How to Write Better AI Prompts in 2026 (10 Examples)

Most people write vague prompts, get vague answers, and conclude the AI "isn't that smart." The truth is usually simpler: the prompt was the problem. Here's how to fix that, with ten before-and-after examples you can copy today.

Why your prompts probably underperform

We tested hundreds of prompts across different models while building and comparing tools for our readers. The pattern was consistent: the gap between a mediocre answer and a genuinely useful one almost always came down to how the request was written, not which model answered it.

The three most common mistakes we found were being too short ("write me a marketing email"), giving no context about who you are or who the output is for, and not specifying the format you actually want back. AI models are good at guessing, but every guess is a chance to guess wrong.

You don't need to learn "prompt engineering" as a technical skill. You need to communicate the way you would with a competent freelancer who has never met you: tell them what you want, who it's for, what good looks like, and what to avoid. That's the whole game.

The four things every good prompt includes

Before the examples, here's the checklist we kept coming back to. You won't need all four every time, but the more you include, the better your results.

If you want a deeper walkthrough of the fundamentals, we cover them in our complete AI prompts guide. This article is the practical, example-first companion to it.

10 examples: bad prompt vs better prompt

1. Writing an email

Bad: "Write an email to a client about a delay."

Better: "Write a short, apologetic but confident email to a client telling them their website project will be delivered 4 days late because of a supplier issue. Keep it under 120 words, no groveling, and offer one concrete next step."

The bad version gets you a generic template. The better version gets you something you can send after a 10-second edit.

2. Summarizing a document

Bad: "Summarize this."

Better: "Summarize this contract in 5 bullet points a non-lawyer can understand. Flag anything that limits my rights or exposes me to fees, and quote the exact clause when you do."

Adding "quote the exact clause" cut down hallucinated summaries dramatically in our testing, because it forces the model to point at real text instead of paraphrasing loosely.

3. Brainstorming ideas

Bad: "Give me some content ideas."

Better: "I run a small ceramics studio and sell mugs online. Give me 10 short-form video ideas that show the making process, each with a one-line hook. Skip anything that requires expensive equipment."

Constraints make brainstorming useful. Without "skip anything expensive," half the ideas will assume you have a film crew.

4. Fixing your own writing

Bad: "Make this better."

Better: "Rewrite the paragraph below to be clearer and 30% shorter. Keep my voice, don't add marketing buzzwords, and don't change the meaning. Show me the rewrite, then list what you changed and why."

"Make this better" invites the model to inflate your text with jargon. Naming what you don't want is as important as naming what you do.

5. Learning something new

Bad: "Explain compound interest."

Better: "Explain compound interest to me like I'm 15, using one real example with actual numbers over 10 years. Then give me one common mistake people make with it."

The "actual numbers" instruction is the difference between an abstract definition and something that sticks.

6. Comparing options

Bad: "Should I use a Roth IRA or a 401k?"

Better: "Compare a Roth IRA and a traditional 401k for a 32-year-old earning $70k who expects to earn more later. Use a short table with columns for tax treatment, contribution limits, and best-fit scenario. Note that you're not a financial advisor."

Tables force structure, and structure exposes gaps in the answer that a wall of text hides.

7. Coding help (yes, even for non-coders)

Bad: "Write me a spreadsheet formula."

Better: "In Google Sheets, I have dates in column A and amounts in column B. Write a formula that sums all amounts from the current month only. Explain each part of the formula in one line so I can adjust it later."

Specifying the tool (Google Sheets vs Excel matters) and asking for an explanation means you can fix it yourself next time instead of coming back.

8. Drafting a plan

Bad: "Help me plan a product launch."

Better: "I'm launching a $29/month budgeting app in 6 weeks with a $500 budget and no team. Give me a week-by-week plan with the 3 highest-impact tasks per week. Be realistic about what one person can do; cut anything that needs paid ads."

Budget, timeline, and team size turn a fantasy plan into a doable one. The phrase "be realistic" genuinely tempers overambitious output.

9. Getting feedback

Bad: "What do you think of my business idea?"

Better: "Here's my business idea: [paste]. Act as a skeptical investor. Give me the 3 biggest reasons this could fail and 2 questions you'd ask before investing. Don't reassure me."

AI defaults to encouragement. If you want honest feedback, you have to explicitly ask it to stop being nice.

10. Repurposing content

Bad: "Turn this blog post into social media posts."

Better: "Turn the article below into 3 LinkedIn posts. Each should stand alone, open with a specific claim (not a question), stay under 150 words, and end with one takeaway. No hashtags, no emojis."

"Open with a specific claim, not a question" is the kind of small instruction that changes the whole output. Most AI-generated LinkedIn posts start with a rhetorical question because nobody told them not to.

The one trick that beats all others: show an example

If you take nothing else from this article, take this. When we added a single example of the output we wanted — a sample email in the tone we liked, a paragraph in our writing style, a formatted list we approved of — the quality jumped more than any other change we made.

This is called few-shot prompting, but you don't need the term. Just paste something and say "match this style." A model that saw one good example of your voice produced usable drafts far more often than one working from description alone. It's the difference between telling someone what you want and showing them.

The trade-off is length. Longer prompts take more effort to write, and for a quick throwaway question they're overkill. Save the detailed prompts for work you'll actually use.

Different prompts want different models

Here's something we ran into constantly: the "best" model depends on the prompt. In our testing, some models were noticeably stronger at long-form reasoning and careful comparisons, others at fast, punchy copywriting, and others at code and structured data. A prompt that produced a brilliant answer on one model produced a mediocre one on another.

For non-technical users, this creates an annoying problem. You'd have to know which model is good at what, keep multiple subscriptions, and manually paste your prompt into the right one. Most people don't want to do that, and honestly, they shouldn't have to.

This is where Panvoxx's Auto Routing earns its place. Instead of you guessing, it reads the type of prompt — a coding request, a creative brief, a long analytical question — and routes it to a model suited for that job. So the coding formula from example 7 and the skeptical-investor feedback from example 9 don't have to compete for the same generalist model. We think this matters more as models specialize further, which they clearly are.

It's not magic, and routing won't rescue a lazy one-line prompt. But paired with the habits above, it removes a decision most people get wrong. If you're weighing your options, our roundup of the best AI platforms in 2026 covers how multi-model access compares to single-model tools.

Common myths worth dropping

"Being polite gets better results." We didn't find meaningful evidence that "please" and "thank you" improve output. Clarity matters; manners don't. Say them if you like, but they're not a technique.

"Longer prompts are always better." No. Beyond a point, extra detail adds noise and the model loses track of what you actually asked. Aim for complete, not bloated.

"You need the newest, most expensive model for everything." Often you don't. For summaries, rewrites, and simple lists, a mid-tier model answers just as well and faster. Many solid options are even free — we round up the good ones in our guide to free AI tools in 2026.

The bottom line

Better prompts aren't about secret phrases or engineering degrees. Give context, name the task, set constraints, and — when it matters — show an example. Do that, and the right model will give you output you can actually use, not just admire.

If you'd rather stop guessing which model fits which prompt, Panvoxx routes each request to a suitable model automatically and lets you try 9 models free for 3 days. You can put these examples to the test yourself and see the difference side by side. Start your free trial here.