The fastest way to get better answers from any AI is to write better prompts, and good prompts share five parts: a clear role, a specific task, the relevant context, the exact output format you want, and one or two examples of what good looks like. Add a step-by-step instruction for anything that involves reasoning, and you will outperform 90% of users without learning a single trick beyond that.
Most disappointing AI output is not the model's fault, it is a vague request. 'Write me a marketing email' leaves the model to guess your product, audience, tone, and goal, so it returns something generic. This guide breaks down the prompt structure that works across ChatGPT, Claude, Gemini, and the rest, the techniques that reliably lift quality, the mistakes that quietly sabotage results, and copy-paste templates you can adapt today.
The Anatomy of a Prompt That Works
Think of a strong prompt as five ingredients. You do not need all five every time, but the more of them you include for an important task, the better the result.
| Ingredient | What it does | Example phrase |
|---|---|---|
| Role | Sets the model's perspective and expertise | 'You are a senior copywriter for SaaS startups.' |
| Task | States exactly what to produce | 'Write a 120-word cold email.' |
| Context | Gives the facts only you know | 'Audience: HR managers at 50-person firms.' |
| Format | Controls the shape of the output | 'Use a subject line, 3 short paragraphs, one CTA.' |
| Examples | Shows the model the target quality | 'Match the tone of this sample: ...' |
Technique 1: Be Specific, Then Be More Specific
Specificity is the single biggest lever. Replace every vague word with a concrete one. Not 'short' but '120 words.' Not 'professional tone' but 'plain, confident, no jargon.' Not 'some examples' but 'three examples, each one sentence.' The model cannot read your mind, and every detail you leave out is a detail it fills in with an average guess. The more constraints you give, the less averaging it has to do.
Technique 2: Show, Don't Just Tell (Few-Shot)
If you can show one or two examples of what you want, do it. This is called few-shot prompting, and it is one of the most reliable quality boosts available. Paste a sample email you liked and say 'write three more in this style,' or give an input-output pair and ask the model to continue the pattern. Examples communicate tone, structure, and standards faster than any description, because the model is extremely good at matching patterns it can see.
Technique 3: Ask for the Reasoning, Step by Step
For anything involving logic, math, planning, or analysis, add 'think through this step by step before giving the final answer.' Letting the model lay out its reasoning before committing to a conclusion measurably improves accuracy on multi-step problems, because it stops the model from blurting a first guess. Modern reasoning models do some of this automatically, but the instruction still helps on harder tasks and on faster, cheaper models.
Technique 4: Assign a Role and a Goal
Telling the model who to be focuses its vocabulary, assumptions, and priorities. 'You are a patient tutor explaining to a beginner' produces a very different answer than 'you are an expert reviewer being blunt.' Pair the role with the goal: 'Your goal is to make this clear to someone with no background.' Role plus goal turns a generic assistant into a specific collaborator.
Technique 5: Iterate Instead of Restarting
The best results almost never come from the first prompt. Treat the answer as a draft and steer it: 'Make the second paragraph shorter.' 'Too formal, loosen it up.' 'Keep the structure but change the examples to retail instead of tech.' Conversation is the feature, not a workaround. People who get great output are usually people who refine three or four times, not people who write one perfect prompt.
The official prompting guides from the major labs all converge on the same advice: be explicit, give examples, specify the format, and let the model reason before it answers. None of it is secret. Almost no one does all four.
The Mistakes That Quietly Ruin Prompts
- Asking two things at once: bundling 'summarize this and also write a tweet and also fix the grammar' into one prompt produces a worse version of each. Do them in separate turns.
- Leaving out the audience: 'explain this' with no audience gets you an answer pitched at no one. Always say who it is for.
- Never giving an out: tell the model 'if you are not sure, say so' to reduce confident-but-wrong answers (hallucinations).
- Vague quality words: 'make it good' or 'make it engaging' mean nothing to a model. Translate them into concrete instructions.
- Forgetting the format: if you do not specify a table, list, or word count, you get a wall of prose you then have to reshape.
Three Templates You Can Steal
- Writing: 'You are a [role]. Write a [length] [type] for [audience]. Goal: [goal]. Tone: [tone]. Must include: [points]. Format: [structure]. Here is an example I like: [sample].'
- Analysis: 'Here is [data or text]. Think step by step, then give me: 1) the three most important takeaways, 2) anything that looks wrong or missing, 3) one recommendation. If you are unsure about a fact, flag it.'
- Learning: 'Explain [topic] to me like I have no background. Use one everyday analogy, then give a concrete example, then check my understanding with one question.'
Do Different Models Need Different Prompts?
The fundamentals are universal, but the feel differs. Some models respond especially well to structure and explicit roles, others to conversational nudging, and the only way to learn the difference is to run the same prompt through a few and compare. A tool like LumiChats, which puts 40-plus models in one place, makes that comparison painless, so you can see which model answers your kind of question best before you settle on one.
01What makes a good AI prompt?
A good prompt includes a clear role, a specific task, the context only you know, the exact output format you want, and ideally one or two examples of good output. For reasoning tasks, also tell the model to work step by step before answering.
02What is few-shot prompting?
Few-shot prompting means giving the model one or more examples of the input-and-output you want before asking it to do the real task. Examples communicate tone, structure, and quality faster than descriptions, and they reliably improve results because the model is excellent at matching patterns it can see.
03Does saying 'think step by step' actually help?
Yes, for tasks involving logic, math, or multi-step analysis. Asking the model to lay out its reasoning before the final answer improves accuracy by stopping it from committing to a quick first guess. It helps most on harder problems and on faster, cheaper models.
04Why does the AI give generic answers?
Almost always because the prompt was generic. If you do not specify the audience, tone, length, and format, the model fills the gaps with average guesses. Add concrete constraints and the output sharpens immediately.
05Should I use different prompts for ChatGPT, Claude, and Gemini?
The core technique is the same for all of them, but each model has a slightly different feel, so the best approach is to test the same prompt across a few and keep the phrasing that works for your tasks.
Prompting is not a dark art, it is a habit. Be specific, show an example, ask for reasoning, set a role, and iterate. Build those five into how you ask, and every AI you touch gets noticeably better, today and as the models keep improving.
