How to Write Prompts for AI Agents: 30 Examples That Get Results (2026)

Most people prompt an AI agent the way they would text a friend: a sentence, a hope, and a vague sense of disappointment when the result is generic. That worked badly with chatbots and it works worse with agents, because an AI agent does not just answer, it acts: it plans steps, uses tools, builds things and makes decisions along the way. A weak prompt does not produce a weak paragraph; it produces the wrong app.

This guide teaches prompt writing for AI agents as a briefing skill rather than a trick. You will get the six-part brief that professional agent users rely on, 30 prompt examples you can copy and adapt for research, software, marketing, operations, documents and personal admin, the techniques that consistently improve agent output, and the mistakes that quietly ruin it. The examples are written for Jobbit, a multipurpose AI agent, but they work with any capable agent, from ChatGPT and Claude to coding tools like Cursor and Claude Code.

Why prompting an agent is different from prompting a chatbot

A chatbot prompt asks for words. An agent prompt delegates an outcome. That changes what the prompt needs to contain:

• A finish line. Agents keep working until something tells them to stop. If you do not define "done", they either stop early or gold-plate.

• Constraints that bound the actions. An agent with a browser and a budget can do a great deal. Say what it may not do.

• Inputs it can use. Files, links, examples, your brand voice, your existing data. Agents are excellent at using what you give them and poor at guessing what you forgot.

• Decision rules. What should it do when a supplier's price is missing, when two sources disagree, when the code will not compile? Tell it, or it will decide for you.

• Permissions. Which steps need your approval: sending emails, spending money, publishing, deleting.

"Prompt engineering" has grown into what practitioners call context engineering: putting the right information, examples and rules in front of the agent so it can succeed. If the vocabulary is new, our AI agent glossary covers it, and what is agentic AI? explains how agents plan and act.

The six-part agent brief

Use this structure for any job bigger than a quick question. You do not need headings, just make sure each element is present.

1. Outcome. What exists at the end. "A deployed booking page for my yoga studio", not "help me with a website".

2. Context. Who you are, who it is for, what already exists. "I run a two-person accounting firm in Manchester serving tradespeople."

3. Inputs. Files, links, examples, data, preferences. "Use the attached price list; match the tone of my current homepage; here are three competitors I admire."

4. Constraints. Scope, budget, time, tools, style, legal. "UK English, no stock photos of handshakes, must work on mobile, nothing that needs a developer to maintain."

5. Definition of done. How you will judge it. "Test the booking flow end to end, show me screenshots, list anything you could not verify."

6. Permissions and escalation. "Draft the emails but do not send them. Ask me before spending anything. If two sources disagree, show both and recommend one."

A brief like this reads like an instruction to a capable freelancer, which is exactly the point: agents now perform at the level of the brief they are given.

30 prompt examples for AI agents

Each prompt below is a complete brief in miniature. Replace the specifics with your own.

Research and analysis

1. "Research the UK market for mobile dog grooming. I am planning to launch in Bristol. Give me the main competitors within 15 miles, their prices, reviews and gaps in their offer, then recommend a positioning and a starting price list. Cite every source and flag anything you could not verify."

2. "Compare Xero, QuickBooks and FreeAgent for a UK sole trader earning about £60k with 40 invoices a month. Table: price, Making Tax Digital support, invoicing, bank feeds, support quality from recent reviews. Recommend one and explain the trade-offs in plain English."

3. "Read these five supplier proposals (attached). Extract price, lead time, minimum order and warranty into a table, highlight hidden costs, and tell me which two to shortlist and what to negotiate."

4. "Monitor my three main competitors' pricing pages weekly and send me a short summary of any change. Only alert me if a price or plan changes; otherwise stay silent."

5. "Summarise the last 200 reviews of my product on Trustpilot into the five most common complaints and five most praised features, with example quotes and a suggested fix for each complaint."

For technique, see how to run deep research with AI.

Building software and websites

6. "Build and deploy a one-page website for my plumbing business in Leeds: services, service area, reviews section, a quote request form that emails me, and click-to-call on mobile. Use my logo (attached), keep the palette navy and white, UK English. Show me the live link and a mobile screenshot."

7. "Create a booking app for my barber shop: three barbers, 30-minute slots, 9 to 6 Tuesday to Saturday, customers get an email confirmation, I get a daily list. Deploy it and give me an admin login. Test a booking end to end before you show me."

8. "Turn this spreadsheet (attached) into a searchable internal tool where my team can filter jobs by status and update them. Password protected, no public access. Explain how to add a user."

9. "Add a Stripe payment step to my existing quote form so customers can pay a £50 deposit. Keep the current design. Test with a test card and show me the receipt email."

10. "Audit this app you built for security and performance: check authentication on every page, form validation, exposed keys, and mobile load time. Fix what you find and list what you changed."

The approach behind these is explained in what is vibe coding? and how to build a SaaS with AI.

Marketing and content

11. "Write a 1,500-word blog post targeting 'best time to book a boiler service in the UK', in the friendly tone of my homepage (link). Include an H2 structure, a comparison table, an FAQ section and a meta description under 155 characters. British spelling. No em dashes."

12. "Turn this blog post into a LinkedIn post, three tweets, an Instagram caption and a 60-second video script. Keep each native to its platform; do not just shorten the article."

13. "Draft a 5-email welcome sequence for new subscribers to my meal-prep service. Goal: first order within 14 days. Each email under 150 words, one clear call to action, subject lines A and B for testing."

14. "Generate six product photos of this candle (attached) on a marble kitchen counter in soft morning light, square format, consistent style, no text. Then create a hero banner using the best one with the headline 'Slow evenings, made simple'."

15. "Review my homepage (link) against my top three competitors for messaging clarity, calls to action and trust signals. Give me a prioritised list of ten changes with the exact copy to use."

Business operations and automation

16. "Every morning at 8, read my inbox, list new client emails with a one-line summary and a suggested reply, and flag anything mentioning 'invoice', 'urgent' or 'cancel'. Do not send anything."

17. "When a new enquiry arrives through my website form, research the company, score it 1 to 5 for fit based on these criteria (attached), draft a tailored reply and add it to my CRM. Send the reply only for scores of 4 and above; otherwise leave it as a draft."

18. "Create an invoice for Client X: 12 hours of consulting at £85 per hour plus £40 travel, due in 14 days, my usual template (attached), and email it to them with a polite note. Then set a reminder to chase if unpaid on day 15."

19. "Build me a weekly report that pulls sales from Stripe, ad spend from Google Ads and website visits from Analytics into one page with the three numbers that matter and a two-sentence commentary. Deliver it every Monday at 9."

20. "Set up a customer support assistant for my Shopify store that answers questions about delivery, returns and sizing using my policies (attached), hands anything else to me with a summary, and never promises refunds."

More ideas in 50 AI automation ideas for small businesses.

Documents, data and admin

21. "Extract every line item from these 40 receipts (attached) into a spreadsheet with date, vendor, amount, VAT and category using UK Self Assessment categories. Flag any receipt you could not read."

22. "Read this 60-page contract and give me the ten clauses I should care about as the smaller party, in plain English, with the page reference and a suggested amendment for each. This is not legal advice; I will take it to a solicitor."

23. "Turn these meeting notes into a decision log, an action list with owners and dates, and a two-paragraph summary for people who were not there."

24. "Clean this customer list: remove duplicates, standardise phone numbers to UK format, fix obvious typos in emails, and show me the rows you changed."

25. "Draft a one-page proposal for the attached brief, using our past proposal (attached) for tone and structure, priced at £4,800 with three milestones. Leave placeholders for anything you do not know."

Personal productivity

26. "Plan a three-day work trip to Edinburgh for the 14th to 16th: train from London around 8am, a hotel under £150 a night near the conference centre, and a dinner reservation for four on the 15th. Show me options and book nothing until I confirm."

27. "Reorganise my week: here is my calendar and my to-do list. Protect two 90-minute focus blocks a day, move anything movable, and tell me what to drop."

28. "Every Friday, summarise what I shipped this week from my task list and calendar into a short update for my clients and a private note of what slipped and why."

29. "Learn how I write from these ten emails, then draft replies to the five unanswered messages in my inbox in that voice. I will review before anything goes out."

30. "I want to learn enough SQL to query my own sales data. Build me a two-week plan with 20 minutes a day, using my actual data, and quiz me at the end of each week."

Techniques that make agent prompts work

• Ask for the plan before the work. "Show me your plan and any questions before you start" catches misunderstandings when they cost seconds instead of hours.

• Give examples, not adjectives. "Like this" with a link or attachment beats "professional but friendly" every time.

• Stage the autonomy. First run: draft and ask. Second run: do and report. Third run: do and only flag exceptions. Trust is earned per task.

• Make verification part of the job. "Test it", "cite it", "show me screenshots", "list what you could not confirm". Agents that must prove their work make fewer claims they cannot back.

• Say what to do when unsure. "If a price is missing, write 'not published' rather than estimating." Without a rule, the agent will guess.

• Iterate in plain English. "Make the header smaller, warmer tone, drop the third section." You do not need to restate the whole brief.

• Save what works as standing instructions. Your business context, voice and rules become the agent's memory, so every future prompt starts from a full brief.

• Split truly separate jobs. One prompt for research, another for the build. Agents handle big jobs by delegating to subagents, but they still need one clear goal at a time; see AI subagents explained.

Prompt mistakes that make agents fail

• No outcome. "Help me with marketing" gives the agent nothing to finish.

• Hidden context. The agent does not know your prices, your customers or your history unless you tell it or it can read it.

• Everything at once. Twelve tasks in one paragraph with no priorities produce twelve mediocre results.

• No constraints. Missing budgets, style rules or scope lead to gold-plated or off-brand work.

• Trusting without checking. For anything published, paid or sent, ask for sources and review before it goes out.

• Restarting from scratch. Correct the agent instead of re-prompting; the context it has built is valuable.

• Prompting for a chatbot. "Write me a paragraph about X" wastes an agent that could have researched X, built the page and scheduled the posts.

Prompting on Jobbit

Jobbit is designed around the brief, not the prompt. Describe the outcome in one message, attach what you have, and the agent plans the work, delegates pieces to subagents, uses a browser, a code sandbox, hosting, image generation and document tools as needed, verifies the result and hands it back finished. Your context and preferences persist, so the tenth job starts with the knowledge of the first nine. When part of a job needs a person, a designer, a bookkeeper, a tradesperson, the Jobbit Pro network supplies the professional inside the same chat with escrow-protected payment. Start free at jobbit.uk.

Frequently asked questions

What makes a good prompt for an AI agent?

A good agent prompt describes an outcome, gives context and inputs, sets constraints, defines what "done" looks like, and says what needs approval. It reads like a brief to a capable freelancer rather than a question to a search engine.

Do I need to learn prompt engineering to use AI agents?

No formal training is needed. The skill is briefing clearly, which most people already do with colleagues. Use the six-part structure in this guide, ask the agent for its plan first, and correct it in plain English; that covers almost everything that used to be called prompt engineering.

How long should a prompt for an AI agent be?

As long as the brief needs and no longer. A good research or build prompt is usually three to eight sentences plus attachments. Length is not the goal; completeness is. If you find yourself writing pages, split the job into separate prompts.

Can I reuse prompts for AI agents?

Yes, and you should. Save prompts that produced good results as templates, and store your business context, voice and rules as standing instructions so the agent applies them automatically. Most of the 30 examples above are designed to be reused with a few words changed.

What is the difference between a prompt and context?

The prompt is the request; context is everything the agent can see while working on it: your instructions, files, examples, tool results and memory. With agents, context usually matters more than the wording of the request, which is why attaching the right inputs improves results more than rephrasing.

Pick one prompt from the list, paste it into Jobbit with your own details, and see the difference a proper brief makes. Start free.