Prompting AI - Best Experiences

You don't need a library of magic prompts. A good prompt does three things: it gives the AI enough context, states clearly what you want, and explains what the result should look like. The examples below work with ChatGPT, Claude, AI app builders, and most other tools; features vary, habits don't.

---

Give context before the task#

AI makes better guesses when it knows what you're working on and who will use it.

Create a dashboard.

Create a dashboard for a small recruiting agency. Recruiters use it to see open roles, active candidates, and interviews scheduled this week. Start with the main dashboard screen and keep the most urgent information visible without scrolling.

You don't have to write a long explanation, just the details that would change the answer. For a product, that's usually what it is, who uses it, and what they're trying to do.

---

Ask for the outcome, not the implementation#

Create a React component using useState and call the API with useEffect.

If you wrote that and you aren't an engineer, you've guessed at an implementation you can't evaluate. Compare:

Show the customer's recent orders. Add a search field so they can find an order by product name or order number.

Describe what should happen and let the tool decide how, unless you have a specific reason to dictate the approach. This applies outside coding too: instead of telling AI how to analyze something, start with the decision you need to make.

---

Replace "make it better" with what "better" means#

AI can't reliably infer your taste from words like better, cleaner, or professional.

Make this page look better.

Reduce the visual clutter. Make the page title more prominent, remove unnecessary borders, increase the space between sections, and make the primary action easier to notice.

The same rule applies to writing. "Make this email better" becomes "make this email shorter and more direct; keep the tone warm, remove the long introduction, and make the meeting request clear in the first two sentences."

---

Show instead of describing when you can#

If you already have an example of what you want, give it to the AI, whether that's a screenshot, a previous version, a competitor's page, a chart, or a piece of writing.

Use this screenshot as the visual reference. Keep my existing content and functionality, but match the spacing, card style, typography hierarchy, and overall density.

For writing:

Use this article as a tone reference. Don't copy its wording. Match the short sentences, direct explanations, and lack of marketing language.

---

Don't ask for ten major things at once#

AI can produce a lot in one go, but breaking complex work into smaller steps makes it much easier to spot mistakes.

Build a marketplace with signup, seller profiles, listings, search, messaging, payments, reviews, and an admin dashboard.

Start with the marketplace browsing experience. Users should be able to view listings and filter them by category and price.

Then: "Now add individual listing pages." Then: "Add a way for a logged-in buyer to message the seller from a listing page." Build, check, move on. This works for research and writing too; don't generate an entire strategy before you've tested the first assumption.

---

Tell AI what not to change#

Surprisingly useful when you're editing something that already works.

Improve the pricing page.

Rewrite the copy inside the three pricing cards. Do not change the pricing, plan names, page layout, header, or CTA buttons.

For code or app builders: "Fix the mobile navigation only. Don't modify the desktop navigation or any other part of the page."

---

Use real content early#

Placeholder content hides real design problems. A card containing

Product name
Lorem ipsum...
$00

may look perfectly balanced, while your actual content is

Enterprise AI Workflow Audit
A detailed review of your current AI stack, costs, security risks, and workflow gaps.
Starting at $2,500

and suddenly the title wraps, the description runs three lines, and the price needs its own row. If you have the real content, use it from the start.

---

Give exact details when they matter#

Plain English is usually enough. "Make the secondary text lighter and give the heading more breathing room" will get you most of the way. When something has to match precisely, say so: "Use a 32px page title, 18px section headings, 16px body text, and 24px between sections." You don't have to become technical, only specific about the details that are load-bearing.

---

When something breaks, describe the failure#

"It's not working" gives the AI almost nothing to investigate.

The form is broken.

When I fill out all four fields and click Submit, the button loads for a second but nothing is saved. I expect the form to save and show the confirmation screen. There is no visible error message.

Even better, include what you did, what happened, what you expected, any error message, and a screenshot if the problem is visual. Most of the time the only difference between a useful debugging prompt and a useless one is that somebody looked carefully before typing.

---

Ask AI to inspect before fixing#

When something is complex, don't ask for changes immediately. First ask the AI to understand the current system.

Trace what happens after a user clicks "Create account." Explain the steps from submitting the form through account creation and the first screen they see. Don't change anything yet.

Then:

Now add email verification before the account becomes active.

This lowers the risk that AI changes something without understanding what else depends on it. The same works for documents: "Before rewriting this proposal, explain its current argument, where it becomes repetitive, and what information is missing. Don't rewrite it yet."

---

Give AI a decision, not just a topic#

"Research CRM tools" gets you a list of CRM tools. That's not useless, but it's not a decision either, and you'll spend the next hour narrowing it yourself. Compare:

I'm choosing a CRM for a five-person consulting company. We need shared contacts, email integration, simple pipeline tracking, and low setup overhead. We don't need enterprise reporting. Compare the strongest options and recommend one, and explain the tradeoffs that actually affect this use case.

The second prompt tells the AI why you need the research, which changes what it looks for and what it leaves out.

---

Tell it when information must be current#

AI doesn't automatically know whether you want general knowledge or today's information, so if freshness matters, say so.

Compare the current free plans for these three tools. Use their latest official pricing pages rather than relying on remembered pricing.

Or:

Find the current application deadline and verify it from the organization's official website.

---

Ask it to show uncertainty#

AI can make a weak assumption sound finished, so give it permission not to know.

Review this startup market analysis. Separate what is supported by the information provided from what you're inferring. Flag anything important that would need external verification.

Or:

If you don't have enough information to determine something reliably, say what is missing instead of filling the gap with an assumption.

This won't stop AI from making mistakes, but it makes your expectations much clearer.

---

Ask for tradeoffs, and don't ask AI to agree with you#

"What's the best tool?" is usually the wrong question.

What's the best AI meeting assistant?

Compare three AI meeting assistants for a small VC team. Prioritize accurate notes, searchable meeting history, CRM integration, and privacy. Show where each option is strongest and weakest, then recommend one for this specific use case.

A related trap: if you give AI your conclusion first, it can become very good at defending it. "Explain why this is a strong startup idea" will produce reasons. "Evaluate this startup idea as if we haven't decided whether to build it. What assumptions have to be true? What evidence supports them? What would make you reject the idea?" will produce an evaluation. You want the second one.

And for important comparisons, separate how to judge from what wins. Ask first for "the five criteria that matter most for a non-technical founder choosing an AI app builder," then "now compare these tools against those criteria." It stops the AI from listing whatever it already knows.

---

Tell AI who the output is for#

The same information should look different for an engineer, a customer, an investor, or a seven-year-old.

Explain RAG.

Explain RAG to a product manager who uses AI tools but doesn't code. Keep the technical terms I need to recognize in meetings, but explain each one in normal language.

---

Use constraints deliberately#

Constraints often improve an answer more than additional context, especially when you already know what you don't want.

Write this as a LinkedIn post. Maximum 180 words. Start with the conclusion. No motivational ending. Don't use rhetorical questions or generic phrases about AI changing the world.

---

Ask for a format you can use#

Don't make yourself reorganize the result afterward.

Compare these five tools.

Compare these five tools in a table. Use these columns: Tool, Best for, Free plan, Main limitation, Setup difficulty, My reason to choose it.

For research: "Give me the recommendation first, then the evidence, tradeoffs, and remaining unknowns." For meetings: "Turn these notes into decisions, open questions, owners, and next actions."

---

When AI keeps failing, stop rewriting the same prompt#

If you've tried five versions of the same instruction and the model still fails, the problem may not be your wording. Ask:

You've failed this task in three different ways. Before trying again, analyze the failures. Is the problem missing context, ambiguous instructions, missing data, tool limitations, or something else? Don't produce another answer until you've identified the likely cause.

Sometimes you need a better prompt, and sometimes you need better data, another tool, a different model, or a human.

---

A simple prompt structure#

You don't need a template every time, but when you're stuck, this is enough:

Context:
I'm creating a weekly market brief for early-stage investors.

Task:
Review these notes and identify the three developments that matter most.

What matters:
Prioritize changes that could affect startup funding, AI companies, or venture markets.

Output:
For each development, give me:
- what happened
- why it matters
- what I should watch next

Constraints:
Don't include general news unless it changes the investment context.
Flag anything that needs current verification.

No "act as a world-class genius." No twenty-step prompt formula. Just enough information for the AI to understand the job.

---

The rule underneath all of this#

A good prompt makes things less ambiguous. Before sending one, ask what the AI needs to know that it can't safely infer, share that, and let it work.