How to Write Better Prompts

Most advice on this topic stops at "be specific," which is true but not very useful. This guide goes a little deeper: a working mental model for why prompts succeed or fail, and a set of techniques you can apply right away.

The one idea that makes everything else click

An AI assistant doesn''t look up answers in a database. It generates a plausible continuation of your text, one piece at a time, based on patterns it learned from enormous amounts of writing. That sounds abstract, but it has a very practical consequence:

The more your prompt looks like the kind of text that precedes a great answer, the better the answer you''ll get.

That''s the whole game. Every technique below is just a way of steering the model toward the response you actually want, instead of leaving it to guess from a vague request.

Technique 1: Give context and a goal

Compare these two:

"Write an email declining the meeting."

"Write a short, warm email declining a meeting invitation from a senior colleague I want to stay on good terms with. I''m declining because of a scheduling conflict, and I''d like to suggest meeting next week instead."

The second one isn''t just longer — it tells the model the tone, the relationship, the reason, and the outcome you want. You''ve removed the guesswork. A good rule: include who it''s for, what it''s for, and what "good" looks like.

Technique 2: Show an example

Telling the model what you want works. Showing it works better. If you want output in a particular style or format, paste one example of it:

"Rewrite these product descriptions in the style of this one: [paste example]. Here are the three to rewrite: ..."

Even a single example dramatically narrows down what you''re after. This is one of the most underused techniques, and it''s especially powerful for matching a voice, a structure, or a formatting convention.

Technique 3: Ask for the format you want

If you don''t specify a format, you''ll get the model''s default — usually flowing paragraphs. So just ask: a bulleted list, a table, a step-by-step guide, three options to choose from, under 100 words, a specific structure. The model is happy to oblige; it just needs to be told.

"Give me this as a table with columns for option, pro, and con."

Technique 4: Break big asks into steps

For anything complex, asking for the final product in one shot often produces something shallow. You''ll get better results by either breaking the task into a sequence of messages, or by asking the model to work through it in stages:

"First, outline the main sections. Then, once I approve the outline, write the draft."

Telling it to "think step by step" or to plan before producing also helps on reasoning-heavy tasks, because it gives the model room to work rather than forcing an instant answer.

Technique 5: Treat it as a conversation, not a vending machine

The first response is a starting point, not a verdict. Rather than rewriting your prompt from scratch when something''s off, just say what to change:

"Good, but make it more concise and drop the second example."

This back-and-forth is where the real quality comes from. People who get great results are rarely writing one perfect prompt — they''re iterating quickly across three or four turns.

Technique 6: Give it an out

Because these systems aim to be helpful and fluent, they''ll sometimes fabricate an answer rather than admit they don''t know. You can counteract this directly:

"If you''re not sure or don''t have the information, say so rather than guessing."

It''s a small addition that noticeably reduces confident nonsense, especially on factual questions.

Technique 7: Provide the material, don''t rely on its memory

If your question is about a specific document, dataset, article, or set of facts, give them to the model directly — paste the text, upload the file, or enable web search. Don''t lean on it to recall the precise contents of something from training. Working from material you''ve supplied is one of the most reliable ways to use these tools, and it sidesteps the fabrication problem almost entirely.

A quick checklist

Before you hit enter on anything important, glance over this:

• Did I give context (who, what, why)?
• Did I state what a good answer looks like?
• Could an example make it clearer?
• Did I specify the format?
• For factual work, did I supply the source or ask it to flag uncertainty?

You won''t need all of these every time. But when a response disappoints you, the fix is almost always one of them.

The takeaway: good prompting isn''t a bag of magic words. It''s the habit of giving the model enough to work with and steering it across a few turns. Get that, and you''ll outperform anyone hunting for the perfect one-line incantation.