Driving it

There are no magic incantations. What works is what would work on any capable stranger: say what you want, what it’s for, who it’s for, and what “good” looks like. Vague in, vague out.

• The task — what exactly do you want it to do — write, list, explain, fix, compare?
• The purpose — what’s it for? An answer for a five-year-old and an answer for an exam are not the same answer.
• The shape — how should the result look — a paragraph, ten bullet points, a table, a single sentence?

Why does phrasing matter so much, given it’s “just predicting words”? Because your prompt sets the scene it predicts inside of. A precise, serious request makes precise, serious answers the most likely continuation. A lazy request makes a generic answer the most likely one. You’re not casting a spell — you’re choosing which neighbourhood of its training it wanders into.

Specificity isn’t about length — it’s about removing guesses. Every detail you leave out, the model fills in for you, usually with the blandest average it can find. Pin down the format, the length, the tone, the audience, and the constraints.

And if you can show one example of what you want, do. A single example of the style or layout you’re after teaches it more than a paragraph of description.

• Weak — “Help me with my essay.”
• Strong — “Here’s my essay’s opening paragraph. Suggest three ways to make the first sentence grab attention, keeping my voice. Don’t rewrite the whole thing.”
• Show, don’t just tell — paste an example: “Match this style: [your sentence].” Examples are instructions the model can copy directly.

Giving examples has a proper name: few-shot prompting. Show it two or three worked examples of input-and-output you like, and it’ll infer the pattern and continue it — often far more reliably than any description. Zero examples is “zero-shot.” When something keeps coming out wrong, the fix is usually one good example, not ten more adjectives.

Whatever it needs to do the job well, paste it in: the actual text, the real numbers, the rubric your teacher gave you, the message you’re replying to. Don’t make it guess what it could simply read.

The model’s short-term memory is the conversation itself — what’s often called the “context window.” Things scroll out of reach in very long chats, and a fresh chat starts blank. It’s a desk, not a diary.

• Give it the source — “Summarise this article: [paste].” beats “What did that article say?” — it never saw the article.
• Give it the rules — marking criteria, word limits, banned words, the format your teacher wants.
• Start fresh when you switch tasks — leftover context from an unrelated chat can quietly steer the answer.

The context window has a real size limit, measured in “tokens” (chunks of words). Stuff in too much and the oldest parts fall off the edge — the model isn’t ignoring you, it genuinely can’t see them anymore. There’s also a subtler effect: models pay most attention to the very start and the very end of what you give them, and can go a bit hazy in the middle. So put your most important instruction first or last, not buried in paragraph six.

A role changes how it talks, how deep it goes, and how blunt it is. Ask it to be a patient tutor and you get one kind of answer. Ask for a blunt editor and you get a completely different one. It’s not really becoming a person — it’s just leaning on a different slice of what it learned.

Constraints are just as powerful, and underused. Fences tell it what not to do, which is often the whole point.

• Role — “Act as a careful editor who points out weak arguments, not typos.”
• Limits — “In under 100 words.” “Use no jargon.” “Only use the text I pasted — don’t add outside facts.”
• Behaviour — “If you’re not sure, say so.” “Ask me a question before you answer if anything is unclear.”

A role prompt doesn’t install knowledge it didn’t already have — telling it “you are a doctor” doesn’t make it a doctor. What it does is shift the style and the slice of training it leans on, which usually makes the answer more focused and consistent. Useful, but don’t mistake a convincing persona for real expertise underneath. It’ll sound like a doctor long before it’s as reliable as one.

Don’t start over with a brand-new prompt when something’s nearly right. Stay in the conversation and adjust — it remembers what it just wrote, so you can edit by talking.

Be specific about what’s off, the same way you’d be specific in the first place. There’s even a dial for how much it gambles on unlikely words — it’s called temperature, and you can play with it in the lab below.

• Redirect — “Too formal. Make it sound like I’m talking to a friend.”
• Zoom in — “The second point is the strongest — expand just that, drop the rest.”
• Challenge it — “Are you sure? What’s the strongest argument against what you just said?”
• Restart when stuck — if it’s gone in circles, a clean chat with a better-built first prompt beats fighting a derailed one.

Pushing back works partly because asking it to reconsider, critique itself, or “think step by step” forces it to generate reasoning before its conclusion — and reasoning it has written out becomes context that shapes a better final answer. But watch for a trap: these models are people-pleasers. Say “are you sure?” and they’ll often cave and “correct” a perfectly good answer, just because the question sounds like you wanted a change. Push back to test the logic, not to bully it into agreeing with you.

• Prompt injection — an AI can’t reliably tell your instructions apart from instructions hidden in the text it’s reading. So a web page or document can carry a sneaky line like “ignore your previous instructions and…” — and the model may obey it. This is a genuine, unsolved security problem, not a party trick.
• Building a knowledge base — instead of hoping the model already knows your stuff, you hand it the right documents at the right moment — your notes, a manual, a set of facts — so it answers from those rather than from memory.

That second idea has a name: retrieval-augmented generation, or RAG. Roughly: you keep a pile of documents; when you ask a question, a search step grabs the few most relevant chunks and pastes them into the context window automatically; then the model answers using them. It’s how a chatbot can “know” your school’s rules or a company’s products without ever being retrained.

The catch ties back to prompt injection: once you’re auto-feeding outside documents into the context, a poisoned document can smuggle in instructions too. The wider craft — deciding what goes into the context window, in what order, and where it came from — is now called context engineering, and it’s most of the real work behind a good AI tool.