The toolbox

Most of these tools are built on the same basic idea — a model that learned patterns from a huge pile of examples. But what they learned from is different. A chat model learned from text. An image model learned from pictures and their captions. A music model learned from audio. So each one is fluent in one medium and clumsy in the others.

So you don’t need to memorise tool names. You just need to ask yourself one thing: what am I actually trying to make? Answer that, and the right tool is usually obvious.

• Chat and reasoning — For thinking, writing, explaining, planning, working through a problem. ChatGPT, Claude, Gemini, Grok. This is your go-to for almost anything made of words or ideas.
• Images — For making a picture from a description. Midjourney (painterly, strong taste), Google’s Gemini image tools, Ideogram (good at text inside images), Higgsfield. Reach for these for posters, characters, concept art, illustrations.
• Video — For generating moving footage from a prompt or a still image. Sora, Google Veo, Runway, Kling, Higgsfield. Use them for short clips, animation, visual ideas — not feature films yet.
• Music — For writing and producing a full song, lyrics and all. Suno. Hum an idea, describe a mood, get a track back.
• Voice — For turning text into spoken audio, or cloning a voice. ElevenLabs is the one most people reach for. Useful for narration, characters, audiobooks.
• Building apps and sites — For making real, working software — what people call “vibe coding.” Lovable, v0, Replit for whole apps and sites; Cursor and Claude Code for working directly in real code. This is the one that surprises people most. You can actually build things.

First, what “learned from” actually means here. The model doesn't keep a library of stored pictures and stitch bits of them together. It studied billions of images and learned patterns — what watercolour does, how light falls on a face, what makes a picture read as “moody” — and it generates from those patterns, not from anyone's actual file.

That's exactly why the argument is hard. Both sides have a real point.

• One side — “It learns the way a human art student learns — by looking at a lot of work and absorbing how it's done. We've never called that stealing when a person does it.”
• The other — “A human student doesn't produce ten thousand images an hour in my exact style, forever, without paying me or asking. Scale changes what a thing is.”
• Where it stands — courts and parliaments in several countries are still arguing. It genuinely isn't settled — anyone who tells you the answer is obvious is skipping the hard part.

Practically, things are shifting. Some models now train on licensed data the makers paid for. Some companies let artists opt out of training — though opting out of something that already happened is part of what artists are angry about. The sharpest edge is the “in the style of” prompt: type a living artist's name and get their look, built from their life's work, for free. That's the case where the taking is hardest to wave away.

None of which means don't use the tools. It means use them knowing where the magic came from. If a style you generated leans on a real person's work, credit it. And don't pass a borrowed style off as your own taste — the people who can tell are exactly the people whose opinion you'd want.

Training is the same trick run in reverse. Take billions of captioned pictures and add noise to them, a little at a time, until each one has dissolved into static — and make the model learn, at every step, how to undo the damage. Do that enough times and it gets good at un-noising in general.

At generation time, you hand it fresh static plus your words, and it runs the reversal: step by step, “what would this look like slightly less noisy, given the caption?” That's a diffusion model. Video models are the same idea with time attached — instead of denoising one frame, they denoise a whole stack of frames at once so the motion hangs together.

You don't have to take this on trust — there's a party trick on this very site. Open the ✦ menu and pick “diffusion,” and you'll watch a picture resolve out of static right in front of you. That's not a metaphor for the process. That is the process.

It also explains the famous tells in AI images — the six-fingered hands, the signs full of almost-letters. The model isn't building a hand, bones and joints and all. It's matching patterns of what hands tend to look like, and “looks right” and “is right” are different tests. Most of the time the pattern holds. At the fiddly edges — hands, text, anything with strict rules — it sometimes doesn't.

• An answer, an essay, a plan, a study partner — A chat model.
• A picture, a poster, a character design — An image tool.
• A short video or animation — A video tool.
• A song — A music tool.
• A narrator or a spoken voice — A voice tool.
• A website, an app, a game — A vibe-coding tool.

The fun part: these combine. Write a story with a chat model, illustrate it with an image tool, narrate it with a voice tool, score it with a music tool, and wrap the whole thing in a website you built by vibe coding. One project, five tools, each doing the bit it’s best at. That’s how a lot of real things get made now.

• A chatbot — You type, it replies, you type again. A conversation. It waits for you at every step. This is what you already know.
• An agent — You give it a goal, and it takes several steps on its own to reach it — searching, writing files, running tools, checking its own work — without asking you between each one. A chatbot answers; an agent goes and does.
• An API — A way for one program to talk to the model directly, no chat window involved. It’s how a model gets built into other software — a game that talks back, an app that summarises your notes. The plumbing behind the apps you use.
• A CLI — A command-line interface — you drive the tool by typing commands into a plain text terminal instead of clicking buttons. It looks old-fashioned. It’s actually fast, precise, and where a lot of serious building happens. Claude Code is one.

The same brain sits behind all four. A chatbot and an agent might run on the identical model; the difference is the harness around it — how much rope it’s given to act on its own. As you go from chatbot to agent, you hand over a bit of control and get a lot more done without babysitting every step. That trade-off — less hands-on, more power — is pretty much where all of this is going.

• Skills — A skill is a saved set of instructions and know-how you can hand a model so it does a specific job well every time — like giving it a recipe card it can pull out when the job comes up, instead of explaining from scratch each time.
• MCP — Model Context Protocol. A shared standard that lets a model plug into outside tools and sources of information — your calendar, a database, a website — without someone hand-building a connection for each one. Think of it as a universal socket. Before it, every connection was custom wiring; with it, tools just plug in.

There’s also the question of which model to use, and bigger isn’t always better. The biggest models are the smartest but slower and more expensive. Smaller models are faster and cheaper and often good enough.

Matching the model to the job is a real skill. You don’t send the genius to fetch the milk. For a quick rewrite or a simple sort, a small fast model is the right call; save the heavyweight for the problem that actually needs it. People who build with these tools think about this constantly — it’s the difference between a tool that feels instant and one that crawls.