This is part of our beginner’s guide to using AI. If you’ve tried ChatGPT or Claude and come away underwhelmed, the tool usually isn’t the problem, the prompt is. A vague request gets a generic answer; a specific one gets something you can actually use. Here’s what actually changes the output.
Give it a role and an audience
Telling an assistant who it’s writing as and who it’s writing for narrows its output dramatically. “Explain compound interest” gets a generic, textbook-style answer. “Explain compound interest to a teenager who’s never had a bank account, using a real example” gets something specific, appropriately simple, and concrete. The role and audience do more work than almost any other single addition to a prompt.
Describe the goal, not just the task
“Write a product description” is a task. “Write a product description that convinces someone comparing three similar products to pick this one, focusing on the two things that actually differentiate it” is a goal. The second version gives the assistant a way to judge its own output against something specific, rather than producing generically competent copy that could describe almost anything.
Set the format up front
If you need a specific structure, a table, a numbered list, a short paragraph under 100 words, say so in the initial prompt rather than asking for a rewrite afterward. Both OpenAI’s own prompt engineering guidance and most practitioners agree on this: specifying format, length, and structure up front saves a full round of back-and-forth compared to describing content first and formatting second.
Iterate instead of starting over
The most underused technique: treating the first response as a draft to refine rather than a final answer to accept or discard. “Make the second paragraph more specific,” “cut this by half,” “make the tone less formal”, each of these builds on what already worked instead of throwing away a mostly-good response over one wrong detail. This is the single biggest difference between people who find these tools frustrating and people who find them genuinely useful.
Give it examples when the format matters
If you have a specific style or format in mind, an existing piece of writing you want matched, an email template your team already uses, pasting an example and asking the assistant to match that style works far better than describing the style in the abstract. This technique, often called few-shot prompting, is one of the most reliable ways to get output that not only meets the objective but also fits the format precisely, and it’s covered in more depth in Anthropic’s prompt engineering documentation.
Common prompting mistakes
Over-specifying every sentence is a common overcorrection: telling the assistant exactly how to phrase each line defeats the purpose of using it in the first place. The better balance is being specific about the goal, audience, and constraints, then letting the assistant handle the actual sentence-level writing.
Another common mistake is assuming a single, perfectly worded prompt should produce a finished result. Even experienced users iterate, the first response is a starting point, and treating it as one rather than expecting perfection on the first attempt removes most of the frustration people report with these tools.
Finally, forgetting the assistant has no memory of context you haven’t provided is a frequent source of bad results. It doesn’t know your company’s specific product unless you describe it, doesn’t know your personal writing style unless you show it, and doesn’t know constraints (a word limit, a required section) unless you state them. Treat every new conversation as starting from zero unless the tool’s memory feature is explicitly enabled and you’ve confirmed it retained what you expect.
A simple template that covers most tasks
For most everyday writing or analysis tasks, this structure covers the essentials: state the role and audience, describe the specific goal, mention any format or length constraint, and provide one example if format matters. It doesn’t need to be elaborate; even a two-sentence prompt following this pattern outperforms a single vague sentence most of the time.
Breaking a complex task into steps
For anything genuinely complex, a research summary with multiple sections, an analysis with several parts, asking for the entire thing in one shot often produces a shallower result than breaking it into stages. Ask for an outline first, review and adjust it, then ask for each section to be filled in against that approved outline. This mirrors how a competent human would tackle the same complex task, and it gives you a checkpoint to redirect before a lot of effort goes into the wrong direction.
Using constraints to improve quality, not just length
A word or length limit is often treated purely as a formatting requirement, but it’s also a genuine quality lever. Asking for “the three most important points” forces prioritization in a way that “list the important points” doesn’t, and the result is usually sharper, not just shorter. The same applies to asking an assistant to argue a specific position, or to explicitly rule out an approach you don’t want considered; narrowing the space of acceptable answers tends to raise the quality of what’s left rather than simply restricting it.
When to start a new conversation instead of continuing
A long, winding conversation that’s drifted through several unrelated topics tends to produce worse results on a new request than starting fresh, since irrelevant earlier context can subtly influence a new answer in unhelpful ways. If you’re switching to a genuinely unrelated task, a new conversation, rather than continuing an existing thread, often gets a cleaner, more focused result.
Related reading
Back to the beginner’s guide to using AI, our directory of best AI writing tools, and what is AI hallucination for why verifying an assistant’s output still matters no matter how well the prompt was written.
Frequently asked questions
Does prompt engineering matter less with newer, smarter models?
It matters somewhat less than it used to, since newer models are generally better at inferring intent from a loosely worded request, but it hasn’t become irrelevant. A specific, well-structured prompt still reliably outperforms a vague one on any model.
Should I say “please” and “thank you” to an AI assistant?
It has no measurable effect on output quality, and doing it or not is entirely a matter of personal preference rather than a technique that changes results.
Is there a difference between prompting ChatGPT and prompting Claude?
The core principles, specificity, examples, iteration, apply to both. Minor differences in how each model responds to certain phrasing exist, but the fundamentals covered here transfer across essentially every major assistant.
