How to get far more out of an LLM
The model finishes your sentence. Learn to start it well.
A language model never answers your question; it only continues your text, and the whole craft of prompting is learning to begin the continuation well.
Ask a model to "write something about our product" and you will get something about a product. Ask it to act as a skeptical enterprise buyer, evaluate a 60-word pitch against three named objections, and reply only in a table, and you get work. The difference isn't intelligence. It's that the second request told the model who to be, what to aim at, and exactly what shape the answer should take. Vagueness in, vagueness out.
Start with role, goal, constraints
Three moves do most of the heavy lifting. Give it a role ("you are a careful copy editor") so it adopts the right voice and standards. Give it a goal: the outcome you want, not just the topic. And give it the constraints that matter: length, audience, format, what to avoid, what must be true. Constraints feel limiting; they are actually the most information you can hand the model in the fewest words. Then show, don't just tell: one or two examples of the exact format you want (this is called few-shot) is worth more than a paragraph describing it.
Two techniques quietly double the quality. First, ask it to work step by step and show its reasoning before the final answer, because a model that commits to a conclusion in its first token can't walk it back, but one that reasons aloud catches its own errors on the way. Second, invite it to ask YOU clarifying questions first. A good model will surface the three things it's guessing about, and answering them up front beats correcting a confident wrong draft afterward.
The model is fluent even when it is wrong. Confidence is its default tone, not a signal of truth.
Three truths the machine won't tell you
It has no memory between chats and cannot look anything up unless you let it. Every new conversation is a stranger who has read most of the internet but knows nothing about you, your files, or today. So paste in the context it needs, and when the task requires live facts or real action, give it a tool: search, a database, a calculator, code. That is what the widget above is really about: tools and latitude change not just the answer but what the model is capable of doing at all.
And it cannot truly do arithmetic. It predicts what a correct calculation looks like, which is not the same as computing it, so it will state 7,431 × 982 with total confidence and be wrong. Hand it a calculator or ask it to write code, and the same model becomes reliable. Two more habits seal it: lower the temperature (the randomness dial) when you need consistent, repeatable output, and always iterate: the first answer is a draft. Tell it what's off, and the second pass is usually the one you wanted.