Writing Prompts

Most AI frustration comes from one-line questions. If you remember only one thing: treat the AI like a brand-new assistant on their first day, who needs a clear brief.

You don't need to memorize a script. You just need five habits. Once these become second nature, your answers will improve dramatically.

Rule 1. Say who, what, and who it's for

Three tiny pieces turn a weak prompt into a strong one: who you are, what you're trying to make, and who it's for.

Compare:

Write an email asking for an extension.

Versus:

I'm a second-year student who's fallen behind because I was sick for a week. I need to email my professor asking for a three-day extension on an essay. She's fair but strict about deadlines and dislikes long excuses. Keep it under 120 words, polite and direct, and propose a specific new date.

Same request. Radically different result. The second one gives the AI the situation, so it can reason about what will work.

Rule 2. Show one or two examples of "good"

AIs are very good at copying. They're bad at guessing. If you want something in your style, show them your style.

Paste in one or two samples of writing you're proud of before asking:

Below are two notes I've written that sound like me. Please write a new one in the same voice, about [topic].

"""
[Paste note 1]
"""

"""
[Paste note 2]
"""

This one habit alone, showing examples, is responsible for most of the gap between "that sounds generic" and "that sounds like me."

Talking to an AI is a lot like briefing a new assistant on their first day. If you only describe the tone you want ("warm but professional"), they have to guess what you mean. If you show them one to three real examples of good work, they can copy the pattern exactly. One example is good; two or three is better, because it lets the AI see what stays constant across them and lock onto it. Showing beats describing, every time.

Rule 3. Tell it to think first

For anything with moving parts (a plan, a comparison, a hard decision), add one sentence at the end:

Think through this step by step before you answer.

Seems like a small thing. It isn't. It changes how the AI reasons. You'll get a more careful answer, and often the AI catches things it would have glossed over.

There's a bigger version of this move. Most AIs now offer a "thinking" (or "reasoning") mode, usually a toggle or a different model in the dropdown. A normal fast model is perfect for most everyday work: rewriting a message, brainstorming ideas, cleaning up a paragraph. But when the task is hard, untangling why some numbers don't add up, planning something with lots of pieces, troubleshooting a stubborn problem, switch to the thinking model. It slows down and retraces its own steps before answering, and it's noticeably better at those knotty jobs. Fast model for most things; thinking model when the problem is hard.

Rule 4. Keep your instructions separate from your content

When your prompt includes stuff to work on (a draft, an email you received, a transcript), wrap that stuff in quotes or brackets so the AI knows what's the task and what's the material.

Please rewrite the paragraph below in a warmer tone. Keep all the facts the same. Don't add anything new.

"""
[paste the paragraph here]
"""

Without the triple quotes, the AI has to guess which part is the instruction and which part is the content. Sometimes it guesses wrong and starts editing your instructions.

Rule 5. Ask for short. Then expand if you want.

AIs tend to over-explain. Cut that off at the knees:

Be concise. Skip the preamble. No warm-up sentence. Just the answer.

If the short answer is missing something, you can always say "expand on point 2." But starting short and expanding is much faster than starting long and deleting.

Putting it together

You don't need to use all five rules every time. For quick questions, one line is fine. But when the answer matters, when it'll go to someone who counts, or onto a page with your name on it, take 30 seconds to set up the prompt properly.

Here's what a full, five-rule prompt actually looks like. Each colored section matches a rule number above.

Put the most important instruction last

Here's a habit that sounds fussy and isn't. When you paste in a long document and then ask the AI to do something with it, the order matters. Models pay closest attention to the very start and the very end of a long input. The middle is where things get skimmed and lost.

So the shape to avoid is: your instruction, then a huge wall of pasted text. By the time the AI finishes reading the wall, your instruction is far away and faint. Flip it. Put the document first, then your instruction right at the bottom, where the AI's attention is strongest:

"""
[the long document goes here]
"""

Now: summarize the three main arguments above in one sentence each.

Ask for a shape, not just an answer

A vague request produces output you have to read carefully to judge. A request with a shape produces output you can check at a glance. Tell the AI the exact form you want the answer in:

• "Give me a table with three columns: option, one upside, one downside."
• "Exactly five bullets, each under 15 words."
• "A yes-or-no, then one sentence of reasoning. Nothing else."

The payoff is speed. If you asked for five bullets and got seven rambling paragraphs, you know instantly it didn't follow the brief, without reading a word of the content. Naming the shape makes the output checkable.

One extra move: "use code" whenever numbers are involved

This is a reflex worth building. The AI reads text in chunks, not digit by digit, so plain arithmetic is one of the few things it quietly gets wrong. If your prompt involves counting, adding, sorting, or percentages, add four words:

Use code to calculate this.

Instead of guessing, the AI writes a tiny program and reports the exact result. "Split a $214 dinner bill three ways, then add 18% tip to each share": with "use code," the numbers are right. Without it, treat any figure it gives you as a rough guess to double-check.

Two things to stop doing

• Stop typing one-liners when the output matters. One-line questions get one-line-quality answers.
• Stop re-explaining your background in every new chat. That's what a workspace is for. Set it up once, then start new chats inside it: see Making Projects.