Prompting is a smaller skill than the market for prompt courses suggests. There are four things that matter, a handful of habits, and a long list of tricks that do nothing.
The four things
1. The material. Paste it. Do not describe it. This is the single biggest difference between a generic answer and a useful one, and it is the step people skip most.
"How do I improve my resume?" → generic advice. "Here's my resume: [paste]. What's weak?" → useful advice.
2. The audience. Who reads the output. "Explain this to a customer who isn't technical" produces a completely different answer from "explain this to me."
3. The shape. Length and format. "Three bullets", "under 100 words", "a short email, no subject line." Unspecified, you get the model's default, which is longer and more structured than you wanted.
4. The bans. What you do not want. "No bullet points", "don't use the word leverage", "no closing summary." Negative instructions work well and are badly underused.
Miss one and the output drifts toward the average of everything the model has seen. That average is exactly what people mean when they say AI writing sounds like AI writing.
Things that do not work
Worth stating plainly, because they circulate constantly:
- "You are a world-class expert with 20 years of experience." Close to nothing. Concrete instructions about audience and format do far more.
- Politeness, threats, offers of money. No meaningful effect on quality.
- "Take a deep breath and work step by step." Some early results suggested small effects; it is not a reliable technique and modern reasoning models do this internally.
- Elaborate formatting rituals — XML tags, ALL CAPS, delimiters everywhere. Structure helps with genuinely complex prompts and does nothing for ordinary ones.
- Very long prompts. Past a point they dilute rather than clarify. Specific beats long.
The habits that matter more than the prompt
Follow up instead of restarting. "Shorter." "Too formal." "Keep paragraph two, redo the rest." A new conversation throws away everything the model learned about your task. This one habit is worth more than any template.
Put your question after the material. Both major model families weight later instructions more heavily. A question buried above forty pages gets less attention than the same question below them.
Ground it when facts matter. "Answer only from the text below. If it is not there, say so." The most effective single instruction against invention.
Ask for a diff, not a rewrite. "Show only what changes." Reviewable instead of a wall you have to compare by eye.
Keep what worked. A note file of five prompts you reuse beats a library of two hundred you do not.
Examples beat adjectives
The most useful single technique for voice and style.
"Write in a professional but friendly tone" means nothing measurable. Instead:
Here are three things I've written. Match this voice — notice the sentence length and how little jargon there is. Now write [X].
Three samples outperform any description, because you are showing rather than naming. The same applies to formatting, structure and level of detail: show one example of what you want.
Do prompts transfer between models?
Mostly, yes. The four-part pattern works everywhere.
What does not transfer is fine-tuned phrasing — the specific incantation someone found for one model version. That is also the thing prompt-course marketing sells hardest, and it goes stale with every release.
Write prompts that would make sense to a competent human assistant. Those keep working.
A prompt worth stealing
The general-purpose shape, for anything substantial:
[Context: one line on what this is and why]
[The material, pasted]
[The task, specifically]
Audience: [who reads it] Length: [limit] Avoid: [the things you do not want]
That is the whole method. Everything else is variation.
Common questions
What makes a good AI prompt? Your actual material, a named audience, a specified length and format, and an explicit list of what to avoid.
Do I need to learn prompt engineering? Not as a discipline. Four principles and a couple of habits cover almost everything.
Do role-play prompts work? Barely. "You are an expert X" is far weaker than concrete instructions.
Should I be polite to the model? It makes no meaningful difference. Do whatever you prefer.
Why do I get different answers to the same prompt? Models sample rather than always choosing the most likely word. Same prompt, different answer, by design.
Do prompts work across different models? The principles do. Model-specific phrasings usually matter less than people think and go stale quickly.
What is the fastest improvement I can make? Paste your material instead of describing it, and follow up instead of starting over.