Practise the work. Build your portfolio.

Understand the idea

A chat assistant is built on a large language model (LLM). It produces likely text from the words you give it and from what it learned during training. Some versions are tuned to “reason” — they work through a problem in steps before answering, which usually helps with maths, code and multi-step planning but takes longer. Many assistants also accept images, PDFs or audio (“multimodal” input), depending on your version and plan.

Everything the model can see at once — your prompt, the pasted document, the earlier messages — sits in its context window. When a conversation grows past that limit the oldest parts drop away, so the tool can forget an instruction you gave earlier. Training also has a knowledge cutoff: the model learned from material up to a certain date and knows nothing after it unless you tell it or it looks it up.

Tools change what is possible. With web search, a file connector or a calculator, an assistant can fetch something real and cite it. Without them it can only recall. An agent goes further: it plans and carries out steps, using tools on your behalf. An automation runs a fixed set of steps on a schedule. Agents and automations act, so anything costly or hard to undo needs a person's approval.

A model writes fluently whether or not it has the facts. A confident, well-written wrong answer is called a hallucination, and it looks exactly like a right one. So judge an answer by its source, not its tone: ask which tool was used, ask for the link or the file, and check every number yourself.