AI Prompting: A Practical Guide to Getting Usable Output

AI prompting is the practice of writing instructions that get a model to produce something you can use. Not something impressive, not something that technically answers the question, but something you would actually put your name on.

The gap between those is where all the difficulty lives, and it is smaller than the amount written about it suggests. Most of what makes prompting work comes down to a handful of ideas.

Why vague prompts produce boring output

A model trained on a very large amount of text has, for any loose question, an enormous number of valid responses. When your instruction does not narrow that space, you get something near the middle of it. The middle of all possible blog posts about productivity is a blog post about productivity that reads like every other one, because that is literally what it is.

This is why “be more creative” rarely helps and “no metaphors, no rhetorical questions, no sentence longer than 20 words” often does. You are not asking for more creativity. You are removing the places the average lives.

Every part of a good prompt does the same job from a different angle: it eliminates answers you did not want.

The four parts

Task. The verb. What is being done. “Summarise”, “critique”, “rewrite”, “list”, “compare”. Ambiguity here poisons everything downstream, because the model has to guess your intent before it can even start.

Context. Everything the model cannot see: the audience, the history, the constraints, what you already tried. This is the highest-return part per word and the one people skip most, usually because it feels like it should be obvious. It is not obvious. The model has no access to your situation.

Format. Length, structure, shape. Specify it in countable terms. “Short” is a judgement you have handed over. “Under 120 words” is not.

Constraints. What to avoid and what to do when information is missing. “If you’re not sure, say so rather than guessing” is close to a free upgrade, because the default behaviour is to produce something confident regardless.

Techniques that earn their keep

There is a long list of named prompting techniques. A few of them matter for ordinary work.

Show examples

Give two or three examples of input and the output you want, then give the real input. This communicates tone, structure, and level of detail faster than describing them, and it handles the things you cannot articulate. If you have ever tried to explain “our house style” in words, you know why this works.

Two or three good examples beat ten mediocre ones. The model will imitate whatever pattern is actually in your examples, including patterns you did not intend, so check them.

Ask for the reasoning before the answer

For anything involving multiple steps, comparison, or arithmetic, asking the model to work through it before answering measurably improves the result. “Think through the steps, then give your answer” is enough. Newer reasoning models do a version of this internally, but stating it still helps when the task has a specific structure you want followed.

The side benefit is that when it gets something wrong, you can see where.

Assign a role, but a specific one

“You are a helpful assistant” does nothing. “You are a copy editor for a technical publication whose readers are engineers; your job is accuracy and concision, not enthusiasm” does something, because it implies a set of judgements.

The useful part is not the job title, it is the standards that come with it.

Let it ask first

For anything with real requirements, this one instruction outperforms most of the others:

Before answering, ask me anything you need to know to do this well. Do not answer until I reply.

You will find out immediately whether your request was as clear as you thought.

Iterate rather than restart

When the output is 70% right, do not rewrite the prompt from scratch. Say what is wrong with what you got. “The third section is too abstract, give me a concrete example there” is faster and more precise than another attempt at a perfect first prompt.

The perfect one-shot prompt is mostly a fantasy. Working prompts are usually the third message.

The failure modes

Asking for too much at once. A prompt that requests research, structure, drafting, and editing produces something mediocre at all four. Split them. Each step also gives you a checkpoint where you can catch a wrong turn before it propagates.

Politeness as instruction. “Could you maybe try to make it a bit shorter?” is not a constraint. “Maximum 150 words” is. You are not being rude to a machine by being precise.

Assuming shared context. You know the project. The model knows the words in your message. Anything you did not say is a blank the model will fill from the statistical average, which is exactly what you did not want.

Trusting confident output. Fluency and accuracy are separate. A model will produce a well-structured, plausible, entirely wrong answer with no change in tone. Anything factual needs checking, and the ones that sound most authoritative deserve it most.

Not saving what worked. People rebuild the same prompt weekly. If a prompt produced something good, it is an asset. Keep it.

Prompting for images is a different skill

Text prompting is about narrowing meaning. Image prompting is about specifying a picture, which means stating things you never say out loud: light direction, lens length, camera height, depth of field. “Nice portrait” is not a specification. “85mm, f/1.8, soft key from camera left at 45 degrees, eye level, shallow depth of field” is.

If you are editing an existing photo rather than generating a new one, there is one extra rule that matters more than all the others: state explicitly what must not change. See Gemini AI photo prompts for how that plays out.

A worked example

Starting point:

Help me write a job posting for a developer.

What you get back is a generic job posting, because that is exactly what you asked for.

After applying the four parts:

Write a job posting for a backend developer at a six-person company building logistics software. Audience is experienced developers who are tired of large-company process. What is genuinely good: no on-call rotation, small codebase, direct contact with the people using the product. What is genuinely hard: no dedicated ops, you own what you ship, the codebase has five years of history and some of it is ugly. Around 300 words. Be honest about the hard parts rather than spinning them. No phrases like “rockstar”, “fast-paced”, “wear many hats”. If you need a detail I haven’t given, mark it clearly rather than inventing it.

Same task, entirely different output, and the difference is almost all context and constraints.

Where to go next

• AI prompts — forty prompts you can copy and adapt
• AI prompt generator — when a tool is worth it and when it is in the way
• Gemini AI photo prompts — image editing specifically

The prompt community is open and account-free if you want to browse what other people are using or keep your own somewhere you will find them again.