Prompting: Getting Good Results

A generative model turns your words into more words by predicting what is most likely to come next, which makes the prompt — the instruction you write — the biggest lever you control over the result. A vague prompt invites a vague, generic answer, because "most likely" is what a next-token predictor drifts toward when nothing pins it down. A specific prompt, with context and a clear target, narrows the output toward something useful. None of this requires knowing how the model works inside; it requires saying clearly what you want, to whom, and in what shape — and adjusting when the first attempt misses. Prompting well is fast becoming an everyday skill, and small changes in how you ask produce large changes in what you get back.

What a prompt actually is

A prompt is everything you hand the model as input: your instruction, plus any background, examples, and constraints. The model has no memory of you beyond the current conversation — it does not know who you are or what you meant but did not say. Whatever context matters must be in the prompt itself.

Be specific, and set format and audience

"Write about climate change" and "Write a 120-word paragraph explaining carbon pricing to a small-business owner, in plain language" pull from very different regions of what the model considers likely. Naming the length, format, audience, and tone removes guesswork and steers the output toward the version you want.

Give context and examples

Showing the model one or two examples of the style you want steers output more reliably than describing that style in the abstract. Background it could not otherwise know — your goal, the reader, the constraints — belongs up front.

Break big asks into steps, then iterate

A request such as "plan and write my report" produces a shallow, everything-at-once answer. Asking for an outline first, then each section, gives you control and better material. And because the system is probabilistic, your first prompt is a draft: read what comes back, see where it drifted, and refine. This is steering, not programming.

Suppose you want help with a cover letter. Weak prompt: "Write me a cover letter." Given almost nothing, the model produces a bland, one-size template. Stronger: "I'm applying for a junior data-analyst role at a mid-size retailer. Here are three facts about my experience [...]. Write a 250-word letter, confident but not boastful, to a hiring manager, emphasising the retail and spreadsheet experience." The second supplies audience, length, tone, and material only you possess — so the output is specific to you, not because the model learned about you, but because you put yourself in the prompt.

More words are not automatically a better prompt. A prompt padded with irrelevant detail, conflicting instructions, and five questions at once can produce a worse answer than a short, sharp one: the model tries to satisfy everything and satisfies nothing well. Specificity means relevant precision, not volume. When a prompt rambles, the fix is usually to cut it to the single clear task, not to add more.

A well-documented finding in AI research is that asking a model to "work through it step by step" can improve its answers on reasoning-style tasks, compared with demanding the final answer immediately. Researchers described this "chain-of-thought" prompting around 2022, and the broad pattern has held across many studies since. The likely reason fits everything above: intermediate steps give the model more relevant context to condition its next tokens on. The effect is not universal — it helps most on multi-step problems — and a convincing-looking explanation can still reach a wrong answer. Treat it as a useful technique, not a guarantee of correctness.

• "The model remembers me between chats." It does not; each new conversation starts blank unless a tool explicitly feeds in history.
• "A longer prompt is always better." Relevance beats length; padding can actively hurt.
• "There is one correct prompt." Prompting is iterative; the same goal has many good phrasings, found by refining.
• "Prompting is programming." You are nudging probabilities, not issuing commands that must execute exactly.

Take one weak, vague prompt and rewrite it three times, each time adding a single thing: the audience, the format, and a worked example. Compare how the likely output changes at each step.

Think Like a Maester: A model answers the prompt in front of it, not the one in your head. If the output is vague, the prompt was vague — fix the instruction before blaming the machine.

1. What is a prompt, and what does the model use it for?
2. Why does naming the audience and format improve results?
3. Why must relevant context go inside the prompt rather than being assumed?
4. What does it mean to say prompting is iterative, not one-shot?
5. Why can a longer, more detailed prompt sometimes produce a worse answer?

A prompt is the instruction and context you give a generative model, and because that model predicts likely next tokens rather than looking things up or remembering you, the prompt is your main means of steering it. Being specific, supplying context and examples, stating format and audience, and breaking large tasks into steps all narrow the model toward the answer you want. Because the system is probabilistic, the first attempt is a draft to refine, not a program to run. Prompting well is a learnable, tool-agnostic skill: say clearly what you want, to whom, and in what shape — then iterate.

• How Large Language Models Work
• Strengths and Limits of Generative AI