Most people use AI tools like a search engine. Short question in, answer out, disappointment follows. Then the conclusion: “AI just isn’t there yet.”
That is usually wrong. The model is capable enough. The question was not.
Prompt engineering sounds like a developer topic. Like Python code and API calls. But it is not. At its core it is simply: communicating clearly what you want. And anyone who has written a good briefing email can do that.
Context matters more than the keyword
The most common weakness in a prompt is missing context. The model does not know who you are, what the background is, who the text is for, or what it should achieve.
Poor: “Write me an email to a client.”
Good: “I am a sales manager at a software company. A client has not responded in three months. I want to follow up, friendly but direct, without being pushy. Maximum five sentences.”
Same task. Completely different result. The second prompt takes 20 seconds longer. The time saved on editing: often several minutes.
Assign a role
“Respond as an experienced project manager” delivers different results than the same question without a role. The model adjusts tone, depth and perspective.
This works especially well when you need a perspective you do not have yourself. “Respond as a sceptical CFO who would need to approve this project” is a good stress test for your own argumentation.
Specify the format
If you do not specify a format you get what the model thinks is appropriate. That is sometimes good, usually too long.
“As bullet points”, “as a table”, “maximum five sentences”, “with one concrete example”, “no introduction”, these instructions cost nothing and save a lot of editing.
Name the constraints
Negative definitions are just as valuable as positive ones. “No jargon”, “no German”, “no bullet points”, “not more than 200 words”, the model follows these if you say them.
Without these instructions it optimises for completeness. With them it optimises for what you actually need.
Store your own examples as context
This is one of the most underrated levers of all.
Anyone who regularly writes similar texts should build a small library of good examples. Three emails that landed well. A proposal that worked. A meeting note that was praised. Feed these as context: “Here are three emails I wrote: [examples]. Now write an email in the same style to...”
This is the fastest way to transfer your own communication style. No style description, no explaining. Just show.
Claude has Projects for this, ChatGPT has Custom Instructions. Set it up once, available permanently. This saves time on every prompt and makes results more consistent.
Iterate instead of starting over
The first output is rarely the best. That is not a failure, that is the process.
“Make it shorter”, “rewrite the third paragraph more directly”, “replace the example with one from retail”, “sounds too formal, more casual please.” Keep the conversation going, do not ask once and give up.
Most good texts from AI tools emerge after three to five iterations. Not after one.
The model can ask questions too
If you are not sure whether the prompt is good enough, you can simply ask the model to ask. “Before you start: what else do you need from me to do this well?” or directly: “If anything is unclear or you need more context, ask me first before you start.”
This sounds simple. But it completely changes the workflow. Instead of getting a result that misses the mark, a brief dialogue emerges that makes the result significantly better. Especially useful for complex tasks where you yourself are not yet sure what you want.
How long should a prompt be?
No universal optimum, but a good rule of thumb: as short as possible, as long as necessary.
Three clear pieces of information beat ten vague ones. The most common mistakes are either too short (no context, poor results) or too long (too many instructions that contradict each other). For most everyday tasks two to four sentences work well if they are well constructed. Those who include examples can make the examples long and the instruction short.
Before and after
Poor: “Write me a meeting note.”
Good: “Write a note for a 45-minute status meeting with three participants from product, sales and IT. Format: decisions at the top, then open items with owners and deadlines, no prose, maximum one page.”
Same effort, completely different result.
What this has to do with consulting
Good prompts and good briefings follow the same rules. Anyone who can clearly explain to a colleague what they need can explain it to an AI model too. Those who cannot will get poor results from both.
The difference: the AI model does not complain when the briefing is unclear. It just interprets. And usually differently than hoped.
One last thought
This article is deliberately general. A sales manager who writes proposals every day needs different prompts than someone in controlling who summarises data. Those building AI agents think about it differently again.
But the core principles apply everywhere: give context, specify format, iterate, use your own examples, let the model ask questions. Those who have these down are already significantly more specific than most people in their daily work, and can build from there for their own role and tasks.
The first step is always the same: stop using AI like a search engine.
— Robert
