Preparing your workspace
Jobbit
Guides11 min read

What Is Prompt Engineering? 10 Techniques That Work in 2026

What is prompt engineering? A plain-English guide to 10 techniques OpenAI, Anthropic and Google recommend in 2026, the myths research has tested, and examples.

What Is Prompt Engineering? 10 Techniques That Work in 2026
Read in:

Prompt engineering is the skill of writing instructions that get an AI model to produce what you need, reliably. OpenAI calls it "the process of writing effective instructions for a model". In 2026 it rests on a handful of plain habits: be specific, give context, show an example, say what format you want and test the result. This guide explains 10 techniques the model makers recommend, and which popular tricks research says to drop.

A cafe owner types "write a post about our new menu" and gets four cheerful paragraphs she would never publish. Her nephew types three more sentences into the same tool, saying who the customers are, how the cafe talks and that it must fit one Instagram caption, and gets something she posts that afternoon. Same AI, same price. The difference has a grand name, prompt engineering, and a reputation for secret formulas. The reality is closer to briefing a new member of staff. The government's 2026 AI adoption survey found 32% of UK businesses now use AI, double the year before, and 67% of those that do say limited skills or knowledge holds them back.

Below: how we checked this guide, the 10 techniques in one table, each technique with an example, what changed with the newest models, the myths that were tested, whether it is still a job and where to learn it free.

Skip the trial and error. Describe the job once on Jobbit and Plan mode shows the steps before the agent starts. Free plan, 100 credits a day, no card.

How we checked this guide

  1. The model makers' own guides, read on 18 November 2026. OpenAI's prompt engineering guide, Anthropic's prompting best practices, Google's prompting strategies for Gemini, Microsoft Learn and IBM.
  2. Independent research. Three reports from the Wharton School's Generative AI Labs that tested common prompting advice many times over, and a 2026 preprint from IBM researchers.
  3. Where the vendors disagree, we say so. The advice on step-by-step reasoning changed during 2025 and 2026.
  4. Examples are ours, written for small firms. Vendor wording is in quotation marks.
  5. UK figures come from the government's AI Adoption Survey 2026 and the Office for National Statistics.

10 prompt engineering techniques at a glance

TechniqueWhat you add to the messageWho recommends itWhen it matters most
1. Be specificLength, audience, purpose, what to leave outAnthropic, Google, MicrosoftAlways
2. Give contextWho you are, who it is for, why you need itAnthropic, Google, OpenAIAnything about your business
3. Set a role"You are an experienced bookkeeper"Anthropic, OpenAITone and expertise
4. Show examplesTwo or three samples of what good looks likeOpenAI, Anthropic, Google, MicrosoftMatching your style or format
5. Separate the partsLabels or tags around instructions and pasted textOpenAI, Anthropic, MicrosoftLong prompts with documents
6. Name the format"A table with 3 columns", "under 80 words"Google, MicrosoftAnything you will reuse
7. Supply the sourcePaste the facts and say "use only this"OpenAI, Microsoft, GooglePrices, policies, figures
8. Break it into stepsOne prompt per stage, check betweenAnthropic, MicrosoftBig jobs
9. Ask for a check"Before you finish, verify against these points"AnthropicSums, facts, code
10. Test and refineRun it again, change one thing, compareGoogle, OpenAI, AnthropicPrompts you use every week

The 10 techniques, with examples

1. Be specific

Microsoft's guide puts it in a sentence: "Be specific. Leave as little to interpretation as possible." Anthropic offers a test: show your prompt to a colleague with no background and ask them to follow it. If they would be confused, the AI will be too.

  • Vague: "Write an email about the price rise."
  • Specific: "Write a 120-word email to existing customers of a dog grooming salon. Prices rise 6% on 1 January. Warm, plain English, no apology, one line thanking them for their loyalty."

2. Give context

Anthropic suggests treating the AI as "a brilliant but new employee who lacks context on your norms and workflows". It also recommends explaining why, since a reason lets the model handle cases you did not list. "Keep sentences short because customers read this on a phone at the counter" works better than "keep sentences short".

3. Set a role

One sentence is enough. Anthropic's guide says setting a role "focuses" behaviour and tone and that "even a single sentence makes a difference". Try "You are a UK employment adviser explaining this to a first-time employer."

4. Show examples

Google's guide is the most insistent: "We recommend to always include few-shot examples in your prompts." Anthropic suggests three to five, varied enough that the model copies the pattern and not one sample. Paste two product descriptions you like and ask for a third in the same voice. The technical name is few-shot prompting.

5. Separate the parts

When a prompt contains instructions plus a pasted document, label them. OpenAI and Anthropic both recommend simple tags or headings so the model can tell your request from your material. Even plain labels work: "INSTRUCTIONS:" and "CUSTOMER EMAIL:".

6. Name the format

Say what shape the answer should take: a table, five bullets, a subject line plus body, under 80 words. This one has evidence behind it. When Wharton researchers removed the formatting instruction from their test prompts, performance fell.

7. Supply the source

For anything factual about your business, paste the facts and tell the AI to use only those. Microsoft says it "strongly" recommends providing grounding data and suggests giving the model a way out, such as "reply 'not found' if the answer is not in the text". This is the single best defence against invented answers. More in why AI makes things up.

8. Break it into steps

A marketing plan in one prompt comes out thin. Ask for the audience first, check it, then the offers, then the calendar. Anthropic notes that chaining prompts is still useful "when you need to inspect intermediate outputs", even though newer models handle more in one go.

9. Ask for a check

Anthropic's guide suggests ending with something like "Before you finish, verify your answer against" your criteria, and says this "catches errors reliably, especially for coding and math". For a quote: "Check that the line items add up to the total and that VAT is 20%."

10. Test and refine

Google's guide says simply: "Prompt engineering is iterative." The Wharton team found that the same prompt can succeed on one attempt and fail on the next, so one good answer proves little. For a prompt you will reuse, run it three times and change one thing at a time. Then save it. Our guide to ChatGPT custom instructions shows where saved instructions live.

What changed with the newest models

Advice from 2023 can now make results worse.

  • Stop asking it to "think step by step". The trick is called chain of thought. OpenAI's guide for its reasoning models says to keep prompts simple and avoid asking for step-by-step reasoning. Microsoft says the technique applies only to non-reasoning models. Anthropic says to rely on the model's built-in thinking. The newest models reason on their own.
  • Drop the shouting. Anthropic notes that its recent models are more responsive to instructions and advises softening aggressive wording such as "CRITICAL: You MUST". Capital letters now cause over-reaction.
  • Expect literal reading. If you want more than you asked for, ask. Anthropic's advice is to "explicitly request it" and not rely on the model to guess.
  • Context is the new word. In September 2025 Anthropic described "context engineering" as "the natural progression of prompt engineering": choosing what information the AI has in front of it, not only how the request is worded. For a small firm that means giving it your price list, your past emails and your policies.

A September 2026 preprint by IBM researchers tested 8 prompting techniques across 27 model set-ups and about 4,300 questions. With the models' reasoning switched on, no technique beat a direct question by more than about half a percentage point. Their test covered multiple-choice questions only, but the direction is clear: clever wording matters less each year, and clear briefing matters as much as ever.

Four myths that were tested

The Wharton School's Generative AI Labs ran common tricks through thousands of repeated trials in 2025.

  • "Offer it a tip." Their August 2025 report is titled plainly: threats and rewards do not work. It found no meaningful overall improvement from threatening models or offering them money.
  • "Be rude" or "be polite". Politeness swung individual answers both ways, and across the whole test the effects cancelled out. Be polite if you like. It is not a technique.
  • "Think step by step always helps." Their June 2025 report found modest gains with more variability on older models and marginal benefit on reasoning models, at a cost of 20% to 80% more time.
  • "There is a perfect prompt." The same prompt gave different results across repeated runs, so testing beats wordsmithing.

Is prompt engineering still a job?

As a job title it has faded. Fortune reported in May 2025 that a role once advertised with salaries up to $200,000 had largely disappeared, quoting an Indeed economist: prompt engineering "is still definitely a good thing to have, but it's not an entire title". Korn Ferry, reading a Microsoft survey of 31,000 workers, said it came second from bottom among new roles firms planned to add.

As a skill it has spread to everyone. The Office for National Statistics found 55% of employees in Great Britain using AI for work or education in mid 2026. The person who briefs the AI well gets more out of the same subscription.

Where to learn it free

  • Anthropic's prompting best practices and its interactive tutorial on GitHub, nine short chapters. The tutorial was written for older models, which Anthropic itself notes.
  • OpenAI's prompt engineering guide, free to read, with a separate page for reasoning models.
  • Google's prompting strategies for Gemini, free to read.
  • DeepLearning.AI's short course with OpenAI, 1 hour 40 minutes, free at the time of writing but aimed at people who know some Python.

For ready-made material, start with our 30 prompt examples for AI agents and 40 templates by industry.

How an AI agent changes the job of prompting

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all respond to the 10 techniques above. An agent that carries out the work moves some of the effort from wording to checking.

  • It asks before it acts. Plan mode on Jobbit lists the steps for your approval, which catches a vague brief before credits are spent.
  • It keeps your context. Projects hold your files and standing instructions, so techniques 2 and 7 are set up once.
  • You pick the model. Paid plans let you choose a lighter model for simple jobs and a stronger one for hard ones. Go costs £7 a month with 2,000 credits. See the plans.
  • It finishes the job. The result is a document, a spreadsheet or a working site, not a paragraph to paste somewhere.
  • A person can take over. Work that needs a professional goes to Jobbit's human network.

One caution as you get better at prompting: instructions can also arrive hidden inside pages and emails the AI reads. Our guide to prompt injection explains the risk.

Frequently asked questions

What is prompt engineering in simple terms?

It is writing clear instructions for an AI so it gives you a useful answer the first time. In practice that means saying exactly what you want, who it is for, what format to use and what facts to rely on, then checking and adjusting.

What are the main prompt engineering techniques?

Be specific, give context, set a role, show examples, separate instructions from pasted material, name the output format, supply your own source text, split big jobs into steps, ask for a self-check and test more than once. OpenAI, Anthropic and Google all recommend most of these.

Does telling an AI to think step by step still work?

Mostly not with current models. OpenAI and Microsoft advise against it for reasoning models, and Anthropic says to rely on the model's own thinking. Research from Wharton in 2025 found only marginal gains on reasoning models and slower answers.

Is prompt engineering a good career in 2026?

As a stand-alone job title it has largely faded, according to reporting by Fortune and Korn Ferry in 2025. As a skill inside other jobs it is valuable, because most office workers now use AI and the quality of the brief decides the quality of the result.

Do I need to learn prompt engineering to use AI for my business?

You need the basics, which take an afternoon. The 10 habits on this page cover most everyday work. Saving your business details and house style as standing instructions removes much of the repetition.

What is the difference between prompt engineering and context engineering?

Prompt engineering is about how you word the request. Context engineering is about what information the AI has when it answers: your documents, past examples and rules. With an agent on Jobbit, Projects hold that context for every task. Try it free with 100 credits a day and no card.

Related guides