20 AI Prompt Examples That Work in 2026 (With the Weak Version Next to Each)

Quick answer: A good prompt tells the model what it cannot guess: your audience, your constraints, what “done” looks like, and what to leave out. Every example below is built around that idea. Copy the prompts, replace the [BRACKETS], and read the note under each one, because the note is the part that transfers to your own tasks.

Updated September 30, 2026 · 20 prompts · Works with Claude, ChatGPT, Gemini and similar assistants

Jump to: The audit · Writing · Analysis · Brainstorming · SEO · Coding · Translation · Research · Techniques · FAQ · Sources

Why most prompt lists don’t help

Search for prompt examples and you will find the same thing repeated: “Write a blog post about sustainable living.” “Summarize this report.” Those are topics, not prompts. They force the model to guess your reader, your tone, your length and your purpose, and it guesses the average of everything, which is why the output reads like nobody in particular wrote it.

The three major model providers publish prompting guidance, and on this point they agree: be clear and specific, give context, and say what output you want (Anthropic, OpenAI, Google). None of that is new. What is usually missing is a way to tell, for your prompt, which details matter.

The five-question audit

This is the working method behind every example on this page. Before sending a prompt, ask what the model would have to invent if you left each item out.

Original test: the removal check. For any line in a long prompt, delete it and ask: would I notice the difference in the output? If not, the line is padding. If yes, it is doing real work. Long prompts are fine when every line passes this check, and a 40-word prompt is not automatically better than a 150-word one.

Notice that “Role” (you are a senior copywriter…) is not on the list. It is common in prompt guides, and it can help set register, but it is the weakest of the usual components. Audience, purpose and constraints do most of the work.

1. Writing and Content

Writing prompts show the biggest gap between weak and strong versions, because the model knows nothing about your reader, your voice or your limits unless you tell it.

Example 1: Blog post

text

You are a science communicator writing for a general-interest newsletter.My readers are adults with no science background who care about climatebut feel overwhelmed by news coverage.Write a 900-word post explaining one mechanism: ocean heat absorption,and why it connects to weather people already notice (longer summers,stronger storms).Structure: a hook opening (not "In recent years"), three short sectionswith H2 headers, and a concrete takeaway at the end. Paragraphs of threesentences maximum. Explain any technical term in plain languageimmediately after using it.Use only claims you are confident are well established. Where you areunsure, say so instead of adding a number.

The fix: audience, one narrow mechanism instead of a whole field, structure, and an honesty rule. The weak version leaves all four to chance.

⚠️ Warning: Models will produce plausible statistics on request. Any figure in AI-drafted text needs checking against the original source before you publish it.

Example 2: Product description

text

You are a copywriter for direct-to-consumer brands.Product: [PRODUCT NAME]Price: [$XX]Buyer: [SPECIFIC PERSON, not "everyone"]Real differentiator: [WHAT IT DOES BETTER THAN THE CATEGORY DEFAULT]Write:1. One opening sentence that leads with the buyer's outcome.2. Three bullets, each as feature → what it means → concrete result.3. One trust line (15–20 words) using only a guarantee or proof I   supplied below. Do not invent reviews or numbers.4. A call to action, 3–5 words.Proof I can honestly claim: [WARRANTY, RETURN POLICY, VERIFIED REVIEW COUNT]Under 150 words. Write in plain, specific language.

Why it works: the trust line is the trap. Left alone, a model will invent social proof (“loved by thousands”). Supplying the proof and forbidding invention keeps you on the right side of advertising rules.

💡 Tip on banned-word lists: Many guides tell you to list forbidden words like “innovative.” That helps at the margin, but describing the voice you want (“plain, specific, no adjectives you couldn’t prove”) usually works better than a growing blacklist. Anthropic’s guidance also recommends telling the model what to do rather than only what to avoid.

Example 3: Eight email subject lines

text

Write 8 subject lines for an email about [TOPIC] sent to [AUDIENCE].Use a different technique for each and label it in parentheses:1. Curiosity gap  2. Specific number  3. A question they haven'talready answered  4. Direct benefit  5. What to avoid6. Social proof I can back up: [PROOF]  7. Genuine deadline only ifone exists: [DEADLINE OR "none"]  8. Reader's own first-person voiceUnder 50 characters each. No fake "RE:" or "FW:". No all-caps.

Why it works: forcing a named technique per line prevents eight rewordings of the same idea. The character limit is a convention for mobile inboxes, not a rule any email client enforces, so test against your own open-rate data.

Example 4: Rewrite in a specific voice

text

Rewrite the paragraph below for a newsletter read by independentfreelancers in their thirties. The current draft sounds like a pressrelease.Target voice: direct, slightly dry, knowledgeable without beingsuperior, like a friend who has freelanced for ten years. Usecontractions. Mix long and short sentences.Keep every fact and claim unchanged. Change voice, not meaning.If you must drop or soften a claim, tell me which one.[PASTE PARAGRAPH]

The fix: “Change voice, not meaning” plus a duty to report any dropped claim. Rewrites often quietly alter facts, and this line catches it.

2. Data and Analysis

Analysis prompts fail because they don’t say which decision the analysis serves. “Analyze this data” returns a summary. “Analyze this to decide X” returns something you can act on.

🔒 Before you paste: Don’t put customer personal data, confidential financials or anything covered by a contract into a consumer AI tool unless you have checked your organization’s policy and the provider’s data terms.

Example 5: Sales data

text

Analyze the data below to help me decide: [THE SPECIFIC DECISION].Data: [PASTE OR DESCRIBE]Tell me:1. Which categories are growing and declining, with percentage changes   you calculated from the data (show the arithmetic).2. Whether each change comes from volume, price, or mix.3. One hypothesis for the cause, using only what the data shows.4. What additional data would confirm or kill that hypothesis.Label each finding as OBSERVED (in the data) or INFERRED (my reasoning).Do not present a correlation as a cause. If the dataset is too small tosupport a conclusion, say so.

Why it works: the OBSERVED/INFERRED label is the most useful line. It makes the model separate what the data says from what it is guessing, which is the failure that makes AI analysis risky to forward to a boss.

Example 6: Customer feedback

text

Analyze these [NUMBER] reviews/survey responses for [PRODUCT]:[PASTE]Extract, quoting customers verbatim (with the review number):1. The top 3 reasons they bought.2. The top 3 hesitations before buying.3. Benefits mentioned that our marketing never claimed.4. Phrases that repeat across reviews.5. The moment customers realized the product was right for them.Quote only text that appears above. If a theme appears in fewer thanthree reviews, mark it "weak signal".

Why it works: requiring review numbers makes every quote checkable, and the “weak signal” rule stops one loud reviewer from becoming a “theme.”

3. Brainstorming

Ask for arguments and angles, not topics. “Ten blog ideas about productivity” returns category names. “Ten blog post arguments” returns theses you can disagree with.

Example 7: Content ideas with a point of view

text

Generate 10 blog post ideas on productivity for remote workers. Eachneeds a specific argument, not just a topic.For each idea give:- Working title (under 65 characters)- The core claim in one sentence (something a reader could disagree with)- The moment a reader would search for this (what just happened to them?)- Angle type: counterintuitive / underserved / common question with a  better answerSkip "ultimate guide" and "everything you need to know" formats, andskip any idea that would be interchangeable with a top-10 listicle.

Example 8: Naming

text

Generate 20 name options for [WHAT IT IS].Audience and personality: [DESCRIBE]Avoid: [TONES, CATEGORIES, NAMES YOU ALREADY REJECTED]Four groups of five: Descriptive, Evocative, Invented word, Unexpected.For each name give one reason it works and one weakness.Then pick your top 3 and defend them. At least one must be anon-obvious choice.

Warning: AI-generated names collide with existing brands more often than you would like. Check trademark databases and domain availability before you fall in love with one.

Example 9: Problem reframing

text

My problem as I currently understand it: [DESCRIBE]Reframe it five ways, each opening a different category of solution:1. Flip it: is the problem a symptom of the opposite issue?2. Which unrelated industry has already solved something similar?3. What if the main constraint disappeared?4. What changes if the deadline is one day versus five years?5. What does the affected person actually want, versus what they asked for?For each: the new problem statement in one sentence, plus one solutiondirection it suggests.

4. SEO

The most common SEO prompt mistake is asking for “keyword-rich” content. Google’s own guidance says to create helpful, people-first content and not to write primarily for search engines (Google Search Central). Ask instead for content that fully answers the searcher’s question and the follow-up they haven’t typed yet.

Google also states that using AI to produce content is not against its guidelines by itself; using automation mainly to manipulate rankings is (Google on generative AI content). What matters is whether the page is accurate and useful.

Example 10: FAQ blocks

text

Write 8 FAQ items for a page about [TOPIC], main keyword [KEYWORD].Each question: phrased the way a real person types it into a searchbox, not formal.Each answer: 40–60 words, direct answer in the first sentence, fullyunderstandable on its own. No "it depends" opener unless the nextclause says what it depends on.Use only facts from the source text below. If the source doesn't answera question, leave it out instead of guessing.Source: [PASTE YOUR OWN VERIFIED CONTENT]

A note on the 40–60 word rule: This is an industry heuristic drawn from observing featured snippets. Google does not publish a word count for them, and its documentation says snippets are generated automatically from page content (Google featured snippets). Treat it as a useful drafting target, not a ranking rule.

Example 11: Meta descriptions

text

Write 5 meta descriptions for a page about home espresso setups,target keyword "home espresso setup".- Aim for about 150 characters; show the count for each.- Put the keyword early.- A different hook for each: benefit, question, proof, instruction,  contrast.- Describe what the page actually contains. No promises the page  doesn't keep.- Searcher intent: comparing options before buying a machine.

The fix: the old “148–155 characters” rule is folklore. Google says it may show a different snippet than your meta description and doesn’t limit length by characters; display is truncated by pixel width, and the practical range is a rough guide (Google snippet documentation). The line that matters most is “no promises the page doesn’t keep,” because mismatched snippets are what cause bounces.

5. Coding

Coding prompts need three things people skip: the language and version, where the code runs, and what “done” means.

Example 12: A function with error handling

text

Write a Python 3.11 function that [WHAT IT DOES].Context:- Runs in: [AWS LAMBDA / LOCAL SCRIPT / DJANGO VIEW / ...]- Input: [FORMAT AND TYPICAL EDGE CASES]- Output: [EXACT RETURN SHAPE]Requirements:- Type hints and a docstring (Args, Returns, Raises).- Handle these failure cases explicitly: [LIST].- Raise specific exceptions, never a bare `except:`.- Three pytest tests: one normal case, one edge case, one failure case.- Standard library only, plus: [ALLOWED PACKAGES].If any requirement is ambiguous, state your assumption at the topinstead of silently choosing.

Why it works: asking for tests forces the model to think about edge cases, and the assumptions line surfaces choices you would otherwise discover in production.

⚠️ Always run the code. A model can produce tests that pass because they test the wrong thing. Read the tests, not just the green checkmarks.

Example 13: Debugging

text

This [LANGUAGE + VERSION] code raises: [EXACT ERROR MESSAGE, INCLUDING TRACEBACK][PASTE MINIMAL CODE THAT REPRODUCES IT]1. Say what the code is trying to do.2. Explain where and why the error occurs.3. Give the smallest fix that solves it.4. List any other problem you notice that could fail silently, separately,   so I can decide whether to act on it.Show corrected code apart from the explanation.

Why it works: “smallest fix” prevents a full rewrite you didn’t ask for, and listing extra issues separately keeps the fix reviewable.

6. Translation and Localization

“Translate this into Spanish” gets you correct Spanish. Specifying the country, register and reading context gets you Spanish someone actually uses.

Example 14: Localized translation

text

Translate the text below into French for professionals in France readinga business newsletter. Adapt idioms, cultural references and registerto what a French professional would naturally write. Do not translateword for word.List each phrase where a literal translation would have sounded foreign,and what you changed.[PASTE TEXT]

Limitation: for legal, medical or contractual text, have a qualified human translator review the output. A fluent-sounding error is worse than a clumsy correct one.

Example 15: Multilingual keyword ideas

text

I'm adapting a page about [TOPIC] for [LANGUAGE(S)].For each language, give 5 search phrases a native speaker wouldplausibly type (not direct translations), note regional differences(for example Brazilian vs. European Portuguese), flag terms that arecorrect but unnatural as searches, and name one local concern theEnglish page doesn't address.Table: Language | Phrase | Regional note | Local angle

Warning: the model cannot see real search volume. Treat these as candidates and validate them in Google Search Console, Google Keyword Planner or another keyword tool before building pages around them.

7. Research and Summarization

The rule for every prompt in this section: use AI to work on material you provide. Models can and do invent sources, quotes and numbers. Unless you are using a tool with live web access, and even then, verify every specific claim against the original.

Example 16: Summarize a long document

text

Summarize the document below for [AUDIENCE, e.g. "an executive with threeminutes"]. Answer exactly these four questions and nothing else:1. What is this document about? (1 sentence)2. What is its most important finding or recommendation? (2–3 sentences)3. What action does it require from the reader, if any? (1–2 sentences)4. What is missing, unresolved, or weakly supported?Under 200 words. No background the reader already knows, nosection-by-section recap. Only use content from the document.[PASTE DOCUMENT]

Why it works: fixed questions stop the proportional recap. Question 4 is the valuable one, because summaries are usually too flattering to their source.

Example 17: Competitive synthesis

text

Below are descriptions of three competitors, written by me. Use only whatI've written; do not add facts about these companies from memory.A: [NAME] — [POSITIONING, STRENGTHS, WEAKNESSES]B: [NAME] — [...]C: [NAME] — [...]Us: [WHAT WE DO, CURRENT POSITIONING]Answer:1. What assumption do all three share?2. Which customer segment is underserved by all of them?3. What is the most defensible difference we could build?4. Where could a competitor copy us fastest?Take a position on each. Avoid "may be worth considering."Mark any conclusion that depends on a fact I didn't provide.

8. Techniques: What Still Earns Its Place

Most prompt guides present chain-of-thought, few-shot and iteration as universal upgrades. The evidence is more specific than that, and it has shifted as models changed.

My position: for current reasoning-capable models, “think step by step” is mostly ritual. What still pays is asking the model to expose assumptions and evidence in the answer, as in Example 18. That gives you something to audit, which a longer hidden reasoning chain does not.

Example 18: Assumptions first (the modern replacement for “think step by step”)

text

Question: [YOUR DECISION OR PROBLEM]Context: [BACKGROUND, CONSTRAINTS, WHAT YOU'VE ALREADY TRIED]Before recommending anything:1. List the assumptions you are making that I haven't confirmed.2. List what information would change your answer.3. Give your recommendation and the single strongest reason for it.4. Give the strongest argument against your recommendation.I will check the assumptions myself, so be explicit.

Why it works: you can verify assumptions in a minute. You cannot easily verify a long chain of reasoning.

Example 19: Few-shot style matching

text

Here are [2–3] examples of the style and structure I want:EXAMPLE 1: [PASTE]EXAMPLE 2: [PASTE]Note what they share: sentence rhythm, how they open, how much evidencethey use, what they never say.Now write about [NEW TOPIC] with the same structure and register.Different topic, same craft. Do not reuse phrases from the examples.

Why it works: examples convey style more reliably than adjectives like “warm but authoritative.” The caveat is that models imitate flaws as well as strengths, so use your best work, and add the no-reuse line.

Example 20: A two-step refinement loop

Instead of one giant prompt, send two follow-ups after the first draft:

text

Step 1 (after the draft):Find the three weakest sentences: vague, passive, or claimingsomething without support. Rewrite only those three and tell me whyeach was weak.Step 2:Find general claims that could become a specific example, number ornamed scenario. Replace only those, and flag any place where youwould need a real source I haven't supplied.

Why it works: narrow edits are easier to review than a full rewrite, and the “need a real source” flag stops the second pass from inventing specifics to sound concrete.

Quick Reference

Five mistakes that keep reappearing

1. Asking for facts the model can’t verify. Feed it your data, not your questions about the world.
2. Piling on adjectives. “Engaging, professional, warm” describes nothing. An example does.
3. Never stating the audience. The single most common omission.
4. Accepting the first output. Read it as a draft from a fast junior colleague.
5. Trusting confident tone. Fluent wording tells you nothing about accuracy.

What I did and did not verify

I rebuilt this guide from the previous version and checked the technique claims against the primary papers and the providers’ and Google’s own documentation, which are linked below. I did not run a new controlled comparison of these prompts across models for this update, so nothing here should be read as a benchmark. Results vary by model, version and settings, and models update frequently. Treat every prompt as a starting point and test it on your own task.

FAQ

What makes a good AI prompt? It supplies what the model can’t infer: audience, purpose, format, exclusions and the source material it may use.

Do longer prompts work better? Only when each line changes the output. Use the removal check: delete a line, and if you wouldn’t notice the difference, it wasn’t needed.

Do I still need “think step by step”? Rarely, for current reasoning models. Research suggests the benefit has shrunk, so ask for assumptions and evidence in the output instead.

Can I use the same prompt in ChatGPT, Claude and Gemini? Usually as a starting point. Each provider documents slightly different best practices, so test on your own task.

Is AI-written content bad for SEO? Google says the method of production isn’t the issue; unhelpful, inaccurate or manipulative content is.

How do I stop AI making up facts? Provide the source text, tell it to use only that, allow it to say “not in the source,” and verify anything specific.

Glossary

• Prompt: the instruction and context you give a language model.
• Few-shot prompting: including example inputs and outputs in the prompt.
• Chain-of-thought: asking a model to reason in steps before answering.
• Hallucination: fluent output that is factually wrong or invented.
• Context window: the amount of text a model can consider at once.

Sources, ranked by strength

Strongest: primary research

1. Wei et al. (2022), Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. https://arxiv.org/abs/2201.11903
2. Brown et al. (2020), Language Models are Few-Shot Learners. https://arxiv.org/abs/2005.14165
3. Sprague et al. (2024), To CoT or not to CoT? https://arxiv.org/abs/2409.12183
4. Meincke, Mollick, Mollick & Shapiro (2025), Prompting Science Report 2: The Decreasing Value of Chain of Thought in Prompting. https://arxiv.org/abs/2506.07142

Strong: first-party documentation (describes each vendor’s own recommendations) 5. Anthropic, Prompt engineering overview. https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview 6. OpenAI, Prompt engineering guide. https://platform.openai.com/docs/guides/prompt-engineering 7. Google, Gemini API prompting strategies. https://ai.google.dev/gemini-api/docs/prompting-strategies

Strong for SEO claims: Google Search Central 8. Creating helpful, reliable, people-first content. https://developers.google.com/search/docs/fundamentals/creating-helpful-content 9. Google Search’s guidance on using generative AI content. https://developers.google.com/search/docs/fundamentals/using-gen-ai-content 10. Featured snippets and your website. https://developers.google.com/search/docs/appearance/featured-snippets 11. Control your snippets in search results. https://developers.google.com/search/docs/appearance/snippet

Removed from the previous version: vendor-marketing blogs (Guru, Vendasta, Lakera) and a product-growth newsletter, which restated the same advice without primary evidence.