Prompt Engineering: A Practical Guide to Writing Better AI Prompts

Most people get mediocre results from AI not because the model is weak but because the instruction is vague. “Write me a marketing email” gives the model almost nothing to work with, so it fills the gaps with generic filler. Prompt engineering is the discipline of removing that ambiguity. This guide covers the structure that works across every major model, the named techniques worth knowing, and the debugging loop that turns a flaky prompt into a dependable one.

The anatomy of a strong prompt

Nearly every prompt that performs well in production contains the same five components. You do not always need all five, but reaching for them in order is the fastest way to improve a result.

1. Role — who the model should act as. “You are a senior tax accountant” pulls the answer toward the right vocabulary and rigour.
2. Task — the single, specific thing to do. One verb, one deliverable.
3. Context — the facts the model cannot know: the audience, the source text, the brand voice, the data.
4. Constraints — length, format, tone, what to avoid, the output schema.
5. Example — a worked sample of the output you want (this is the most under-used and most powerful element).

Compare a weak prompt with an engineered one:

Weak:
"Summarise this article."

Engineered:
"You are an editor writing for busy executives.
Summarise the article below in exactly 3 bullet points.
Each bullet must be one sentence, lead with the key number,
and avoid jargon. If a claim has no supporting data, omit it.

Article:
<<<{article_text}>>>"

Core techniques worth naming

Zero-shot, one-shot, and few-shot

A zero-shot prompt asks for the task with no examples. A few-shot prompt includes two to five worked examples of input → output. Few-shot is the most reliable way to lock down a specific format or style, because the model pattern-matches against your examples rather than guessing.

Chain-of-thought

Adding “think step by step” or asking the model to show its reasoning before the final answer measurably improves performance on maths, logic, and multi-step tasks. For a deeper treatment, see our dedicated guide on chain-of-thought prompting.

Output formatting and delimiters

Wrap input text in clear delimiters (triple quotes, XML-style tags, or fenced blocks) so the model cannot confuse instructions with data. When you need machine-readable output, ask for “valid JSON matching this exact schema” and provide the schema. This single habit eliminates most parsing failures in automation pipelines.

Role and persona via the system prompt

Durable rules — tone, persona, hard constraints — belong in the system prompt, not repeated in every message. See system prompt vs user prompt for when to use each.

The debugging loop

Treat a prompt like code. When the output is wrong, do not rewrite from scratch — isolate the failure and fix one variable at a time:

• Wrong format? Add an example and an explicit schema. Lower the temperature.
• Made-up facts? Supply the source text and add “answer only from the text below; if it is not there, say so.”
• Too generic? Add audience, role, and at least one concrete constraint.
• Inconsistent across runs? Lower temperature, pin the format, and test the prompt on a fixed set of inputs before you trust it.

Stop rewriting: build a prompt library

Once a prompt works, do not let it disappear into your chat history. Save it as a reusable template with named variables. A team that shares a vetted library moves far faster than one where every person re-discovers the same prompt. This is exactly what GeraPrompts exists for — see our guide on building a prompt library and browse vetted, runnable prompts you can adapt instead of starting from a blank page.

A copy-paste checklist

Before you send a prompt, confirm:
[ ] Role:        who should the model be?
[ ] Task:        one clear deliverable
[ ] Context:     everything the model can't know
[ ] Constraints: length, format, tone, what to avoid
[ ] Example:     a sample of the exact output
[ ] Delimiters:  data wrapped, separated from instructions
[ ] Schema:      if machine-read, specify exact structure

Master these and you will get more from any model — GPT-4, Claude, Gemini, or an open-weights model running locally.