Prompt Patterns: Few-shot, CoT, Roles

You already know the core mechanic: a model continues text plausibly. That one fact explains almost every prompting failure you will ever hit.

Ask for "a summary of this report" and you get a summary — a competent, forgettable, middle-of-the-road summary. The model didn't malfunction. It gave you the most plausible continuation of a vague request, and the most plausible continuation of a vague request is average.

Think about what "summary" actually covers. A one-line summary for a CEO. A bullet list of action items. A customer-facing note. A technical abstract. The model has seen all of them. You named none of them. So it produced the blurry centre of all of them at once.

Prompting is specification

It is not magic words. There is no phrase that unlocks a better model. The job is much less mystical than that:

Remove ambiguity until only the answer you want is plausible.

Three patterns do most of that work:

• Few-shot — show examples instead of describing.
• Chain of thought — ask for reasoning before the answer.
• Roles — set the frame the answer is written from.

They stack, and most strong production prompts are all three at once. Let's take them one at a time.