Prompt Craft

Imagine hiring a brilliant temp who has read half the internet but knows nothing about your company, your boss, your customers or what you did yesterday. Each morning they forget everything. Every instruction you give them has to carry all the context they need. That is roughly the position of an AI assistant when you open a new chat — and it is why the same model can look dazzling for one person and useless for another. Prompt engineering sounds technical, but at work it mostly means writing a good brief.

The anatomy of a strong prompt

The Novice prompting page introduced clear asks and iteration. At practitioner level, it helps to think of a prompt as having up to six parts. Not every task needs all of them, but when a result disappoints, one of these is usually missing.

1. TaskWhat you want done, in one plain sentence, starting with a verb: summarise, draft, compare, rewrite, critique.
2. ContextWhy it is needed, who it is for, what they already know, what has been tried.
3. MaterialThe text, data or notes to work on — clearly separated from your instructions.
4. ExamplesOne to three samples of the style, format or level you want.
5. Constraints and success criteriaLength, tone, what to avoid, what must be included, and how you will judge the result.
6. Output formatBullets, table, headed sections, email, JSON — spelled out.

Here is the difference in practice.

Roles and personas — useful, within limits

“You are an experienced HR manager…” is probably the most famous prompting trick. A role tells the model which register, vocabulary and priorities to adopt, and it is genuinely useful for style: a role as “a patient primary-school teacher” produces simpler explanations than “a professor of physics”.

Debated Whether a role makes the model more accurate is less clear. A 2024 study testing many personas across thousands of factual questions found that adding a persona did not reliably improve accuracy, and that which persona helped looked largely random. Declaring the model “a world-class lawyer” does not give it legal knowledge it lacks.

Structure: keep instructions and material apart

When you paste a long email thread and add “summarise this” at the end, the model has to guess where your instructions stop and the material begins. Clear separators — called delimiters — remove the guesswork. Three common styles work across assistants:

• Headings: lines such as ## Instructions and ## Text to summarise.
• Fences: triple quotes or dashes around the material.
• XML-style tags: <report> … </report>. Anthropic’s guidance recommends these, and they work well with other models too. The tag names are yours to invent; consistency is what matters.

Structure also helps with prompt injection: if a pasted document contains a line such as “ignore previous instructions”, tags make it easier for the model to treat that as material rather than as your command. It reduces the risk; it does not remove it.

Show, don’t only tell: examples

Describing a style is hard; showing it is easy. Giving the model a few worked examples is called few-shot prompting, and the 2020 GPT-3 paper that popularised the term for prompting showed how strongly models pick up a pattern from a handful of demonstrations. Consensus All three major vendors’ guides list examples among the most effective techniques.

Two cautions. First, models copy examples closely — including their length and quirks — so vary them if you want variety. Second, if your examples share a hidden pattern (all positive, all about Kuala Lumpur), the outputs will too.

Constraints, success criteria and output formats

Constraints narrow the space of possible answers: word limits, reading level, what to leave out, which facts must stay unchanged. Success criteria go one step further and tell the model how the result will be judged — “a new staff member should be able to follow this without asking anyone” is a criterion the model can aim at.

State positively what you want, not only what you don’t. “Use plain sentences a secondary-school student can follow” works better than “don’t be too complicated”. And ask for a specific format — it is one of the cheapest ways to make output usable.

Think first, then answer

For problems with several steps — a schedule with clashing constraints, a budget, a tricky decision — asking the model to reason before answering often improves the result. This is chain-of-thought prompting; a 2022 paper found that simply adding “Let’s think step by step” improved accuracy on many reasoning tasks.

Many assistants now include “reasoning” or “thinking” modes that do this automatically As of 2026. With those, the magic phrase matters less; what still helps is asking for a plan first so you can correct the approach before the model commits to it.

Chains and reusable templates

Big jobs go better as a sequence of smaller prompts — prompt chaining. Each step produces something you can check before it feeds the next. Tap each step of this example chain for a 1,500-word proposal.

Once a prompt works, save it as a template with blanks in square brackets. Many assistants let you store these as saved prompts, custom instructions or project instructions — see Working with Context. For a set of ready-made templates to start from, see the Prompting Cheat Sheet.

That last line — inviting questions — is one of the most underused tricks. It turns a guess into a short conversation.

Key terms on this page

Sources & further reading

• Anthropic, Prompting best practices — clarity, examples, XML tags, roles, thinking and chaining (accessed 2026).
• OpenAI, Prompt engineering guide (accessed 2026).
• Google, Prompt design strategies, Gemini API documentation (accessed 2026).
• Brown, T. et al., “Language Models are Few-Shot Learners” (2020).
• Kojima, T. et al., “Large Language Models are Zero-Shot Reasoners” (2022) — the “Let’s think step by step” paper.
• Zheng, M. et al., “When ‘A Helpful Assistant’ Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models”, Findings of EMNLP (2024).