How to Use LLMs: A Beginner's Guide

A large language model is trained on an enormous amount of text to do one thing: predict what comes next. That single ability, at enough scale, produces summarising, translating, explaining, coding, reasoning and conversation as side effects. It is not a database and it is not a search engine — it has no list of facts it looks things up in.

Within one conversation, everything you type is in context and the model can use it. Between conversations, it remembers nothing unless the product explicitly stores it. So when it 'forgets' something important from yesterday, that isn't a bug you can prompt your way around — you have to re-supply the context.

• Open with the context, not the question. Give the situation before the ask.
• If the answer is wrong about your setup, restate the setup — it genuinely doesn't have it.
• In a long chat, earlier turns can fall out of the effective context. Re-state the key constraint rather than saying 'as I mentioned'.

Everything the model can 'see' at once is the context window: your messages, any attached files, and its replies. It is finite, and quality degrades near the edges — most noticeably in the middle, where details get lost in a long document.

• Put the most important instruction at the end. Attention is strongest there.
• For long documents, work in sections rather than pasting everything at once.
• Start fresh when the topic changes. A clean context beats a long messy one.
• For repeated work, keep a short reusable preamble with your role, constraints and preferences.

• Transforming text: reformat, shorten, expand, change tone, translate, restructure.
• Explaining something at whatever level you need, including analogies you wouldn't have thought of.
• Drafting from an outline you provide.
• Code that follows a pattern you show it, especially when you include the types and the test cases.
• Brainstorming breadth — many options fast, which you then filter yourself.
• Being a tireless sparring partner for arguing, practising or rehearsing.

• Facts about anything after their training data, and anything specific to your organisation.
• Arithmetic and precise counting without a tool attached.
• Knowing what they don't know. Confidence is uniform, whether right or wrong.
• Consistency over a long piece without being told to check its own earlier claims.
• Anything where you need it to have seen your private data — it hasn't.

• For drafting, rewriting and everyday work, a mid-tier model is the sweet spot.
• For hard reasoning, long documents and tricky maths, use the strongest model available and budget more time to verify.
• For current events, use a model with search or retrieval attached. Recall alone will invent.
• For volume, speed or privacy, a smaller or local model is usually the right trade.

The practical advice: stop reading model comparisons and pick two — one fast everyday one, one strong one — and learn those well. Switching constantly costs more than any quality difference.