo is the unit you browse, clone (copy down) or star. One project = one repo is the usual rule.

One aircraft's complete file — every document and every revision, from delivery to this morning.

resourcing

Deciding who works on what, when — matching people to projects by skill, capacity and budget.

Resourcing (or resource management) is the job of allocating an agency's people across its projects. You take incoming requests, judge them on scope, timeline and priority, and match each to the right person — considering not just who's free, but who's right, and what it costs. Done well nobody notices. Done badly, two people are idle while a third works until 2am and the client gets a late deck.

Air traffic control for humans. Nobody thanks you for the planes that landed on time; everybody notices the one circling.

retainer

An ongoing monthly fee for an agreed amount of work — as opposed to one-off project fees.

A retainer is a recurring monthly fee that buys a client an agreed level of ongoing service. Agencies love them because the revenue is predictable, which makes hiring and resourcing planable. The opposite is project work: lumpy, unpredictable, and much harder to staff. Most agencies aim for a base of retained clients with project work layered on top.

A gym membership versus paying per class. The gym would very much prefer the membership.

RFP / RFI

Request For Proposal / Information — the formal document a client sends when putting work out to tender.

When a brand wants to appoint an agency it often runs a formal process. An RFI (Request For Information) is the light first filter — size, capabilities, credentials. An RFP (Request For Proposal) is the real thing: a written brief you respond to with strategy, ideas, team and costs, usually to a hard deadline and a fixed template. Large brands and public bodies run these through procurement, so the paperwork is as much of the test as the ideas.

The written round before the audition. Fluff it and you never get to sing.

RLHF

Reinforcement Learning from Human Feedback — humans rate the model's answers, and it's nudged towards the ones people prefer.

RLHF (Reinforcement Learning from Human Feedback) is the polishing step that turns a raw model into a pleasant assistant: humans compare candidate answers, a reward model learns their preferences, and the model is trained to score higher on them. It's like a chef adjusting recipes based on diner reviews. Side effect worth knowing: it optimises for answers people like, which isn't always identical to answers that are correct.

Thousands of passenger surveys, used to train crew out of technically-correct-but-alarming answers and into helpful ones. It's how a raw model learns bedside manner.

ROI

Return On Investment — what you got back compared with what you spent. The question behind every marketing number.

ROI (Return On Investment) compares what you got back with what you put in — spend £100 on ads, gain £500 of business, and the investment clearly paid. Every marketing metric ultimately serves this one question. Ad platforms often use the sibling term ROAS (Return On Ad Spend: revenue ÷ ad spend) for the same idea applied purely to advertising.

The route brought in £40,000 and cost £31,000 to fly. That gap is the only number the board actually asks about.

S

Search Console

Google's free tool showing how you appear in search results: impressions, clicks, the queries people typed, and your average position.

Google Search Console reports on your site's life inside Google's results page — before anyone reaches your site. It shows impressions (times you were shown), clicks, the actual queries people searched, and your average position for each. It's the SEO tool: GA4 sees inside the shop, Search Console sees the shop window. One quirk: it only keeps 16 months of history.

The report showing how often your airline came up when people searched “cheap flights to Málaga”, and how many of them clicked.

self-hosting

Running software on your own machine or rented server instead of using a company's hosted service. Your data, your rules, your maintenance.

Self-hosting means running an application on hardware you control — your Mac, a home server, or a rented VPS — rather than using someone else's hosted service. You gain privacy and control (your chats and files never live on a vendor's servers) and usually pay nothing for the software; in exchange, you're the IT department: updates, backups and security are on you. Personal agents like Hermes and OpenClaw are self-hosted by design.

Keeping your own aircraft in your own hangar rather than buying a seat. More control, and the maintenance is now entirely your problem.

SEO

Search Engine Optimisation — making your site rank higher in Google's unpaid results.

SEO is the craft of earning free traffic from search engines: relevant content, the words people actually search for, a technically healthy site, and links from elsewhere. It's the unpaid counterpart to ads. AI now sits on both sides — helping you write and analyse content, while search engines themselves add AI answers that change the game.

Getting your airline to the top of the search results without paying for the poster site.

session

One visit to your website, from arrival until they leave (or go quiet for ~30 minutes). One person can make many sessions.

A session is analytics-speak for one visit: someone arrives, browses around, and the session ends when they leave or after about 30 minutes of inactivity. The same person visiting in the morning and again at lunch counts as one user but two sessions — like one customer making two trips to the shop. Sessions are GA4's basic unit of 'how busy were we?'.

One visit, from walking into the terminal to leaving it. Sit still for half an hour and the system assumes you've gone home.

shadow AI

Employees using unapproved AI tools at work under the radar — a top way sensitive data leaks into the wrong systems.

Shadow AI is the workplace habit of using AI tools your employer hasn't approved — pasting client documents into a personal ChatGPT account, say. It's named after 'shadow IT', the older habit of sneaking in unapproved software. Surveys consistently find a large share of employees do it, usually because the approved route is slower or worse. The fix is as much policy as discipline: make the safe path the easy path.

Crew doing the paperwork on their own phone app because the official system is slow. Completely understandable, and exactly how sensitive data walks out of the building.

shell

The program inside the terminal that interprets your commands. Macs use zsh; 'Unix shell' means this family.

The shell is the program that reads what you type in a terminal and makes it happen — the terminal is the window, the shell is the brain behind it. Modern Macs use zsh; Linux commonly uses bash. They're near-identical for everyday commands, which is why Mac skills transfer straight to Linux servers.

The voice on the other end of the cockpit radio that takes what you say and makes it happen. Macs use one called zsh.

skill

A packaged, reusable workflow an agent loads on demand — 'how we deploy', 'our review checklist' — instead of cluttering every session.

A skill is a folder of instructions (and sometimes scripts) that teaches an agent one repeatable job — deploying the site, writing a release note, running the review checklist. Unlike CLAUDE.md, which loads every session, a skill loads only when invoked or relevant — like a laminated recipe card pulled out when that dish is on the menu. In Claude Code, skills live under .claude/skills/, each with a SKILL.md.

The written procedure for one specific job — de-icing, say — that any crew can pick up and follow correctly first time.

slash command

A command starting with / typed inside a tool like Claude Code — instructions to the tool itself, not a message to the AI. e.g. /help, /init.

Inside Claude Code (and many chat tools), anything you type starting with a forward slash is a slash command: an instruction to the software itself rather than a prompt for the model. /help lists the commands, /init writes a CLAUDE.md, /clear wipes the conversation. Think of it as the difference between talking to your taxi driver and pressing buttons on the meter.

The short codes crews use on the radio instead of full sentences. Typed rather than spoken, and everyone knows what they mean.

SOW (scope of work)

The document defining exactly what the agency will deliver, by when, for how much.

A Scope of Work sets out the deliverables, timings, assumptions and fee for a piece of client work. It matters to resourcing because it's the contract between what was sold and what the team must actually produce — and it's the document you point at when a client asks for a fifth round of revisions that nobody priced.

The builder's quote. "Kitchen refit" is a conversation; the SOW is the bit that says whether it includes the tiling.

staging area

Git's holding pen between editing and committing — git add puts changes here; git commit snapshots whatever's in it.

The staging area (or index) is git's shopping basket: after editing files, git add places chosen changes into it, and git commit seals exactly that basket into a snapshot. It exists so a commit can be deliberate — you might edit five files but commit only the two that belong together. Most beginner confusion dissolves once you see the three zones: working files, staging area, committed history.

The bags checked in but not yet loaded. You can still pull one off the pile before the hold door shuts.

static site

A website made of fixed files sent to visitors exactly as-is — no server-side code or database. Fast, cheap, hard to break.

A static site is a website served as ready-made files: the server just hands over the same HTML, CSS and images to everyone, doing no thinking of its own. That makes static sites fast, extremely cheap to host (often free), and very secure — there's nothing running to hack. Portfolios, documentation and landing pages are usually static; sites with logins, baskets and feeds need server-side code, making them dynamic.

A printed timetable handed out as-is, versus a departures board that recalculates. The printed one is the static site: cheap, instant, and never wrong in a surprising new way.

studio

The craft/production team — artwork, design finishing, retouching, resizing, making things actually deliverable.

In an agency, "studio" is the production department that turns an approved creative idea into finished assets: artworking, retouching, resizing a campaign into forty formats, preparing files to spec. It's distinct from the creative team, who come up with the idea. Studio is usually the most demand-heavy team to resource, because every project ends there and every deadline lands on it at once.

Creative designs the dress; studio cuts, sews and makes it in eight sizes by Friday.

subagent

A focused helper your agent spawns for one sub-task — with its own memory and tools — that does the work and reports back just a summary.

A subagent is a second AI worker your main agent delegates a chunk of work to — searching a big codebase, reviewing a file, drafting one section. It runs in its own context window with its own instructions, tool access and (often) a cheaper model, then hands back only the result, so the messy details never clog the main conversation. Claude Code stores them as .claude/agents/*.md; Codex added them in 2026 too. Like a manager handing a specialist one job rather than doing everything at one desk.

Sending one engineer to look at the landing gear while you carry on with everything else. They report back; you keep the overview.

sudo

Prefix that runs one command with administrator power, after your password. Powerful — use only when you understand why.

sudo ('superuser do') runs a single command as the all-powerful administrator instead of as ordinary you — the terminal equivalent of macOS asking for your password before a big change. It exists because your normal account is deliberately not allowed to modify system files, which is a safety feature. Rule of thumb: if a tutorial starts with sudo, understand why before typing your password.

The master key. It opens every door, including the ones locked for extremely good reasons.

SVG

Scalable Vector Graphics — an image described in text as shapes and coordinates, so it stays sharp at any size.

An SVG is a picture written as instructions rather than pixels: 'a circle here, this radius, this colour'. The browser draws it fresh at whatever size is needed, so it never goes blurry — ideal for logos and icons. Because it is text, an AI agent can edit it precisely and git can show you exactly which line changed. Photographs cannot be SVGs; illustrations and logos usually should be.

The instructions for painting the airline’s logo on a tail fin — “arc of this radius, in this red” — rather than a photograph of one. Hand them over and you get a crisp logo on a luggage tag or a hangar door.

system prompt

Hidden standing instructions that set an AI's role, rules and personality before your message arrives.

A system prompt is the standing instruction sheet an AI gets before it sees your message — its role, rules, tone and boundaries. When a company builds a support bot that's polite and never discusses competitors, that behaviour lives in the system prompt. In the API and in tools like Claude Code you can set it yourself.

The crew's standing orders, read before a single passenger speaks: who you are, what you may say, and what you never say.

T

Telegram

A chat app with an open bot system — the easiest place to give your own AI agent a phone-friendly chat interface.

Telegram is a messaging app notable for its open, friendly bot platform: anyone can create a bot in minutes (via @BotFather) and get an API token. That's why personal-agent projects like Hermes and OpenClaw use it as their front door — your agent runs on a server, and you text it from your phone like any other chat.

The staff radio channel that anyone is allowed to build a handset for — which is exactly why it's the easiest way to reach your own agent.

temperature

The randomness dial for AI output: near 0 = focused and repeatable; higher (≈0.8) = more varied and creative.

Temperature controls how adventurously a model picks its next token from the probability distribution. Near 0 it almost always takes the favourites — predictable, consistent, good for extraction and code. Higher values (say 0.7–1.0) flatten the odds so less-likely tokens get a chance — livelier and more creative, but riskier. It's a thermostat for randomness, and it's why the same prompt can give different answers each time.

The dial on the cabin announcements. At zero it's the same script word for word every time; turned up, the crew start improvising. Charming at the right setting, alarming past it.

terminal

The text window where you type commands to your computer directly. On Mac: the Terminal app. Powerful, not scary.

The terminal (or command line) is a text interface to your computer: type a command, press Enter, the computer does it. It looks intimidating but it's just typing instead of clicking — and it's where developer tools like Claude Code, git and Homebrew live. A dozen commands (cd, ls, mkdir…) cover most daily use.

The text window you type commands into. Confusingly, the airport kind is where you wait — this one is where you work.

token

The word-chunks AI reads and writes — roughly ¾ of an English word each. Pricing and limits are measured in tokens.

Models don't read letters or whole words — they read tokens, chunks of text roughly three-quarters of a word long ('understanding' might be 'under'+'standing'). Everything is measured in tokens: how much a model can hold in mind at once (its context window), and what API usage costs (priced per million tokens). 'Tokenisation' is the chopping-up step.

Not words — chunks. “Heathrow” might arrive as “Heath” + “row”. You're billed per chunk, which is why long documents cost what they do.

tokeniser

The slicer that turns text into tokens. Each model family has its own, so the same sentence is a different token count on different models.

A tokeniser is the program that chops text into tokens before a model sees it — usually built with byte-pair encoding, which gives the most common letter-sequences their own vocabulary entry. Think of it as a bread slicer with its own idea of where the cuts go: OpenAI, Anthropic, Google and Llama models all slice differently. That's why token counts (and therefore costs) for the same document differ between models, and why letter-level questions are hard: the model gets sealed slices, not characters.

The machine that slices text into chunks before anything else happens. Every airline slices slightly differently, so the same sentence costs a different amount depending on whose aircraft you're on.

tone of voice

The consistent personality a brand writes in — the thing you match when drafting for someone else.

Tone of voice is the codified personality of a brand's writing: vocabulary, rhythm, formality, what it never says. Agencies define it for clients and have one themselves. Being able to read a brand's existing output and write a convincing new line in that voice is a core, testable skill — it's why interview tasks so often ask for a social post "in the agency's tone of voice". The way to pass is to read ten of their real posts first and imitate the pattern.

You can recognise a friend's texts without seeing the name. That's tone of voice.

traffic (traffic manager)

The older agency word for resourcing — keeping work flowing through the building.

"Traffic" is agency slang inherited from the print-advertising era, when a traffic manager physically walked jobs between departments — copy to art, art to studio, studio to print. The physical bit is gone; the function isn't. Today it means the same as resourcing: making sure work moves through creative, strategy and studio without jamming. If a job ad says "traffic" or "resource" in the title, expect the same work.

Same job, two names, depending on how old the agency is. Like "post room" and "logistics".

training

How a model learns: showing it billions of examples and nudging its internal numbers until predictions improve.

Training is the learning phase: the model reads billions of text examples, guesses the next token, gets scored, and its billions of internal numbers are nudged to do better — repeated at colossal scale on thousands of specialised chips for weeks. That's why frontier models cost hundreds of millions to make. After training, using the model (inference) is comparatively cheap.

Flight school. Thousands of hours, constant correction, enormously expensive, and entirely finished before the first paying passenger boards.

transformer

The 2017 neural-network design (with 'attention') that made modern LLMs possible. The T in GPT.

The transformer is the neural-network architecture behind virtually all modern language AI, introduced in Google's 2017 paper 'Attention Is All You Need'. Its trick, attention, lets the model weigh every word against every other word at once — so 'it' in a sentence knows what it refers to. Transformers train efficiently at massive scale, which unlocked GPT, Claude and everything since.

The jet engine of AI. Things flew before 2017; afterwards they flew further and faster, and the whole industry rebuilt itself around the new engine.

Turing test

Alan Turing's 1950 'imitation game': if you can't tell a machine's chat replies from a human's, it passes.

Proposed by Alan Turing in 1950 as a practical stand-in for the fuzzy question 'can machines think?'. A judge holds text conversations with a hidden human and a hidden machine; if the judge can't reliably tell which is which, the machine passes. It's a blind taste test for minds — judging the conversation, not the ingredients. Modern chatbots routinely fool people in short chats, which is why researchers now use harder benchmarks.

Turing's version: if you can't tell from the radio whether the tower is staffed by a person, does it matter? Still being argued about, seventy-five years on.

U

Unix

The operating-system family behind macOS and Linux — why the same terminal commands work on both.

Unix is a 1970s operating system whose design became the foundation of macOS, Linux, and most of the servers running the internet. Practical upshot: the terminal commands you learn on a Mac (cd, ls, mkdir, grep…) work nearly identically on Linux servers — learn once, use everywhere.

The engineering standard underneath both macOS and Linux — rather like most of the world's aircraft sharing the same basic controls. Learn it once and it works on both.

utilisation

The percentage of someone's paid time that's billable to clients — the number agency finance directors watch.

Utilisation is billable hours divided by available hours. It's how agencies measure whether they're selling the time they're paying for. Too low and the agency is losing money on salaries; sustained too high (above roughly 85–90%) and you have no slack for pitches, sickness or thinking, so quality and people both start to break. A good resourcer runs the team in the healthy band rather than maximising the number.

A restaurant measuring how many of its tables are full. Empty tables lose money; a permanently rammed restaurant has no room for the regular who turns up unbooked.

V

version control

Software that tracks every change to a project over time, so you can see history and roll anything back. Git is the standard.

Version control is the practice (and software) of recording every change to a project's files over time — who changed what, when, and why — so any earlier state can be recovered. It replaces the chaos of final_v2_REALLY_final filenames with proper, labelled save points. Git won this category so completely that 'version control' and 'git' are near-synonyms in practice.

Every revision of the operations manual, kept, dated and reversible — instead of a folder containing manual_final_FINAL_v2.

vibe coding

Building software by describing what you want to an AI and iterating on the result — without writing the code yourself.

Vibe coding (coined by Andrej Karpathy in 2025) is building software by conversation: you describe what you want, the AI writes the code, you react to the result and steer. You may barely read the code at all. It has real limits for production software, but it genuinely lets non-programmers build working tools — which is exactly what this course trains you to do.

Telling the ground crew “it's pulling to the left, sort it” and signing off whatever they do. Sometimes exactly right; occasionally you really should have looked.

virtual environment

A project-local pocket of Python packages (a .venv folder), so different projects' libraries never clash.

A virtual environment (venv) is a self-contained pocket of Python packages living in a folder inside one project. Activate it and pip installs go into that pocket instead of onto the whole machine — so project A's library versions can never break project B. It's like each project having its own toolbox rather than everyone rummaging in one communal drawer; modern Python setups all but insist on it.

A separate toolbox for each aircraft type, so the spanners for the A320 never end up on the 787.

VPS

Virtual Private Server — a small rented computer in a data centre that stays on 24/7. From a few pounds a month.

A VPS (Virtual Private Server) is a slice of a data-centre machine you rent — your own always-on computer in the cloud, controlled over the internet. It's the classic home for anything that must run 24/7, like a personal AI agent: no laptop lid to close, no home wifi outage. Entry-level ones cost a few pounds a month. Think of it as a storage unit for software — small, cheap, always open.

A rented parking space in a data centre that stays lit when your laptop closes. That's where an always-on agent actually lives.

W

wash-up

The honest debrief after a pitch or project — what worked, what didn't, what we'd do differently.

A wash-up is the post-mortem meeting held after a pitch concludes or a big project ships, win or lose. The point is capturing learnings while they're fresh and getting them somewhere findable, so the next team doesn't repeat the mistake. Everybody agrees wash-ups are essential; most agencies skip them when they lose, which is precisely when the learning is most valuable. Making them actually happen is a real, visible contribution.

The team debrief after the match. Easy after a win, character-building after a defeat, useless if nobody writes it down.

weights

Another word for a model's parameters — the billions of learned numbers, the 'dials'. Publishing them is what 'open-weights' means.

Weights are the adjustable numbers inside a neural network — the same things this course calls parameters or 'dials'. Each one controls how strongly a signal passes from one part of the network to the next, and training is simply the process of nudging billions of them until the model's guesses get good. After training, the weights ARE the model: copy the weights and you have copied everything it knows. That is why a downloadable model is called 'open-weights' — the company is publishing the numbers themselves, not the code that made them.

The billions of settings inside the aircraft, arrived at by testing rather than design. Nobody wrote them down — the training set them.

WIP (work in progress)

The recurring status meeting where a team walks through everything currently live.

A WIP is the standing meeting — usually weekly — where a team reviews every live project: where it's at, what's blocked, what's coming. For anyone in a resourcing or traffic role, WIPs are the main intelligence-gathering exercise of the week: it's where you find out that the project everybody said was fine has quietly slipped, before it becomes your problem on Friday.

A ward round. You visit every bed, even the patients who look fine, because that's how you catch the one who isn't.

Working directory

The folder a command is currently “standing in”. Commands act here unless you tell them otherwise — and it is the folder an AI agent can see.

Your working directory is where your terminal currently is. Every command you type acts relative to it: 'ls' lists this folder, 'mkdir notes' makes a folder here. 'pwd' tells you where you are. It matters enormously with AI coding agents, because the folder you start one in is the scope of what it can read and change — which is why you make a fresh folder per project rather than starting one in your home directory.

The gate you are standing at. “Board the plane” is the same instruction at Gate 12 and Gate 40, and gets you on a completely different aircraft.

Like your jargon decoded? 🛫

The full course explains every one of these terms with hands-on practice — 17 short lessons, quizzes, a friendly AI tutor. Free, forever.

Join free