How to Write Better AI Prompts

Prompting is not a secret language. There is no magic phrase that unlocks a better model, and the long lists of prompt tricks that circulate online mostly describe the same handful of ideas in different costumes. What actually works is unglamorous: giving the model enough context to do the job, saying clearly what a good answer looks like, and iterating instead of expecting one shot to land.

These patterns hold across assistants and survive model upgrades, because they address how the tools fail rather than how any one product is built.

Why vague prompts produce vague answers

A language model predicts a plausible continuation of what you gave it. If your request is compatible with a hundred different reasonable answers, you will get the most generic one — the average of everything the model has seen. The fix is not a cleverer instruction; it is narrowing the space of acceptable answers.

Compare:

Weak: Write a blog post about remote work.

Better: Write an 800-word post for engineering managers at companies of 50 to 200 people who are deciding whether to keep a hybrid schedule. Argue that the decision should be made per team, not company-wide. Use a plain, direct tone. No statistics unless you are confident they are real.

The second version is not more polite or more magical. It simply removes ambiguity.

Seven patterns worth learning

1. Give role and audience

Telling the model who it is writing for does more work than telling it who to be. Audience determines vocabulary, assumed knowledge and length. Instead of act as an expert marketer, try explain this to a small business owner who has never bought advertising before.

2. Show an example of what good looks like

This is the single highest-leverage technique. If you have one paragraph, one email or one row of output in the style you want, paste it and say: match this format and tone. Models are far better at imitation than at interpreting adjectives. Two or three examples work better than an elaborate description.

3. Specify the format explicitly

Ask for what you actually need: a table with these three columns, a bulleted list of no more than six items, a JSON object with these keys, a two-paragraph summary with no preamble. Adding no preamble or no closing summary is worth doing routinely — it removes the filler most assistants add by default.

4. Supply the source material

Most factual errors come from asking a model to recall rather than to read. If the information exists in a document, an email thread or a page, paste it in or attach it, then ask the question against that text. Add a constraint like: if the answer is not in the material I gave you, say so rather than guessing.

5. Ask for reasoning before the answer on hard problems

For analysis, comparison or anything with multiple steps, ask the model to work through it before concluding. Try: list the relevant factors first, then weigh them, then give your recommendation. This tends to surface flawed assumptions where you can see them. For simple tasks it just adds noise, so skip it there.

6. Constrain what you do not want

Negative constraints are underused. Useful ones include: do not invent statistics or citations; do not use the words leverage, seamless or robust; keep every sentence under 25 words; do not suggest options that require new software.

7. Iterate in the same conversation

The biggest gains come after the first response. Rather than rewriting your prompt from scratch, react to what you got: the second section is too abstract, rewrite it with a concrete example; cut this to half the length; you assumed a budget, ask me instead. Conversation is the interface. Treating it as a search box wastes most of the value.

A prompt structure that works for most tasks

• Task: one sentence on what you want produced.
• Context: who it is for, what has already been decided, what constraints exist.
• Source material: the actual documents, data or notes, clearly separated from your instructions.
• Format: length, structure, tone, and anything to omit.
• Quality bar: what would make this good or unusable — for example, must be accurate about the refund policy, must not promise a delivery date.

You do not need all five every time. For a quick question, one line is fine. For anything you will actually send to another human, all five take about ninety seconds and save far more.

Common mistakes

• Politeness padding. Please and thank you cost nothing but add no capability. Spend that attention on specifics instead.
• Asking for a source after the fact. If a model produced a claim from memory, asking it to cite a source often produces a plausible-looking reference that does not exist. Provide sources up front, or verify independently.
• Stacking ten requests into one prompt. Quality drops as instruction count rises. Split into steps and chain them.
• Accepting the first draft. First responses are usually a starting point, not output.
• Assuming it remembers. Memory features vary and are not guaranteed. In a new conversation, restate the context that matters.

Build a small prompt library

If you do the same task weekly, do not rewrite the prompt each time. Keep a plain text file of prompts that worked, each with a short note about what job it does. Over a few months this becomes more valuable than any technique list, because it is tuned to your work, your tone and your standards. When a prompt underperforms, edit it in place rather than starting over.

The practical takeaway

Better prompting is mostly better specification. State the task, give context and real source material, describe the output format, say what would make the answer unacceptable — and then iterate in conversation rather than hoping for a perfect first response. If you adopt only one habit, make it showing an example of what good looks like. It outperforms every clever phrasing trick, and it will keep working no matter which assistant you use or how the models change.