Prompting That Actually Works: A Practical Guide
Treat the model like a capable new hire: give it the goal, the context, an example of good, and the format you want back. Specific beats clever every time.
TLDR
There is a myth that prompting is an incantation. The reality is ordinary and useful: a good prompt is just clear delegation, the same way you would brief a new teammate.
The four things every prompt needs
Give the model what you would give a person:
- Goal. What does done look like? Name the outcome, not just the task.
- Context. Who is this for, what came before, what should it avoid?
- An example. One sample of good is worth a paragraph of description.
- Format. Ask for the shape you want back: a table, three bullets, a short draft.
Be specific, then iterate
Vague in, vague out. "Make this better" gives you a coin flip; "tighten this to 120 words, keep the second paragraph, drop the jargon" gives you something usable. When the first pass misses, correct the one thing that was off and run again.
The best prompters are not wordsmiths. They are clear communicators who would be easy to work for.
Let the model think out loud
For anything with reasoning, ask for the steps before the answer. You catch flawed assumptions early, and the final response improves because the model reasoned its way there.
