How to Write AI Prompts That Get Real Results

To write an AI prompt that gets real results, give the model five things: a role, one clear task, the context it needs, the output format you want, and an example. Most weak outputs are not a model problem. They are a prompt that left too much to guesswork. Here is how to write prompts that come back usable the first time.

TL;DR

• A reliable prompt has five parts: role, task, context, format, and an example. Missing parts get filled in by the model's guesses.
• Be specific. "Make this better" gets you nothing. "Cut this to 120 words, keep the data, drop the adjectives" gets you the edit.
• Show, do not just tell. One example of the output you want beats three paragraphs describing it.
• Name the output format every time, or you will spend your day reformatting.
• Treat prompting as iteration, not a coin flip. Change one thing, see what moves.

The five parts of a prompt that works

You do not need every part for every task, but when an output keeps missing, it is almost always because one of these is absent.

1. Role

Tell the model who it is. "You are a copy editor who cuts ruthlessly and never changes the author's meaning" sets a stance the rest of the prompt can lean on. The role is not magic, but it primes tone, priorities, and what the model treats as important.

2. Task

State the one thing you want done, as a verb. Not "thoughts on this email?" but "rewrite this email to be three sentences shorter and more direct." One prompt, one job. If you are asking for three different things, you will get three half-done things.

3. Context

Give the model what it cannot know on its own: the audience, the goal, the constraints, the source material. The model is sharp but it has no idea who this is for or what you are optimizing for unless you say so. Most "the AI does not get it" moments are missing context, not missing intelligence.

4. Format

Say exactly what the output should look like. A bulleted list. A table with these columns. A 50-word summary. JSON with these keys. If you do not specify the shape, you get the model's default shape, which is rarely the one you can paste straight into your work.

5. Example

This is the highest-leverage part and the one people skip. Paste one example of a great input and its ideal output. The model pattern-matches off it instantly. A single good example routinely outperforms a long paragraph of instructions, because you are showing the target instead of describing it.

The mistakes that quietly wreck your outputs

Four habits cause most bad results:

• Vague verbs. "Improve," "optimize," and "make it pop" mean nothing concrete. Replace them with the actual change: shorter, in this tone, with these removed.
• Stacking jobs. Asking for a summary and a rewrite and three subject lines in one prompt splits the model's attention. Break it into steps.
• No format. Leaving the shape unstated guarantees a reformatting tax on every output.
• Burying the ask. The most important instruction hidden in the middle of a wall of text gets the least weight. Lead with it, or repeat it at the end.

A before and after

Here is the same request, weak and strong.

Weak: "Can you help me write a better LinkedIn post about our new feature?"

Strong:

You are a B2B content writer who values clarity over hype.

Task: write one LinkedIn post announcing our new feature.

Context: the feature lets small teams schedule social posts a week ahead. Audience is solo founders and small marketing teams. Goal is signups, not applause. Voice: plain, confident, no buzzwords.

Format: a hook line, three short paragraphs, and a one-line call to action. Under 150 words. No hashtags.

Example of the tone I want: [paste a post you like].

The weak version makes the model guess at audience, goal, length, tone, and shape. The strong version removes every guess, so the first draft is close instead of generic. That difference is the whole game. If you want a head start, the prompts in Secret Codes Vol. 1 are built on exactly this structure.

Iterate like an engineer, not a gambler

A bad prompt is not a dead end, it is a first draft. When the output misses, do not rewrite the whole thing and reroll. Change one variable and watch what moves:

1. Output too long? Add an explicit word limit.
2. Wrong tone? Add a role and a one-line voice description.
3. Ignored a rule? Move it to the top and state it once, clearly.
4. Format off? Paste an example of the exact shape you want.

Changing one thing at a time tells you which lever actually mattered, so your prompts get better instead of just different. Save the ones that work. A small library of proven prompts beats clever improvisation every single time.

The bottom line

Writing AI prompts that get real results is not a trick or a secret phrase. It is the discipline of removing guesswork: a role, one task, the context, the format, and an example. Get those right and an ordinary model gives you sharp, usable output on the first try. Get lazy on them and even the best model hands you something generic.

For the practical AI moves worth making each week, join the OpusJake newsletter. And if you want to put this to work, opusjake.ai is where the prompt packs and playbooks live.