What is prompt engineering?

Writing instructions that reliably get useful output from an AI model. In practice it is far less mystical than the name suggests: most of it is specifying the audience, the format and the constraints that you were holding in your head and never wrote down. There is no secret vocabulary, and anyone selling you one is selling something.

Do I need to say 'you are an expert' at the start?

Rarely, and it is the most over-used pattern in circulation. Role framing helps when the role genuinely changes the answer — a tax specialist and a general accountant would answer differently — and does nothing when it does not. 'You are a world-class expert' before a request to summarise an email is noise.

What is few-shot prompting?

Showing the model two or three examples of the output you want instead of describing it. It is the single highest-leverage technique available, because one good example carries more information about tone, structure and length than a paragraph of adjectives ever will.

Do prompt packs and marketplaces work?

Mostly not, for two reasons. They were written for a specific model version and prompt behaviour changes as models change, and they cannot contain the context that makes a prompt work — your audience, your material, your constraints. A short personal file of prompts that worked on your own tasks beats a thousand generic ones.

What is chain of thought, and do I still need it?

Asking the model to reason step by step before answering. It mattered a great deal on older models. Current reasoning models often do it internally, so adding 'think step by step' to every prompt is frequently redundant — though asking for the reasoning to be shown is still useful when you need to check the logic rather than just the answer.

Why does the same prompt give different answers?

These models are probabilistic by design, so identical input can produce different output. That is a feature for creative work and a problem for anything you need repeated reliably. If consistency matters, give examples, tighten the format, and check the result rather than assuming it matched last time.

How long should a prompt be?

As long as the necessary context, and no longer. Long prompts fail in a specific way: the instructions compete, and the model satisfies some while quietly dropping others. If output is ignoring part of your request, the usual cause is too many instructions rather than too few.

The Complete Prompt Engineering Guide (2026)

Prompt engineering has attracted more mystique than it deserves. There is no secret vocabulary, no magic phrase, and the thousand-prompt packs being sold are mostly worthless.

What actually works is unglamorous: say what you would have had to say to a competent freelancer who cannot read your mind. This guide is that, made systematic — plus the parts that have genuinely changed as models have improved.

The six-part structure

Not every prompt needs all six. Most bad prompts are missing three.

Part	The question it answers	Example
Task	What, exactly, in one verb?	Rewrite · Summarise · Critique · Draft
Context	What does it need to know?	“This goes to customers who already complained once”
Material	What is it working on?	The actual document, pasted
Format	What shape should output take?	“Five bullets, under 15 words each”
Constraints	What must it avoid?	“No adjectives. Keep every number unchanged.”
Audience	Who reads the result?	“A busy finance director who skims”

Compare:

Make this email better.

with:

Rewrite this email for a customer who has already complained once and is still angry. Keep it under 120 words, no apologies after the first sentence, and do not promise a refund. Plain language — they are not technical.

Same task. The second one can only be answered one way, which is exactly the point. Every unstated decision is one the model makes by averaging, and the average of all possible business emails is the bland output people complain about.

The single most effective technique

Show, don’t describe.

Describing a tone takes a paragraph and still misses. One example nails it:

Rewrite these product descriptions in the style of these two examples: “Cast iron. Heavy. Lasts forever.” “Wool-lined. Warm at -10°C. Not waterproof.” Now do these five: …

This is called few-shot prompting, and it beats every other technique for consistency. Two or three examples are usually enough. If you find yourself writing “professional but friendly, authoritative yet approachable”, stop and paste two sentences you actually like instead.

Iterate on the prompt, not the output

The common failure loop is: bad output → ask for a fix → slightly different bad output → repeat until you give up.

The productive loop is: bad output → work out which of the six parts was missing → add it → generate again.

If the result was too long, you did not specify format. Too generic, you did not specify audience. Wrong emphasis, you did not give context. The model is generally doing exactly what you asked; the gap is in the asking.

Break big tasks into steps

Asking for a finished 2,000-word article in one go produces the most average version of that article, because you delegated every structural decision at once.

Do this instead:

Outline only. “Give me an outline. Do not write prose yet.”
Fix the outline yourself. This is the step people skip and the one that decides quality — it is where your judgement enters.
One section at a time. “Write section two only, in my voice, short sentences.”
Then criticise. “What have I claimed here that I cannot back up?”

Step four is the most under-used prompt in existence. Models are markedly better at finding weaknesses in text than at avoiding them while writing.

Techniques worth knowing

Ask for the opposite. “Argue against this decision” surfaces more than “is this a good decision”, because the second invites agreement and these models are agreeable by default.

Ask what is missing. “What would a specialist notice that this draft does not address?” is reliably more useful than asking for improvements.

Constrain by exclusion. “No adjectives”, “no bullet points”, “do not use the word innovative” produce sharper writing than any positive instruction about style.

Name the format precisely. “A table with columns X, Y, Z” or “valid JSON with these keys” beats “organise this nicely”.

Give it the real material. Paste the actual document rather than describing it. Every current assistant handles long inputs well, and a real example removes a hundred guesses.

What has stopped working

Prompt advice ages badly. Three patterns that were genuinely useful and are now mostly noise:

“You are a world-class expert in…” Role framing helps when the role changes the answer. Stacking superlatives does not. Modern models do not produce better summaries because you told them they were world-class.

“Think step by step” on everything. This mattered enormously on older models. Current reasoning models often do it internally, so bolting it onto every prompt is frequently redundant. Still useful when you want to see the reasoning to check it — which is a different goal.

Emotional pressure. “This is very important to my career” and tipping promises circulated widely. Treat any such trick as unproven unless you have tested it on your own task, and be aware these behaviours change between model versions without notice.

The general rule: test rather than believe. Run the same task with and without the technique, on your own material. That takes two minutes and beats any list, including this one.

Knowing when to stop prompting

The genuinely valuable skill is recognising when the prompt is not the problem.

The model lacks the information. No prompt retrieves a document it has never seen. Upload the file, or use a tool built for sourced answers — Perplexity for cited research, or NotebookLM for questions about your own documents.
The task needs the wrong medium. Asking a chat assistant for a poster with readable text will fail regardless of phrasing. Use Ideogram — see how to write AI image prompts.
It needs to act, not answer. Editing files across a repository is an agent’s job, not a prompt’s. See best AI coding tools.
You would be faster writing it. Sometimes true, and worth admitting.

Prompting differs by tool

The six-part structure transfers everywhere. The details do not.

ChatGPT — the largest ecosystem; Custom GPTs let you save standing instructions. Start with our ChatGPT beginner’s guide and prompts that save time.
Claude — strongest on long documents and keeping your voice; Projects hold context so you stop re-explaining. See the Claude beginner’s guide.
Gemini — Gems save custom instructions, and Deep Research is a different product from the chat box. See the Gemini beginner’s guide.
Image and video — prompts describe a picture or a shot, not a topic. Image prompts and what video credits actually buy.

Build your own file, not a collection

Keep a short document of prompts that worked on your own tasks, with the tool and the date beside each one.

That file will outperform any prompt pack, for a reason worth understanding: downloaded prompts cannot contain your context — your audience, your material, your constraints — and that context is where nearly all the quality comes from. The prompt is just the container.

Revisit it occasionally. Model behaviour changes, and a prompt that was excellent in March may need a sentence adjusting by August. Dating them is what makes that visible.

The short version

Task, context, material, format, constraints, audience.
Show examples instead of describing style.
Fix the prompt, not the output.
Outline → correct it yourself → section by section → criticise.
Test techniques on your own work rather than believing lists.
Know when the prompt is not the problem.

Where to go next

ChatGPT prompts that save time — ready-made, for real tasks
How to use Claude · ChatGPT · Gemini
How to write AI image prompts
Best AI chatbots — pick the tool before tuning the prompt