Prompt Basics

Not magic words. Prompting is task specification, and four reusable structures cover most of what works.

Prompting has a mystique it does not deserve. There are no magic words, and lists of “secret prompts” mostly document coincidences.

What actually works is unglamorous: specify the task precisely enough that the desired output is the obvious continuation.

That framing comes straight from the mechanism. The model is predicting what comes next. Your prompt sets up the conditions. A good prompt makes the answer you want the most likely thing to follow.

Four structures

Almost every effective prompt uses some combination of these.

1 · Say what the output should look like

The single highest-return habit. Most disappointing output comes from an underspecified target.

Weak: Summarize this article.

Better: Summarize this article in three bullet points, each under 20 words, focused on what changed rather than background.

Length, format, focus, audience, what to exclude. Every unspecified dimension is one the model guesses at.

2 · Show an example instead of describing one

For anything with a specific shape, examples beat descriptions. Two examples of your desired output communicate more than three paragraphs of instruction, and they eliminate ambiguity you did not know was there.

This is the highest-leverage single technique available. See Show, Don’t Tell.

3 · Give the reasoning room

For multi-step problems, ask for the steps before the answer. Each token gets one forward pass, so written-out intermediate work is genuine additional computation rather than performance. See Making a Model Think Step by Step.

4 · Supply the context it cannot have

The model does not know your codebase, your company’s terminology, or last week’s decision. If the answer depends on information outside its training, include it.

When that information is too large to include wholesale, retrieve the relevant part — that is what RAG is for.

Structure the prompt itself

For anything beyond a sentence, separate the parts:

Task: Extract every action item from the meeting notes below.

Output format: One line per item, "OWNER: action".
If no owner is named, write "UNASSIGNED".

Notes:
---
{notes}
---

Delimiters matter more than they look. They prevent the model from confusing instructions with content — and they close a real security gap, since text inside the content block might itself contain something resembling an instruction.

Order matters too, for a mechanical reason: stable content at the beginning is what makes prefix caching effective. Put fixed instructions first, variable input last.

What consistently helps

Positive over negative. “Respond in one paragraph” works better than “don’t write multiple paragraphs.” Describing the target is more reliable than fencing off alternatives.

Explicit permission to decline. Adding “if the notes do not contain this information, say so” measurably reduces invention. Without it, the generation loop commits to answering and fills the space.

One task at a time. Four requests in one prompt gets you four mediocre answers. Split them.

A role, when the domain is ambiguous. “You are reviewing this as a security engineer” usefully shifts what gets emphasized. This is genuine framing, not a personality trick — and it does not add expertise the model lacks.

What does not help

Politeness or urgency. Please, thank you, and “this is very important to my career” are folklore. Any effect is marginal and unreliable.

Piling on adjectives. “Expert, world-class, highly detailed” adds tokens, not capability.

Longer for its own sake. Past the point where the task is specified, extra text dilutes attention and costs money.

When it is not the prompt

Prompting cannot fix everything, and misattributing failures wastes a lot of time.

If the model lacks the information, you need retrieval. If it lacks the capability, you need a stronger model. If it needs to act on the world, you need tools. If output format must be guaranteed rather than merely likely, you need structured output constraints.

Your Prompt Isn’t Working. Now What? covers telling these apart.

What to remember

• Prompting is task specification, not incantation — make the output you want the obvious continuation.
• Four structures: specify the output, show examples, allow reasoning space, supply missing context.
• Use delimiters, put stable content first, prefer positive instructions, and explicitly permit “I don’t know.”
• Politeness, urgency, and adjective stacking do nothing.
• Some failures are capability, information, or format problems, not prompt problems.

Next: What Is a System Prompt?