How-To Guides · 8 min read · September 26, 2026

How to Write Better AI Prompts

Turn a vague request into a brief with a job, a source, a shape, and a limit, then revise by naming the failure instead of asking for better.

Published by the GuideTopics Editorial Team.

A weak prompt is a missing brief

"Write a blog post about AI tools" is not a brief. It is a topic. The model fills the gaps with whatever is most common in its training, which is why the result sounds like every other post: a definition, a list, a soft conclusion. Better prompts are not magic phrases. They are the same information you would give a freelancer who has never met you.

This guide builds one brief from the pieces that change the output: the job, the source, the shape, the limits, and the revision pass. You can use it in ChatGPT, Claude, Gemini, or a local model. The ChatGPT, Claude and Gemini comparison helps you pick where to paste it. The brief itself is where the skill lives.

Start from a bad answer you already have

Open a chat and paste a short, lazy request. Use something you actually need this week, not a demo. Read the answer and mark three failures in the margin, or in a note beside the chat.

Those three marks become the next prompt. You are not "being nicer to the AI." You are closing the gaps it used.

The four lines that do most of the work

Write the next attempt as four short blocks. Labels help you notice when one block is empty.

Job. Who is writing, for whom, and what the piece must do. "You are drafting an email from a studio owner to a client who missed a call. The email has to reschedule without apologizing three times."

Source. The only facts the model may use. Paste them. If there is no source, say "do not introduce facts I did not give you; ask me instead." This single line prevents a large class of made-up specifics.

Shape. Length, format, and what done looks like. "120 words. Three short paragraphs. No bullet list. End with one proposed time, not three."

Limits. What to avoid. "Do not mention discounts. Do not say the word synergy. Do not add a postscript."

Here is the lazy version and the brief version of the same task.

Lazy: "Write a blog post about standing desks."

Brief: "Draft a 400-word post for people who already own a standing desk and keep forgetting to raise it. Audience: office workers who sit most of the day. Point: a timer and a default desk height beat motivation. Use only these facts: the writer sets a 25-minute timer, the desk stays at elbow height when standing, and the chair stays at the desk so sitting is still allowed. No health statistics. No product names. Four paragraphs, no list, no question in the last line."

The brief is longer than the lazy line. It is shorter than editing a generic draft for an hour. That trade is the whole practice.

Show the shape once

If you need a format the model has not seen from you, paste one example and label it as an example, not as source facts. "Match this shape, do not copy the content: Problem in one sentence. What we tried. What we kept." Two examples are enough if the format is odd. Five examples usually means you have not said the rule.

For voice, paste your sentences, not adjectives. "Friendly and professional" is how every draft already sounds. Two paragraphs of your own writing are a pattern the model can lean on. Tell it to match the rhythm, not to quote you.

Ask for the work in steps when the task is stacked

Some tasks are really three tasks. "Research this, draft the page, and make a chart" will blur together, and the chart will cite numbers from the draft instead of from the source. Split them.

  1. Ask for questions you still need to answer before a draft is possible.
  2. Paste the answers.
  3. Ask for an outline of headings only.
  4. Approve or rewrite the headings.
  5. Ask for the draft under those headings, using only the source you pasted.

This is slower on a two-sentence email and faster on anything you would be ashamed to publish unedited. The business guide uses the same split for team work, where the cost of a wrong fact is higher than the cost of an extra turn.

Revision is a second prompt, not a vibe

"Make it better" returns a longer version of the same draft. Name the change.

That last one is useful because it separates criticism from rewriting. If you let the model do both, it will sand off the criticism and hand you a smoother version of the problem.

When you agree with a criticism, write the fix yourself or ask for a rewrite of that paragraph only. Whole-draft rewrites tend to undo the sentences you already liked.

What not to put in the prompt

Do not paste passwords, private keys, medical details, or a customer's full record into a consumer chat. A clever prompt does not make that safe. If the material is sensitive, use a plan your organization has reviewed, or do not use a cloud model. The local model guide is the path when the file cannot leave the machine.

Do not paste a block of instructions you found online that tells the model it is a world-class expert with a secret method. The costume does less than the source and the shape. If a prompt template is mostly adjectives, delete the adjectives and put the facts back.

Do not ask the model to imitate a living writer's style by name and then publish the result as yours. Use your own samples. If you are studying how a writer builds a paragraph, ask for an analysis of a passage you have the right to quote, and keep the analysis in your notes.

A checklist you can reuse

Before you hit enter, look for empty slots.

If you write this checklist at the top of a note, you can duplicate the note per project. Store the winning brief next to the final piece. Next time the job rhymes, you will edit a brief instead of starting from "write a post about."

The glossary is there when a word in someone else's prompt tutorial is doing more work than it should. You do not need a new vocabulary to write a clear assignment.

Worked repair of a muddy answer

Suppose the model returned a list of seven tips for a client update, and you needed a single paragraph the client can forward. The repair prompt is specific:

"Rewrite the update below as one paragraph, under 90 words, that a project manager can forward to their director. Keep only items that are in the update. Drop the tips format. Do not add a status that is not written here. Update: the homepage copy is in review, the photo shoot moved to Thursday, and the contact form is still broken on mobile."

Read the result against the three facts. If a fourth fact appeared, the prompt still has a hole, or the model ignored the hole. Say so: "You added a fact. Remove it." Short corrections beat starting over, as long as the thread has not filled up with failed drafts. If it has, open a new chat and paste only the brief and the last good text.

Practice on work you are about to do

Pick tomorrow's smallest real assignment. Write the four blocks. Run it. Fix one failure with a revision prompt that names the failure. Save both prompts in the project folder. That pair is worth more than a folder of prompts you have never used.

If you write for publication, take the same brief into the tools for writers and decide which pass the assistant is allowed to touch. Prompt skill does not replace that decision. It makes the decision enforceable.

#prompt engineering #AI prompts #ChatGPT tips #AI writing

Recommended on Amazon

The Coming Wave

The Coming Wave
Mustafa Suleyman. A clear look at how frontier AI changes work, power, and everyday tools.

AI Superpowers

AI Superpowers
Kai-Fu Lee. How the AI landscape is shifting, and how to position your work inside it.

All-in On AI

All-in On AI
Thomas H. Davenport & Nitin Mittal. How companies put generative AI into real operations, not just demos.

2k to 10k

2k to 10k
Rachel Aaron. A short, practical book on writing faster while keeping the work specific.

As an Amazon Associate I earn from qualifying purchases.