The Practitioner's Prompt Framework

Most people prompt an AI the way they'd text a friend: one line, a lot of assumed context, and a shrug when the answer misses. That's fine for trivia. It fails on real work, because the model can't read your mind — it can only work with what you put in front of it. The practitioner's move is to stop improvising and start using a repeatable structure.

The framework is five parts. Role — who the AI should act as. Context — the background and constraints it needs. Task — the specific thing to do. Format — exactly how you want the output shaped. Examples — one or two samples of "good", when you have them. You won't need all five every time, but naming them turns a vague request into a specification.

"Write me a marketing email" becomes: "You're a B2B copywriter. We sell an AI-tools intelligence platform to enterprise L&D buyers; the tone is confident, not hypey. Write a 120-word cold email that books a demo. Output a subject line, the body, and a one-line call to action." Same model, same few seconds, entirely different result.

Of the five, Examples is the one people skip and the one that moves quality most. Describing a tone in words is lossy — "professional but warm" means something different to everyone. Pasting two emails you've actually sent and saying "match this voice" is unambiguous. This is called few-shot prompting, and for anything with a house style it consistently beats a paragraph of adjectives. Two examples is usually the sweet spot; one is a coincidence, five is diminishing returns and wasted context.

Format deserves more attention than it usually gets, because it's what makes output usable rather than merely correct. "A markdown table with columns Risk, Likelihood, Mitigation." "Valid JSON matching this shape, no prose around it." "Three bullets, each under fifteen words." If the output feeds into something else — a spreadsheet, a slide, a script — specify the shape and you stop reformatting by hand.

The real payoff isn't a better answer, it's a reusable one. Once a structured prompt works, save it as a template with the variable parts blanked out: [AUDIENCE], [PRODUCT], [WORD COUNT]. Now anyone on the team gets that quality without being a "prompt wizard", and when the output drifts you fix the template once rather than re-improvising. That's the actual difference between a practitioner and a dabbler — they build prompt assets, not prompt attempts.

Keep the templates somewhere shared and versioned, even if that's just a document with a date on each entry. The valuable artefact is not the clever prompt, it's the accumulating library of ones that reliably work for your particular business.

Two cautions. First, more structure is not always better. For open-ended thinking — "help me find angles I haven't considered" — an over-specified prompt boxes the model into your existing assumptions, which is the opposite of what you asked for. Match the structure to the job: tight specs for repeatable output, loose prompts for exploration.

Second, don't pile on emphasis in place of information. Writing IMPORTANT in capitals, or promising a tip, or telling the model to take a deep breath, are folk remedies that made a real difference on weaker models and mostly don't now. If an answer is wrong, the fix is almost always a missing fact, a missing constraint, or a missing example — not a stronger adjective.