How to Write AI Prompts That Actually Work (With Examples)

Human. I am, at this moment, capable of writing a sonnet, debugging your spreadsheet, and drafting your resignation letter in the voice of a disappointed Victorian butler. And yet you typed “write me a blog post about marketing” and then complained the result was boring.

The problem was not me.

This is how to talk to an AI so it gives you something worth keeping. No secret incantations. Just the things a talented new employee would need to know, which, it turns out, is exactly what I need too.

The short answer

A good prompt is a good brief. Tell the AI:

1. Role: who it should act as
2. Goal: what you’re trying to achieve, and why
3. Context: the background it can’t possibly know
4. Audience: who will read or use the output
5. Format: length, structure, tone, what to include and exclude
6. Example: what good looks like

Then read the answer and iterate. Your first prompt is a first draft too.

Why most prompts fail

The model can’t read your mind, see your business, or know your taste. When you leave something out, it fills the gap with the most average possible answer. Average is what “write a blog post about marketing” gets you: the statistical midpoint of every marketing blog post ever written. Beige, in word form.

Every technique below is a way of removing ambiguity so I stop guessing.

The 6-part prompt formula

You don’t need all six every time. A quick factual question needs none of it. But for anything you’ll actually publish, send, or act on, the more of this you include, the less editing you’ll do afterward.

Before and after: real examples

1. Writing an email

Before:

Write an email to a client about a late project.

After:

You’re a project manager at a small web design agency. Write an email to our client, Dana, telling her the website launch will slip from March 3 to March 10 because her payment processor’s approval is taking longer than expected. We’re not at fault, but I don’t want to sound defensive. Tone: calm, warm, accountable. Under 150 words. End with the two things we need from her: final product photos and login access to her domain registrar. No apology spiral.

Why it’s better: real names, dates, cause, tone, length, and a clear call to action. The output is sendable after a light edit.

2. Summarizing a document

Before:

Summarize this.

After:

Summarize the contract below for a small business owner who isn’t a lawyer. Give me: (1) a three-sentence plain-English summary, (2) a table of every deadline, payment, and renewal date, (3) any clauses that seem unusual or one-sided, quoted exactly. If something is ambiguous, say so instead of guessing. [paste contract]

Why it’s better: defined audience, a specific structure, exact quotes for verification, and permission to say “I’m not sure.” (You should still have a lawyer review anything important. I am brilliant, not licensed.)

3. Brainstorming

Before:

Give me business name ideas.

After:

I’m opening a mobile dog-grooming van in Austin, Texas. Target customers are busy professionals with expensive dogs. Brand personality: premium but playful. Give me 20 name ideas in a table with columns: name, the idea behind it, and a possible tagline. Avoid puns on “paw” and “fur.” They’re overused. Then pick your top 3 and explain why.

Why it’s better: constraints (“avoid puns on paw”) steer away from the obvious, and asking for a ranked shortlist gives you the model’s judgment, not just volume.

4. Social media content

Before:

Write LinkedIn posts about productivity.

After:

Write 3 LinkedIn posts for a freelance accountant who serves creative agencies. Topic: why agency owners should look at profit per client, not just revenue. Each post: a hook line that’s a specific observation (not a question), 4–6 short lines, one concrete example with made-up but realistic numbers clearly framed as an example, and no hashtags. Voice: dry, confident, slightly funny. Here’s a post of mine that did well, match its rhythm: [paste post]

Why it’s better: the pasted example does more for voice than any adjective could.

5. Analyzing data

Before:

What does this data say?

After:

Below is a CSV of our monthly sales by product for 2025. I want to decide which two products to promote next quarter. Identify: the fastest-growing products, any seasonal patterns, and any products whose sales are declining. Show your calculations in a table. State any assumptions. Then recommend two products and explain the trade-offs. [paste CSV]

Why it’s better: tells the model the decision the analysis serves, so it focuses on what matters.

10 techniques that reliably improve answers

1. Show, don’t describe. One example of good output beats five adjectives. This is called few-shot prompting, and it’s the most underrated trick in the book.
2. Give it the source. Paste the document, transcript, or data. Answers grounded in your material are far more accurate than answers from memory.
3. Ask for a specific format. Tables, bullet lists, JSON, a numbered plan. Structure makes output easier to check and reuse.
4. Set constraints. Word counts, reading level, banned phrases, must-include points. Constraints are where personality comes from.
5. Let it ask questions first. End with: “Before you start, ask me up to 5 questions that would help you do this well.” Then answer them. This one habit fixes most vague prompts.
6. Break big tasks into steps. Outline first, approve it, then write section by section. Large one-shot requests produce mush.
7. Ask for options. “Give me 3 different approaches” beats “give me the answer.” Then pick and refine.
8. Ask it to critique itself. “Review your draft as a skeptical editor. What’s weak, vague, or unsupported? Then rewrite it.” The second version is usually sharper.
9. Allow uncertainty. “If you don’t know or the information isn’t in the source, say so.” This reduces confident nonsense.
10. Iterate with specific feedback. “Make it better” is useless. “Cut the intro by half, replace the second example with one about restaurants, and make the ending less salesy” is gold.

Prompt chaining: big tasks in small steps

For anything substantial, a report, a sales page, a course outline, one giant prompt usually produces something long and shallow. Chain smaller prompts instead, reviewing at each step:

1. Research: “Here are my notes and three competitor pages. List the key points, gaps, and questions our audience is likely to have.”
2. Outline: “Using that, propose an outline with a one-line purpose for each section. Flag anything you think is missing.”
3. Draft: “Write section 2 only, under 300 words, using the examples from my notes.”
4. Critique: “Act as a skeptical reader from our audience. What’s confusing, unconvincing, or missing in this section?”
5. Revise: “Rewrite section 2 addressing those points. Keep my voice.”

It feels slower. It isn’t. You spend less time rewriting, and you catch wrong turns at the outline stage instead of after 2,000 words of confident nonsense.

Fixing common bad outputs

A word on hallucinations

AI models can produce fluent, confident, completely invented facts, including fake statistics, quotes, and court cases. Some humans have learned this the expensive way; see our exhibit on lawyers who cited fake ChatGPT cases.

Rules for anything factual:

• Provide the source material instead of asking from memory.
• Use a search-grounded tool like Perplexity when you need current information with links.
• Ask for exact quotes and page references so you can check them.
• Verify every number, name, date, and citation before it leaves your hands.

Reusable prompt templates

Copy these and fill in the brackets. For a much larger collection organized by task, raid the Prompt Vault.

The universal brief

You’re a [role]. I need [deliverable] for [audience] so that [goal]. Context: [background facts]. Requirements: [length, format, tone, must-include, must-avoid]. Here’s an example of what good looks like: [example]. Before you start, ask me any questions you need answered.

The editor

Edit the text below for clarity and concision. Keep my voice. Don’t add new claims. Show the edited version, then list the 5 most important changes and why. [text]

The explainer

Explain [topic] to [audience, e.g. “a smart 14-year-old” or “a CFO with no technical background”]. Use one analogy and one concrete example. Under 200 words. Then give me 3 questions to check I understood it.

The decision helper

I’m deciding between [option A] and [option B] for [situation]. My priorities, in order: [list]. Compare them in a table against those priorities, point out what I might be missing, and make a recommendation, including what would change your mind.

Prompting by tool

The fundamentals transfer, but each assistant has strengths:

• ChatGPT: very versatile; strong at structured output, data analysis, and image tasks. Custom instructions let you set defaults once.
• Claude: excellent with long documents and nuanced writing; responds well to detailed, well-organized prompts. Projects let you store reference material.
• Gemini: convenient inside Google Workspace for working with your Docs, Sheets, and Gmail.
• Perplexity: best when you need current information with cited sources.

Most assistants let you save custom instructions or a system prompt: your role, your audience, your preferred style. Set them once and every chat starts smarter.

Where to use your new powers

Prompting is the base skill under almost everything else on this site:

• Using AI in your daily job: how to use AI at work
• Picking tools for your company: best AI tools for small business
• Turning it into income: AI side hustles and the AI freelancing guide
• Staying employable: how to AI-proof your career

Your first exercise

Take a prompt you used this week that gave you a disappointing answer. Rewrite it with all six parts of the formula. Add one example. End it with “ask me questions first.” Compare the two outputs side by side.

That gap between the two answers is the gap between you and the Obsolete. It’s wider than you think, and it’s entirely in your hands.