Academy
1. What Is a Prompt — And Why Yours Might Be Failing
Understand that a prompt is everything the model receives, and that vague instructions like 'make it better' fail because the model can only optimize what you name.
A prompt is not just the sentence you type. It is everything the model receives for one request: your message, the conversation history before it, any attached documents, and the hidden instructions built into the product you are using. Prompting is the craft of making that combined input clear enough that the most plausible answer the model can produce is also the one you actually want. When you understand that the model sees all of these layers at once, you start writing differently — because you realize you are shaping an entire context, not just asking a question.
A good prompt can carry up to five parts: a goal (what you want and what it is for), context (who you are, who will read the output, what already exists), format (length, structure, language, tone), constraints (what to avoid, what must remain true), and optionally an example of the expected result. You rarely need all five in every request. But when an answer disappoints you, the cause is almost always a missing part. Before blaming the model, check which of the five you left out.
The single most common prompting mistake is vague feedback. Telling a model to "make it better" gives poor results because "better" is not a direction — it is every direction at once. The model can only optimize what you name. Say exactly which dimension needs to change: shorter, warmer, more formal, with one concrete example per point, aimed at a reader who has never seen the product. Each of those is a clear axis the model can act on. Stack several if you need to, but never leave the axis unnamed.
Before you send a prompt, read it back and ask: "Could someone interpret this in a way I don't want?" If the answer is yes, add the missing part — usually a format, a constraint, or a description of the reader. Naming the audience and the standard you expect beats inventing a persona.
When the output format matters — a table, a bulleted list, a JSON object — show it. Describe the structure you need ("a table with three columns: Feature, Benefit, Limitation") and, if precision matters, paste a short example of what a correct row looks like. Models follow examples more reliably than they follow abstract descriptions. This is one of the simplest upgrades you can make to any prompt: stop describing the format and start demonstrating it.
Open a conversation and write a prompt that is deliberately vague — for example, "Write me something about onboarding." Send it, read the result, then rewrite the same request adding a goal, an audience, and a format constraint. Compare the two outputs and notice which missing part made the biggest difference.
A prompt is everything the model sees, not just your last message. Weak results almost always trace back to a missing part: goal, context, format, constraints, or example. "Make it better" fails because the model can only optimize dimensions you name — so name them. Show the format you want instead of describing it, and always ask yourself whether your request could be misread. Precision in, precision out.