Short answer: better AI output comes from giving the model context, a clear role, concrete examples and explicit constraints — not from memorizing "magic" prompt phrases. The same handful of patterns work across ChatGPT, Claude, Gemini and almost any chatbot, because they address the real bottleneck: a language model can only work with what you put in front of it. This guide gives you those patterns, shows which ones matter most for which jobs, and hands you example phrasings you can copy today.
The core insight is simple. The AI isn't reading your mind; it's predicting the most probable useful response to your words. Vague prompts get vague, average answers. Specific prompts get specific, useful ones. Almost every "the AI is bad at this" complaint is really an under-specified prompt — and the fix is rarely a clever incantation. It's clearer communication.
How we evaluated the techniques
There is no shortage of "100 ChatGPT prompts" lists. What is missing is an honest sense of which moves actually change output quality versus which are folklore. To keep this grounded, we leaned on three things rather than vibes:
- The official vendor guidance. The major labs publish prompt-engineering documentation that broadly agrees on the fundamentals. See OpenAI's prompt engineering guide, Anthropic's Claude prompt engineering overview, and Google's Gemini prompting strategies. When three competing labs independently recommend the same move, it's signal, not fashion.
- Repeatability across tasks. A technique earns a spot here only if it helps on more than one kind of work — writing, coding, analysis, extraction — not just one demo.
- Cost of getting it wrong. Some techniques are high-impact and nearly free to apply (adding context). Others are powerful but easy to overdo (stuffing in ten examples). We weighted by impact-per-unit-effort, which is what the charts below capture.
A quick word on what the scores are not: they are qualitative, directional weightings drawn from the guidance above and hands-on use, not benchmark numbers. Prompting quality is genuinely hard to measure precisely because it depends on the task, the model and the user. Treat the visuals as a map of where to spend your attention, not a leaderboard.
The five patterns that matter most
If you only internalize five things, make it these. They cover the large majority of everyday wins.
1. Give it context
Tell the model who the output is for, what it's about, and why it exists. Compare:
Weak: "Write a product description for my candle."
Strong: "Write a 60-word product description for a hand-poured soy candle called 'Rainy Sunday.' Audience: gift shoppers aged 25 to 40 who value calm and self-care. Tone: warm and sensory. Mention the lavender-and-cedar scent and the 40-hour burn time. Avoid clichés like 'perfect gift'."
The second version removes the guesswork. Context is the single highest-return thing you can add, and it's almost free. If you're using AI to draft long-form content, the same rule scales up — our walkthrough on how to use AI to write blog posts is mostly an exercise in front-loading context before you ask for a single paragraph.
2. Assign a role
Telling the model who to be focuses its knowledge and tone. "You are an experienced copy editor specializing in plain-English business writing" produces sharper editing than no role at all. Roles work because they anchor both the style and the depth of the response, and they prime the model to draw on the relevant slice of what it knows. Be specific: "a B2B SaaS copywriter" beats "a writer," and "a tax accountant who explains things to non-experts" beats "an expert."
3. Show examples (few-shot)
If you want output in a particular format or voice, show one or two examples of what good looks like. This is the single most underused technique. "Rewrite these in the style of the two examples below" beats any amount of describing the style in the abstract, because the model can pattern-match a concrete sample far more reliably than it can interpret an adjective like "punchy." Two or three examples is usually the sweet spot; past that you get diminishing returns and a longer, slower prompt.
4. Set explicit constraints
State length, format, what to include, and what to avoid. "Answer in exactly three bullet points, under 15 words each, no jargon" gives the model a target to hit. Without constraints you get the model's default register, which tends to be too long, too hedged and weirdly fond of the word "delve." Negative constraints matter too: telling it what not to do ("don't invent statistics," "don't use exclamation marks") is often as useful as the positive instruction.
5. Ask for the format you'll actually use
Need a table? Ask for a table. Need JSON, a checklist, an email, a tweet thread? Say so. "Output as a markdown table with columns Name, Pro, Con" saves you reformatting and quietly forces the model into more structured thinking. If you're piping the result into a spreadsheet or another tool, name the exact schema; this is the difference between a tidy import and an afternoon of cleanup, as anyone who has tried to get AI to fill a spreadsheet will recognize.
Which technique for which job
The five patterns aren't equally important for every task. Asking for a strict output format matters enormously when you're extracting structured data and far less when you're brainstorming. The matrix below shows where each move pulls its weight.
| Technique | Writing | Coding | Analysis | Data extraction |
|---|---|---|---|---|
| ★Give context | ✓ | ✓ | ✓ | ✓ |
| Assign a role | ✓ | ~ | ✓ | ✕ |
| ★Show examples (few-shot) | ✓ | ✓ | ~ | ✓ |
| Set constraints | ✓ | ✓ | ~ | ✓ |
| Request exact format | ~ | ✓ | ~ | ✓ |
| Step-by-step reasoning | ~ | ✓ | ✓ | ✕ |
The pattern is worth saying out loud: context helps everywhere, examples and constraints help almost everywhere, and the more specialized moves (strict formatting, explicit reasoning) earn their place on technical and structured tasks. If you're unsure where to start, start with context.
Patterns for harder tasks
Once the basics are second nature, a few more advanced moves handle the jobs that trip people up.
Break complex jobs into steps
For multi-part work, ask the model to reason step by step, or tackle one stage at a time. "First outline the structure. I'll approve it, then you write each section." Staging beats asking for everything at once and getting a mediocre blob you then have to dismantle. For reasoning-heavy problems, simply adding "think step by step before giving your answer" measurably improves accuracy on the kinds of tasks where the model would otherwise blurt out a wrong conclusion in the first sentence.
Ask it to ask you
A genuinely powerful move: "Before you answer, ask me any questions that would help you give a better response." This surfaces the context you forgot to provide and prevents wasted generations. It is especially useful at the start of a big task, when you don't yet know what you don't know. The model frequently asks the exact question that reveals the real requirement.
Give it room to think
For analysis or judgment calls, let the model work through the problem before committing to an answer, rather than demanding the conclusion in line one. Quality improves when it doesn't have to be right immediately. This is the same principle behind the "reasoning" modes now built into the major chatbots — you're just invoking it manually.
Iterate instead of restarting
Treat the first output as a draft, not a verdict. "Good, but make it shorter and cut the second point" refines faster than rewriting the prompt from scratch. The conversation is the tool. Most people give up after one underwhelming response when two follow-ups would have gotten them there. The skill isn't writing one perfect prompt; it's steering.
Ground it in your own material
When facts matter, paste in the source material and tell the model to use only that. "Using only the document below, answer the question. If the answer isn't there, say so." This is the manual version of what retrieval-augmented tools do automatically, and it's the most reliable defense against confident-sounding invention. Pairing this with a research-grade tool helps too — our Perplexity vs Gemini comparison digs into which assistants cite real sources versus which improvise.
Where to spend your effort
Not every technique costs the same to apply. Adding a line of context is trivial; building a clean few-shot example set takes real work. The map below plots impact against effort so you can pick your battles.
The takeaway from both charts is the same and a little deflating for prompt-hack enthusiasts: the cheapest moves are also the most valuable, and the elaborate ones (bribes, ALL-CAPS threats, "you are the world's greatest expert") sit in the dead zone. Spend your effort on context and a couple of good examples before anything else.
A quick reference you can keep
| Goal | Pattern | Example phrase |
|---|---|---|
| Relevant output | Add context | "Audience is X, tone is Y, purpose is Z" |
| Right voice | Assign a role | "You are a senior X who writes for Y" |
| Specific format | Show examples | "Match the style of these two samples" |
| Controlled length | Set constraints | "Under 100 words, no jargon, no clichés" |
| Usable structure | Request format | "Output as a markdown table with columns A, B, C" |
| Complex tasks | Stage the work | "Outline first, then I'll approve, then write" |
| Missing info | Let it ask | "Ask me questions before you answer" |
| Factual accuracy | Ground it | "Use only the text below; say so if unsure" |
Do the patterns differ between ChatGPT, Claude and Gemini?
This is the question we get most, and the honest answer is: far less than the marketing implies. The five core patterns work everywhere because they address how language models use input in general, not a quirk of one product. You do not need a separate "skill" for each chatbot.
That said, a few real differences are worth knowing:
- Long context and document handling. Some models accept enormous inputs, which makes the "paste the whole source and ground the answer" technique more practical. If you routinely feed in long documents, that capability matters more than any prompt trick.
- System prompts and custom instructions. Most tools let you set persistent instructions (a standing role, tone and constraints) so you don't repeat them every time. Setting these once is one of the highest-leverage things you can do.
- Built-in reasoning modes. Several chatbots now have an explicit "think harder" mode. When available, it can replace the manual "reason step by step" instruction for tough problems.
- Tone defaults. Each model has a house voice out of the box. Your constraints and examples override it, which is exactly why specifying them matters more than choosing the "right" model.
If you're picking a writing-focused tool rather than a raw chatbot, the prompting still matters but the surrounding workflow matters more — see our Jasper review and Notion AI review for how purpose-built tools bake some of these patterns into templates so you don't have to type them every time.
Mistakes to stop making
- Treating prompts as one-shot. The best results come from a short back-and-forth, not the perfect single sentence. If you only change one habit, make it this one.
- Being polite-but-vague. "Can you maybe help with some marketing stuff?" wastes the model's capability. Direct and specific is kinder to both of you.
- Over-stuffing. Cramming ten unrelated requests into one prompt produces a muddy answer. One clear job at a time, then move on.
- No success criteria. If you can't say what "good" looks like, the model can't hit it. Define it before you press enter.
- Chasing magic words. "Act as the world's best expert and you'll be tipped 1000 dollars" is folklore. Clear context and examples do the real work.
- Trusting facts you didn't ground. Models will produce confident, fluent, wrong statements. For anything load-bearing, supply the source or verify the claim. If you're on the other side of this — checking whether something was machine-written — our guide on how to detect AI-generated text covers the tells and the limits of detectors.
The bottom line
Prompting well isn't a secret skill, and it definitely isn't a list of clever phrases. It's clear communication applied to a literal-minded, eager-to-please assistant that can't read your mind. Give it context, set a role, show examples, state your constraints, and refine through iteration. Those few patterns transfer across every major AI tool and will do more for your output than any "ultimate prompt pack" you can buy.
The practical test: write the prompt you'd give a sharp new hire on their first day — someone capable but with zero knowledge of your situation, your audience or your standards. Tell them who it's for, show them an example of what good looks like, set the boundaries, and be ready to give one round of feedback. Do that, and you'll get most of the way there with any model, today and after the next dozen releases.