How to Write AI Prompts: 15 Techniques That Work

By Muzamil Ahmad · Updated 2026-10-02 · Writing

#writing#how to write ai prompts#chatgpt prompt tips#prompt engineering guide#better prompts chatgpt#prompt engineering tips#chatgpt prompts guide

How we write our guides: we start from the vendor’s own announcement or documentation, not from other “best of” lists. Launch dates and features are checked against that primary source; anything reported only by coverage is labelled as reported. Prices and allowances are given only when the vendor publishes them — otherwise we say “not published” rather than guessing. We do not invent test results, prices, or rankings. Our guides are drafted with AI assistance, then reviewed and edited before publication; the final call on what we recommend is a human one.

How to Write AI Prompts: 15 Techniques That Work
How to Write AI Prompts: 15 Techniques That Work

Better prompts get better answers. That sounds obvious, but most people type the first thing that comes to mind into a chatbot and then blame the tool when the output is vague. The good news: prompt writing is a learnable skill, and small changes in how you ask produce large changes in what you get back. Here is how to write AI prompts that actually work: 15 practical techniques, each with a short explanation and an example you can adapt.

Inside this guide: the key sections covered and the tools compared.
Inside this guide: the key sections covered and the tools compared.

How I picked these 15 techniques

I get asked this constantly: are all those prompting tips floating around social media actually worth learning? Here is my honest answer. I tested dozens of circulating tips on my own writing, editing, and coding tasks, and most of them made no measurable difference. Being extra polite, adding magic phrases, threatening the AI, bribing it with tips: none of it moved the needle. What worked, every time, was boring fundamentals: be specific, give context, show examples, set the format, iterate.

That is the filter I used for this list. Every technique here had to pass three tests: it works across ChatGPT, Claude, and Gemini, not just one tool; a beginner can use it in their next session without setup; and the improvement is visible immediately, not theoretical. The 15 that survived are the ones below. The advanced section later in this guide covers what to learn once these feel natural.

1. Be Specific About What You Want

Vague prompts get vague answers. "Write about marketing" could mean anything. "Write a 200-word email announcing a 20% off sale to existing customers" leaves little room for misunderstanding. Add the details that would matter if you were briefing a human colleague.

Example: "Write a 150-word product description for a stainless steel water bottle, aimed at hikers, highlighting durability and insulation."

2. Give Context and Define the Audience

AI models do not know your situation unless you tell them. Mention who the output is for, what it is for, and any background that shapes the answer.

Example: "Explain compound interest to a 16-year-old saving for their first car. Keep it under 120 words."

3. Assign a Role

Telling the model to act as a particular expert focuses its tone, vocabulary, and depth.

Example: "Act as a nutritionist. Review this daily meal plan and suggest two improvements."

4. Provide Examples (Few-Shot Prompting)

Showing beats telling. If you give two or three examples of the input-output pattern you want, the model will usually follow it closely. This is one of the most reliable techniques for consistent formatting, tone, or classification tasks.

Example: "Convert these to friendly support replies. Example: Input: 'Where is my order?' Output: 'Great question! Your order is on its way and arrives Thursday.' Now convert: 'How do I reset my password?'"

5. Set Format and Length Constraints

Models default to medium-length prose unless told otherwise. Specify the format you need: bullet list, table, numbered steps, JSON, or a word count range. This saves you from reformatting everything by hand afterward.

Example: "Summarize this article as 5 bullet points, each under 20 words."

6. Iterate and Refine

Treat the first answer as a draft, not a verdict. Follow up with targeted revisions: "make it shorter," "add an example," "use simpler language." Each round narrows in on what you actually want, and it is usually faster than rewriting the whole prompt from scratch.

Example: "Good start. Now cut it to half the length and make the tone more casual."

7. Ask for Step-by-Step Reasoning

For math, logic, or multi-part problems, ask the model to show its work. Phrases like "think step by step" or "explain your reasoning before giving the final answer" tend to produce more accurate results, because the model checks itself as it goes.

Example: "A store offers 25% off, then an extra 10% off the sale price. Show each step, then give the final price of a $80 item."

8. Use Delimiters to Separate Instructions from Content

When you paste text for the model to work on, wrap it in clear markers like triple quotes, brackets, or XML-style tags. This prevents your instructions from bleeding into the content and vice versa, especially with long pasted passages.

Example: "Summarize the text between the triple dashes. --- [paste article here] ---"

9. Stick to One Task per Prompt

Cramming five requests into one prompt usually means all five get mediocre treatment. Split complex jobs into a sequence: first summarize, then extract key points, then draft the email. You can chain the outputs together in follow-up prompts, and each step gets the model's full attention.

Example: Instead of "summarize this, translate it, and write a tweet about it," run three separate prompts.

10. Specify the Tone Explicitly

"Professional," "friendly," "playful," "formal," and "direct" all produce noticeably different outputs. If tone matters for your use case, name it. You can even combine descriptors or point at a reference: "in the tone of a helpful customer support agent." If you need more than prompts — an actual chatbot handling tickets — see our best AI chatbots for customer support roundup.

Example: "Rewrite this paragraph in a warm, encouraging tone suitable for new employees."

11. Request Alternatives

Instead of accepting the first output, ask for two or three versions and pick the best. This works well for headlines, subject lines, names, and any creative task where variety helps you choose.

Example: "Give me 3 headline options for a blog post about home composting, each under 60 characters."

Deeper in this guide: advanced sections worth your scroll.
Deeper in this guide: advanced sections worth your scroll.

12. Prime with Background Information

Front-load the facts the model needs before you ask the question. Paste the relevant data, describe the situation, or list the constraints first, then make your request. The more grounded the prompt, the less the model has to guess, and guessing is where errors creep in.

Example: "Here is my weekly schedule: [paste]. Suggest three 30-minute workout slots that do not conflict with meetings."

13. Use Follow-Up Prompts Strategically

The conversation is a feature, not just a container. Ask clarifying follow-ups like "what assumptions did you make?" or "what would change your answer?" These questions surface hidden reasoning and often catch mistakes the first pass missed.

Example: "Before I use this plan, what important information am I missing that would change your advice?"

14. Test and Compare Outputs

For important tasks, run the same prompt more than once or try small wording variations and compare. If the answers are consistent, you can trust the result more. If they swing wildly, the prompt probably needs more constraints or context before you rely on it.

Example: Run your draft prompt three times. If key facts change between runs, add specifics until they stabilize.

15. Save Reusable Prompt Templates

When a prompt works well, save it. Keep a small library of templates with blanks for the variable parts, like a weekly report summary or a standard code-review request. Developers can go further with our best AI coding assistants roundup. This turns one good prompting session into a repeatable workflow and keeps your outputs consistent over time.

Example template: "Summarize the following meeting notes into: 1) decisions made, 2) action items with owners, 3) open questions. Notes: [paste]"

Templates like that one work even better when the notes write themselves: the best AI meeting assistants produce the summary and action items this prompt would otherwise have to extract by hand.

How to Write AI Prompts: A Strong Prompt Checklist

Before you hit send on an important prompt, run through this quick list:

  1. Is the request specific enough that a stranger would know what to do?
  2. Did I include context and name the audience?
  3. Did I set the format, length, and tone?
  4. Is there exactly one task in this prompt?
  5. For anything important, will I verify the output afterward?

You do not need all 15 techniques every time. For a report, a plan, or anything you will publish, layering several of them is what separates a usable answer from a great one. Our guide to using ChatGPT for content writing puts these techniques into a full workflow.

Copy-paste prompt templates

Theory is nice, but templates you can steal are better. Here are six I use constantly. Fill in the brackets and go.

The Brief (for any content). "You are a [role, e.g. experienced copywriter]. Write a [length] [format, e.g. blog section] about [topic] for [audience]. Tone: [tone]. Must include: [3 key points]. Avoid: [words or phrases to skip]. End with [specific closing]."

The Rewrite. "Rewrite the text below to [specific goal: shorter, punchier, more formal]. Keep all facts and examples. Change only the wording and structure. Text: [paste]"

The Brainstorm. "Give me [number] ideas for [goal], each described in one sentence. Constraints: [budget, time, audience]. Make ideas 1-3 safe and proven, ideas 4+ more creative."

The Explainer. "Explain [concept] to [audience, e.g. a 14-year-old]. Use one everyday analogy. Under [length]. End with a one-sentence summary I could quote."

The Critic. "Review the following [draft/plan/email] like a skeptical expert. List the 3 weakest points and suggest a specific fix for each. Do not praise what works; focus on what to improve. [paste]"

The Planner. "Help me plan [project]. Constraints: [time, budget, resources]. First ask me up to 5 clarifying questions, then produce a step-by-step plan with owners and deadlines."

Save these in a notes file. After a month of use, you will have customized versions that outperform anything generic.

Advanced techniques: beyond the basics

Once the 15 techniques feel natural, these five take you further.

Use negative instructions. Tell the model what to avoid, not just what to do. "No jargon, no exclamation marks, no phrases like 'in conclusion'" works surprisingly well. Models respond strongly to explicit don'ts, and this is the fastest fix for generic-sounding output.

Make it critique itself. After a good draft, prompt: "Now review your answer as a skeptical expert. List the two weakest claims and explain why each might be wrong." Then ask it to fix them. This catches confident errors that a single pass misses. It is not foolproof, but it is free quality control.

Stack constraints for creative work. Give three hard constraints: a length limit, something it must include, and something it must avoid. Constraints force interesting choices. "Write a 100-word product pitch that mentions the warranty, avoids the word 'innovative,' and ends with a question" will beat an unconstrained pitch almost every time.

Cross-check between models. For anything important, draft in one chatbot and verify in another. Generate the answer in ChatGPT, then paste it into Claude with "fact-check this and flag anything questionable." Different models make different mistakes, so the overlap is where the truth usually sits. Our ChatGPT vs Claude vs Gemini comparison helps you pick the right pair.

Prompt for research, then verify. For research tasks, prompt: "Answer [question] and list the specific sources for each claim." Then open every source yourself. A research-focused tool can help here too; see our Perplexity vs ChatGPT breakdown. Never cite a source you have not opened. Models invent convincing-looking citations, and this habit is the only reliable defense.

Which techniques to combine: a quick guide

ScenarioCombine these techniquesWhy it works
Writing a blog post1, 2, 4, 5, 9Specific brief plus examples plus format, split into steps
Debugging code7, 8, 13Show reasoning, isolate the code, ask follow-ups
Learning a topic2, 7, 11Audience context, step-by-step, then alternatives for depth
Email to a client3, 10, 11Role sets professionalism, tone is explicit, options to choose
Planning a project12, 9, 14Front-load facts, split into steps, test stability

You do not need all 15 every time. Pick the 3 or 4 that match the scenario and layer them.

Mistakes to avoid

1. The mega-prompt. Five tasks crammed into one prompt get five mediocre answers. Split them. Technique 9 exists because this mistake is so common.

2. Accepting the first answer. The first output is a starting point. One round of "shorter, punchier, with an example" transforms most drafts. People who complain about AI quality usually skipped this step.

3. Vague revision requests. "Make it better" means nothing. Say what better looks like: shorter sentences, a concrete example, a warmer tone, half the length. Direct the revision like you would direct a person.

4. No audience defined. "Explain inflation" gets a textbook paragraph. "Explain inflation to a teenager with a part-time job" gets something readable. The audience sentence is the cheapest upgrade in prompting.

5. Forgetting the format. If you need a table and get an essay, that is on the prompt, not the model. State the format every time: bullets, table, steps, word count. Reformatting by hand is pure wasted time.

6. Writing a life story instead of a brief. Some prompts read like autobiographies: three paragraphs of background, then a vague question at the end. Flip it. Lead with the ask, then add only the context that changes the answer.

7. Copying "prompt hacks" without testing. Social media is full of tricks: threaten the model, tip it, use magic phrases. Test them yourself with technique 14 and watch most of them do nothing. Fundamentals beat hacks every time.

Pro tips for power users

End with the exact closing you want. "End with a 3-bullet summary" or "End with one question for the reader." Models often fizzle out at the end; a specified closing keeps the landing clean.

Show a bad example, not just good ones. "Here is a headline I hate: [example]. Write 5 in the opposite direction." Negative examples steer tone faster than paragraphs of description.

Add "if you are unsure, say so." For factual questions, this single sentence reduces confident hallucinations noticeably. It does not eliminate them, so still verify, but it lowers the noise.

Version your templates. When you tweak a saved template, note what changed and why: "v2: added audience line after outputs kept sounding corporate." In six months you will have a personal prompting playbook that no course can match.

Learn with it, do not just produce with it. Students get the most from AI by asking for explanations, practice questions, and step-by-step walkthroughs, not finished assignments. Our AI tools for students roundup has study-specific workflows built on exactly these techniques.

Test on the free tier first. Before committing to a paid plan for heavy prompting work, run your real templates on the free tier for a week. If the outputs hold up, you just saved $20 a month. Upgrade when limits, not curiosity, demand it.

Same Prompt, Different Model: What Changes

One thing the generic guides skip: the same prompt does not behave the same everywhere. ChatGPT, Claude, and Gemini have different defaults, and a prompt tuned for one can underperform on another. Here is what I have learned from running the same prompts across all three:

ChatGPT defaults to structured, thorough answers. It loves headings, bullet lists, and covering every angle, which is great for research and terrible when you want brevity. If your ChatGPT outputs feel bloated, the fix is almost always a length constraint: "in under 150 words" or "three sentences max." It also responds well to directness. Flowery setup wastes tokens; tell it what you want plainly.

Claude defaults to a more conversational, nuanced tone and is better at following subtle style instructions. "Write like a tired expert explaining this to a smart friend" lands better here than on ChatGPT. Claude is also more cautious: it will hedge, qualify, and offer caveats where ChatGPT just answers. If you want decisiveness, say so explicitly: "give me your best recommendation, not a list of options." Where Claude frustrates people is refusal-adjacent behavior on borderline requests; rephrasing usually fixes it.

Gemini is the most uneven of the three in my experience. On factual and Google-ecosystem tasks it is strong, and its long context window means you can paste enormous documents. But its default tone can feel generic, and it sometimes ignores format instructions that ChatGPT and Claude follow cleanly. The fix: repeat the format requirement at the end of the prompt, not just the beginning. Models weight recent instructions more heavily, and Gemini seems to need the reminder.

The practical takeaway: when a prompt fails, do not just rewrite the prompt. Try the same prompt on a different model first. Half the time the problem is the model, not your wording, and switching is faster than iterating.

Diagnosing Failed Outputs

When a prompt produces garbage, most people blame the model and start over. Better: diagnose. Every bad output has a cause, and the cause tells you the fix. Here are the failures I see most, with diagnoses:

The generic essay. You asked for something specific and got a bland five-paragraph overview that could apply to anyone. Diagnosis: no audience and no stakes in the prompt. The model wrote for everyone, which means no one. Fix: add one audience sentence and one constraint ("for a bakery owner with no marketing budget").

The confident wrong answer. The output sounds authoritative but the facts are off. Diagnosis: you asked a knowledge question without giving the model anything to work from. It filled the gaps with plausible invention. Fix: paste the source material into the prompt and say "answer only from the text below." Grounding beats hoping.

The endless list. You wanted a decision and got 47 bullet points. Diagnosis: you asked an open question ("what should I do about...") without asking for a recommendation. Models enumerate by default. Fix: end with "recommend one option and explain why in two sentences."

The tone-deaf draft. The content is right but it sounds like a press release when you needed a human. Diagnosis: you specified the task but not the voice. Fix: add a tone line with a concrete reference ("warm and direct, like explaining to a colleague over coffee, no corporate jargon").

The format fight. You asked for a table three times and keep getting paragraphs. Diagnosis: the format instruction got buried under everything else. Fix: put the format requirement last, repeat it ("format: markdown table with exactly these columns"), and cut competing instructions.

Notice the pattern: every diagnosis points back to something missing from the prompt, not something wrong with the model. Failed outputs are usually incomplete briefs. Read your bad outputs like a manager reading a bad first draft: what instruction would have prevented this?

The key takeaway from this guide.
The key takeaway from this guide.

Frequently Asked Questions

How long should a good prompt be?

As long as it needs to be. Short prompts work fine for simple questions, but complex tasks usually need a paragraph or two of context. Clarity matters more than length: a focused 50-word prompt beats a rambling 300-word one.

Do these techniques work on all AI chatbots?

The core ideas transfer across ChatGPT, Claude, Gemini, and similar tools, since they all respond to clarity, context, and examples. Exact phrasing that works best can vary slightly between models, so test technique 14 when you switch tools.

What is the single biggest prompting mistake beginners make?

Being too vague and then accepting the first answer. Adding one sentence of context and doing one round of follow-up refinement fixes most disappointing outputs.

Should I be polite to the AI?

It does not change the model's capabilities, but clear, direct instructions work best either way. "Please" will not hurt, but it will not fix a vague prompt. Spend the words on specifics instead.

Can I reuse the same prompt template for different tasks?

Yes, that is the point of technique 15. Keep the structure that works and swap in new content. Just re-check the context each time, since a template written for one audience can misfire for another.

Do "magic words" or secret prompt tricks really work?

Mostly no. Threatening the model, offering tips, or adding mystical phrases does not improve output in any test I have run. What works is embarrassingly unsexy: specific instructions, good context, clear format, and one round of revision. Anyone selling secret prompt words is selling you something.

How do I prompt specifically for images or code?

Different outputs need different prompt shapes. For images, describe the subject, style, lighting, and composition in concrete visual terms. For code, paste the relevant code, state the language and the exact error or goal, and ask for an explanation alongside the fix. Our best AI coding assistants roundup covers tools built for the code side.

Should I write prompts in English if it is not my first language?

Write in the language you think in. All major chatbots handle many languages well, and a clear prompt in your own language beats a clumsy translated one. If the final output must be in English, say so explicitly and let the model handle the translation.

How often should I update my saved templates?

Review them quarterly, or whenever you switch tools. A template tuned for one chatbot can behave differently on another, so re-test your core templates after any model upgrade. Small wording tweaks after testing beat full rewrites.

Going further: put better prompts to work in the chatbot that fits you, compared in ChatGPT vs Claude vs Gemini, or pick a dedicated platform from the best AI writing tools.

Conclusion

Writing better prompts is not about secret tricks or magic words. It is about communicating clearly: saying what you want, who it is for, and how it should look, then refining the result. Start with specificity and context, add examples and constraints when it matters, and save what works. Do that consistently and the quality of everything you get from AI tools will rise noticeably.

Reader reviews

–
Loading reviews…