How to Write AI Prompts: The Four Parts That Decide the Output

Most disappointing AI output is not a model problem. It is a prompt that asked a stranger to do a job without telling them anything about it, then judged the result.

There is no secret phrasing. The single biggest predictor of output quality is how much of the relevant context you actually supplied, and almost everyone under-supplies it. This guide covers the structure that fixes that, how to repair a bad answer in one follow-up, and which popular techniques stopped earning their keep.

The four parts

A prompt that works has four components. Missing any one produces a predictable kind of failure.

Compare the two versions of the same request.

Weak: "Write a product update email."

Strong: "You are writing to existing customers of a B2B analytics tool. Here is the changelog [paste]. Two of these fix bugs customers complained about in support, marked with an asterisk. Write a 120-word email that leads with the fix people asked for, mentions the other changes in one sentence, and does not use the word 'excited'."

The second is not cleverer. It pasted the changelog, named the audience, set a length, and banned the word that would have made it sound like everyone else's email.

Context is the whole game

The model knows an enormous amount about the world and nothing about your situation. Every fact you withhold, it will invent a plausible substitute for.

Paste the source material. The thread, the document, the data, the previous version, the example you like. A prompt with 500 words of pasted context and a one-line instruction beats a beautifully engineered instruction with no context, every time.

Two practical habits:

Show an example of the output you want. One sample of the tone or structure does more than three paragraphs describing it.

State what to avoid. Models default to the average of what they have seen, which is why so much output reads the same. Naming the clichés you do not want is the fastest way out of that register.

Fixing a bad answer

The follow-up is where quality actually arrives, and most people restart instead of correcting.

Do not re-prompt from scratch. Say what was wrong with the specific answer. "Too long, cut to half" or "you assumed we sell to consumers, we sell to hospitals" gets you a good second attempt. Rewriting the original prompt throws away everything that was right.

That last one is genuinely useful and underused. Asking a model to mark its own weak points surfaces the claims worth checking, which is faster than verifying everything.

What stopped mattering

Several widely-taught techniques earned their reputation on older models and now add length without much benefit.

Elaborate role-play preambles. A short role line helps. Three paragraphs of persona does not.

Explicit "think step by step". Current reasoning models do this natively. Adding the phrase is mostly harmless and mostly redundant.

Offering tips or threatening consequences. This was always folklore.

Very long instruction stacks. Twenty rules produce a model tracking twenty rules rather than doing the job. Five that matter beats twenty that mostly do not.

What did not stop mattering: pasting the source material, naming the audience, and specifying the format.

Tools

The structure above is portable across assistants, so pick one and get fluent rather than sampling. Our comparisons cover the best AI chatbots, AI writing tools and AI search engines if your work is research-heavy and you need sources shown.

Save the prompts you reuse. Five reusable prompts you actually run each week beat any course on the subject.

Pitfalls

Asking for several things at once. Split it. One task per prompt, then chain.

Accepting the first answer. It is a draft. The second exchange is where the work happens.

Not saying who it is for. "For a customer" and "for our CFO" produce different, and differently useful, answers.

Trusting numbers. Models state recalled, inferred and invented figures with identical confidence. Check anything you will act on.

What these tools actually cost

We price every tool we review, so this is measured rather than estimated. Across 429 tools, 293 publish a price and 33% offer a free tier. Among content creation tools, the median entry plan is $15 a month, which sits below the $24 median across every category we price.

The spread matters more than the median. Half of the content creation tools sit between $9 and $24, and the range runs from $5 to $99. A quoted "starting at" price near the bottom of that range usually means per-seat add-ons land on top of it.

FAQ

What is prompt engineering?

It is the practice of structuring what you send a model so the output is usable: supplying context, constraining the task, and specifying the format. The term oversells it slightly, because the skill is much closer to briefing a competent colleague clearly than to programming, and the biggest wins come from context rather than clever phrasing.

Do I need different prompts for different models?

Less than you would expect. The four-part structure works across every mainstream assistant, because it addresses what the model does not know rather than how a particular one is built. Differences show up at the edges, mainly in how much material you can paste and how each handles very long instructions.

Why does AI writing sound the same everywhere?

Because a prompt with no context or examples returns the average of everything the model has read, and that average has a recognisable register: even paragraph lengths, hedged enthusiasm, a tidy summary at the end. The fix is to supply your own material and explicitly name the patterns you do not want.

How long should a prompt be?

As long as the context requires and no longer. A one-line instruction with a pasted document is often ideal. What does not help is length made of instructions rather than information: five clear constraints beat twenty, and a wall of rules makes the model manage rules instead of doing the task.