Prompting for Real Work

Most disappointing output comes from a prompt that would also have confused a competent colleague. These four patterns fix the majority of it, and they compound: used together they change the quality of what comes back far more than switching models does.

Give the context you would give a new colleague

A model cannot see your situation. Say who the output is for, what it will be used for, and what the reader already knows. This single addition changes more than any clever phrasing.

Weak:   Write about our new course.

Better: Write a 120-word description of a 12-week data analytics
        course for final-year students in India who have used
        Excel but never written SQL. It goes on a landing page
        above the fold. Plain language, no hype, no exclamation
        marks.

Show one example of what good looks like

Describing a style is slow and imprecise. Pasting one example of it is fast and exact. One good example usually beats three paragraphs of instruction, and two examples establish a pattern the output will follow closely.

Constrain the output, not the thinking

Be strict about the shape of the answer and loose about how it gets there. Format constraints are cheap to specify and make the result immediately usable.

• State the length in words or items, not "short".
• Ask for a specific structure: a table with named columns, JSON with named keys, five bullets.
• Name the audience and the reading level.
• Say what to leave out, which is often more useful than what to include.

Iterate on the answer, do not restart

When the output is close but wrong, most people rewrite the prompt from scratch and lose the good parts. Correct in place instead: name the specific problem and ask for a revision of that one thing.

"Keep the structure. The second paragraph reads like marketing copy.
 Rewrite only that paragraph in the plain register of the first."

Know when not to use it

Language models are strong at drafting, restructuring, summarising and explaining. They are unreliable for anything requiring a fact you cannot verify, arithmetic you have not checked, or a citation you have not opened. Use them for the first draft and the reshaping, and keep the verification for yourself.