Prompting Basics: How to Ask AI for What You Actually Want
Market & numbers — every figure sourced
Prompting Basics: How to Ask AI for What You Actually Want
Most people get mediocre answers out of AI for one reason: they treat it like a search box instead of a brand-new contractor who is smart, fast, eager, and completely without context. The model does not know your business, your audience, your constraints, or what "good" looks like to you. Your prompt is the entire briefing. Get the briefing right and the output quality jumps without you touching the model at all.
This is the highest-leverage skill in the whole AI stack. You do not need a bigger model. You need a better ask.
The mental model: AI is a new hire on day one
Imagine you hired a brilliant generalist this morning. If you walk by their desk and say "write something about our product," you get generic mush. If you hand them a one-page brief with the audience, the goal, the format, two examples of what you like, and a hard constraint, you get something usable. Same person. Different brief.
Everything below is just how to write that brief.
The four ingredients of a strong prompt
Both major model providers converge on the same core components. A well-engineered prompt almost always includes a clear instruction (the role and the goal), context (the background the model needs), examples (samples of what good looks like), and a defined output format. OpenAI and Anthropic both teach this same skeleton in their official guides (OpenAI, Anthropic).
1. Be specific. Vague in, vague out.
This is the single biggest lever. The more specific you are about context, outcome, length, format, and style, the better the result. OpenAI's own guidance uses this exact contrast: instead of "Write a poem about OpenAI," write "Write a short inspiring poem about OpenAI, focusing on the recent DALL-E product launch, in the style of a classic poet" (OpenAI best practices).
Weak: "Write a marketing email."
Strong: "Write a 120-word cold email to a small-business owner who runs a local HVAC company. Goal: book a 15-minute call. Tone: direct, no fluff, no exclamation points. End with one clear call to action."
2. Give context the model cannot guess
Tell it who you are, who the reader is, and any facts it needs. The model has no access to your situation. If the answer depends on your industry, your price point, your region, or last week's launch, you have to say so. Anthropic frames the whole job as "context engineering": designing the entire context window (instructions, documents, examples, and the query) as one deliberate package (Anthropic).
3. Show examples (this is the cheat code)
Telling the model what you want is good. Showing it is better. This is called few-shot prompting: you paste 2 to 5 examples of input-and-ideal-output before your real request. Examples clarify subtle requirements that are hard to put into words, like tone, structure, or edge cases. OpenAI suggests starting with zero examples and adding them when the output drifts; Anthropic recommends including 3 to 5 relevant, diverse examples up front. If you only remember one technique from this page, remember this one, because the research shows the gains come primarily from including examples at all.
4. Specify the format
Say exactly how you want the answer shaped: a bulleted list, a table with named columns, JSON, a 3-paragraph essay, a tweet. If you do not specify, the model picks for you, and it often picks wrong. "Return your answer as a markdown table with columns: Step, Owner, Deadline" beats "give me a plan" every time.
Make it think before it answers
For anything involving reasoning, math, logic, or multi-step planning, add a simple instruction: "Think through this step by step before giving your final answer." This is chain-of-thought prompting, and the effect is not subtle. In the foundational study, a 540-billion-parameter model jumped from about 18 percent to nearly 57 percent accuracy on a grade-school math benchmark just by being prompted to show its reasoning (Wei et al. 2022, figures confirmed in this summary).
A caveat worth keeping honest: the biggest gains from "show your work" show up on math and symbolic reasoning. On many everyday writing and lookup tasks, the lift is small, and one cross-model evaluation put the average accuracy gain from reasoning-in-examples at roughly 4.23 percent. So use chain-of-thought for hard reasoning, not as a magic phrase you bolt onto everything.
The reusable template
Steal this structure for almost any task:
- Role — "You are a senior copywriter for a B2B SaaS company."
- Task — "Write a landing-page headline and three subheadlines."
- Context — "Product: invoicing software for freelancers. Audience: solo designers who hate bookkeeping."
- Constraints — "Headline under 10 words. No jargon. American spelling."
- Examples — paste 2 headlines you admire.
- Format — "Return as a numbered list. Headline first, subheads indented."
Iterate. The first prompt is a draft, not a verdict.
Treat prompting as a conversation, not a vending machine. If the output is off, do not start over. Tell the model what was wrong and ask it to fix that specific thing: "Too formal. Rewrite punchier and cut it to half the length." You converge on great output in two or three turns far faster than by crafting one perfect mega-prompt.
Common mistakes that kill output quality
- Asking for too many things at once. Split a 5-part request into steps; quality drops when you cram everything into one ask.
- Negative-only instructions. "Don't be boring" gives the model nothing to aim at. Say what TO do: "Open with a surprising statistic."
- No format, then complaining about the format. If you did not specify a table, you cannot be mad you got prose.
- Assuming it knows your context. It does not. Spell out the facts every time, or paste them in.
- Accepting the first answer. The first draft is rarely the best one. Push back.
The 60-second upgrade
Next time you are about to send a lazy prompt, stop and add three things: who the output is for, what format you want it in, and one example of good. That alone will out-perform 90 percent of how people use these tools. Specificity is free, and it is the whole game.
---
Sources: OpenAI prompt engineering best practices, OpenAI Platform guide, Anthropic prompt engineering, Wei et al. 2022 (arXiv:2201.11903).