How to Write AI Prompts That Get Great Results

If you've ever typed "write me a marketing email" into ChatGPT and gotten back something that sounds like a press release written by a fridge, the problem almost certainly isn't the AI. Learning how to write AI prompts - with real context, clear instructions, and an example or two - is the highest-leverage skill a business owner can pick up in 2026, and it takes an afternoon, not a course. This guide gives you the framework, the common mistakes, and ten copy-paste templates built around real Okanagan businesses.

Why your AI results are mediocre (it's usually the prompt)

Vague prompts create what researchers call a priming problem. Ask "How can I improve my business?" and the model has nothing to anchor on, so it defaults to generic filler - "sell more products," "improve customer service." The single most documented mistake in prompt writing is assuming the AI has context it doesn't possess: your industry, your audience, your tone, your goal. It knows none of that unless you say it.

The data backs this up. Over 83% of surveyed users report that specific, structured prompts improve their results, and a 2025 arXiv study on prompt detail found output quality rises with detail - peaking at moderate detail. Dumping every fact about your company into one prompt doesn't help either; you want a well-briefed paragraph, not a memoir.

The most common prompting mistakes, in rough order of damage:

1. No context - the AI doesn't know who you are, who you're writing for, or why
2. No format instructions - you wanted three bullet points; you got 800 words
3. Asking for five things in one prompt - quality drops on all five
4. No examples - the model guesses at your voice and guesses wrong
5. Accepting the first draft - then blaming the tool
6. Pasting confidential client data into consumer AI tools - one of the top documented small-business AI mistakes

How to write AI prompts: role, context, task, format

Every good prompt framework - and the internet is drowning in acronyms - boils down to the same anatomy: role, context, task, constraints, format.

The two-minute version is RTF (Role, Task, Format), which honestly solves about 80% of daily business use:

• Role: who the AI should act as. "Act as an email marketer for a family winery in East Kelowna."
• Task: exactly what you want, with specifics. Not "write something about marketing" but "write a 300-word newsletter announcing our fall wine club release to existing members."
• Format: what the output looks like. "Subject line plus three short paragraphs. Warm, unfussy tone. No exclamation marks."

Add context (who the audience is, what they already know, what you're trying to achieve) and constraints (word count, things to avoid, what "done" looks like) and you've covered everything the fancier frameworks like CRAFT and RISEN offer.

Watch the difference:

Weak: "Write a social media post about our winery event."

Strong: "Act as a social media manager for a small family winery in East Kelowna, BC. Write an Instagram caption promoting our patio-season release party on July 18. Audience: locals aged 30-55 who follow Okanagan wine accounts. Goal: RSVPs through the link in bio. Tone: warm and local, not salesy. Under 100 words. Don't use hashtag spam - maximum three hashtags. Don't invent details about the wines."

Same tool, wildly different output. Note the negative constraints at the end - telling the model what not to do ("don't invent statistics," "no exclamation marks") dramatically reduces off-target output.

Anthropic's own prompting documentation says it in three lines: state the task explicitly, say who the output is for, and define what "done" looks like - the same anatomy whether you're learning how to prompt Claude, ChatGPT, or Gemini.

Give examples: the single highest-leverage trick

If you take one thing from this guide, take this: show the AI an example of what good looks like. Practitioners call this few-shot prompting, and it's repeatedly cited as the most powerful single technique in prompt engineering for beginners. Even one good example transforms output quality.

The business translation is beautifully simple: paste your best past work and say "match this."

• Writing a newsletter? Paste your two best past newsletters: "Write the next one in this voice."
• Quoting a job? Paste a past quote your client loved: "Use this structure and tone for the new quote below."
• Replying to reviews? Paste three replies you were proud of: "Respond to this new review in the same style."

Anthropic's guidance: 2-5 relevant, diverse examples; when prompting Claude specifically, wrap them in <example> tags so the model separates the demonstration from the instruction.

Iterating: treat the AI like a smart new hire, not a vending machine

The vending-machine mindset - insert prompt, receive perfection - is where most people quit. Treat the AI instead like a smart new hire on day one: capable and fast, but needing direction and feedback. Research on ChatGPT use found iterative, multi-prompt sessions with a critical eye significantly outperform one-shot prompting. In practice:

1. One task per prompt. Don't ask for a blog post, five captions, and an email at once - chain them: outline, then draft, then adapt.
2. Ask what it needs. Before a big task, try: "Before you write this, ask me any questions that would improve the result."
3. Give feedback like a manager. "Shorter. Warmer. Cut the jargon. The second paragraph is the real opener - start there." The model takes revision notes better than most humans.
4. Ask it to critique itself. "Review your draft above. What's weak? Rewrite it fixing those weaknesses."
5. For logic-heavy work, ask it to think first. "Think step by step before answering" still earns its keep on math, pricing, and multi-step planning tasks.

Ten copy-paste prompt templates for common business tasks

These are AI prompt examples for business you can steal today - written as real Okanagan scenarios rather than "[INSERT PRODUCT]" mad-libs. Swap in your details.

1. Winery club newsletter - "Act as an email marketer for a family winery in East Kelowna. Write our monthly wine club newsletter announcing [release]. Audience: existing club members who visit 2-3 times a year. Tone: warm, knowledgeable, zero pretension. 250 words, subject line included. Here's a past newsletter to match: [paste]."

2. Landscaping quote follow-up - "Act as the owner of a landscaping company in West Kelowna. Write a follow-up email to a homeowner who received our quote for [job] five days ago and hasn't replied. Friendly, no pressure, one clear next step. Under 120 words."

3. Kelowna event promo - "Act as a community events promoter. Write a promo post for [event] in Kelowna on [date]. Audience: local founders and professionals. Goal: registrations. Include what attendees will walk away with. 80-120 words, no hashtag spam."

4. Google review reply - "Act as the owner of [business] in [city]. Write a reply to this Google review: [paste]. If positive: thank them specifically, no copy-paste energy. If negative: acknowledge, take it offline, stay classy. Under 80 words."

5. Cold email with a hook - "Write a cold email to [role] at [company type] in the Okanagan offering [service]. Open with one specific, non-obvious insight about their industry or region - not flattery. One clear ask. Under 130 words. Sound like a helpful expert, not a pitch deck." (That one unique insight is the documented resistance-lowerer in AI prompts for sales emails.)

6. Sales objection handler - "I'm selling [thing] to [people]. Their main problem: [problem]. Our solution works because [reason]. The biggest objection I hear: [pushback]. Write copy that addresses the objection head-on and sounds like a helpful expert, not pushy."

7. Blog post outline - "Act as a content strategist for [business]. Outline a blog post targeting the search 'how do I [customer question]'. Include an H2 structure, the questions a real searcher wants answered, and a suggested title under 60 characters."

8. Job posting - "Write a job posting for a [role] at [business] in Kelowna, BC. Emphasize [what makes you a good employer]. Honest about the hard parts of the job. Plain language, no 'rockstar/ninja' clichés. Include salary range [range]."

9. Meeting-to-action-items - "Here are my raw meeting notes: [paste]. Extract: decisions made, action items with owners and deadlines, and open questions. Format as three bulleted lists. Don't add anything that isn't in the notes."

10. Plain-language rewrite - "Rewrite this for a customer with zero industry knowledge: [paste]. Grade 8 reading level, keep it accurate, keep my key points, cut the jargon. Same approximate length."

Why the local flavour? Role-priming with a specific business and place measurably sharpens output - and the scenarios are real: BC has 306 licensed grape wineries, more than 200 of them in the Okanagan Valley.

How to write AI prompts for long documents, data, and files

The bigger unlock for a business is feeding the AI your actual documents - leases, supplier contracts, financial exports, past proposals - and prompting against them.

Two rules change everything here:

• Document first, question last. Anthropic's docs note that putting long documents at the top of the prompt and your instructions at the end can improve response quality by up to 30%. Paste the contract, then ask "What are the three biggest risks to me in this agreement?"
• Be specific about the extraction. "Summarize this" gets mush. "List every date, deadline, and dollar amount in this contract as a table" gets something you can act on.

Practical limits as of mid-2026 (these change often): Claude.ai accepts files up to 30MB, 20 per conversation, with a 200K-token context window - roughly 500 pages - versus ChatGPT's 128K. One quirk: for PDFs over 100 pages, charts and images are only analyzed in the first 100, though the text beyond is still processed. Details in the Claude help centre.

When prompts aren't enough: projects, custom instructions, and agents

If you're re-typing the same context every morning, stop. There's a ladder of tools above the prompt box, and the major consumer plans that include them cluster around $20/month.

System prompts explained in one analogy: a system prompt is the AI's permanent job description - the employee handbook set behind the scenes - while your prompt is the individual task or customer question. Custom instructions are effectively your personal system prompt.

Claude Projects shine for sustained document-heavy work - they use retrieval to pull the relevant excerpts from your uploaded files, so a big knowledge base doesn't overflow the context window. Load your services, pricing, brand voice guide, and three best past proposals into one project, and every chat inside it starts fully briefed.

Is prompt engineering dead in 2026? The trick-hunting era is: 82% of IT and data leaders say prompt tricks alone aren't sufficient for production AI, and the professional craft has moved to context and flow engineering. But for a business owner, clearly articulating role, context, task, and format is precisely the part that survived - it's clear thinking, written down. To see where the ladder tops out, our guides on what vibe coding is and building an app with AI without writing code show what happens when prompting becomes building; then compare Claude Code vs Cursor.

Build your own prompt library in 30 minutes

A prompt that worked is an asset. Most businesses throw those assets away daily. Here's the 30-minute fix:

1. Minutes 0-5: pick a home. A shared Notion page, a Google Sheet, or a pinned Slack channel - whatever your team already opens daily. Notion's marketplace even has a free AI prompt library template.
2. Minutes 5-10: set up six fields per prompt. Name, use case, the prompt itself, one example output, owner, last-tested date.
3. Minutes 10-25: seed it with ten prompts. Customize the templates above, plus any prompt from your chat history that produced work you shipped.
4. Minutes 25-30: make a rule. Any time a prompt produces work you'd happily publish, it goes in the library within 24 hours.

The payoff: dramatically faster onboarding and a consistent brand voice across every channel - prompt templates for small business that compound with every entry.

The local stakes are real, too. Only about 8% of Canadian SMEs have adopted AI, versus 29-42% in the Nordics - so the Kelowna business that gets systematically good at prompting this year is competing against neighbours who mostly haven't started. Canada's new AI for All strategy even includes free AI literacy training, so levelling up costs time, not money.

Key takeaways

• Mediocre AI output is usually a context problem, not a model problem - state your industry, audience, tone, and goal every time
• The minimum viable prompt is Role, Task, Format; add context and constraints for anything that matters
• One good example beats a paragraph of description - paste your best past work and say "match this"
• Iterate like a manager: one task per prompt, ask the AI what it needs, give revision notes
• For long documents, put the document first and the question last - it can improve quality by up to 30%
• Moving from repeated prompts to custom instructions and Projects installs your context once instead of retyping it forever
• A shared prompt library turns individual wins into a team asset in 30 minutes

Frequently asked questions

Why does AI give different answers to the same prompt?

Models pick the next word probabilistically among likely candidates (controlled by a setting called temperature), so identical prompts can branch into different outputs. Vague prompts amplify this variability; specific, structured prompts produce far more consistent results.

Do I still need prompt engineering in 2026?

The magic-phrase era is over - 82% of IT and data leaders say prompt tricks alone don't cut it for serious systems. But the durable core of prompt engineering is clearly articulating role, context, task, and format, and that skill matters more than ever. Think "clear requirements," not "secret incantations."

What is a system prompt?

It's the AI's permanent job description, set behind the scenes by the tool's developer - defining how the assistant behaves before you type anything. Your message is the individual task on top of it. Custom instructions in ChatGPT or Claude act as your personal layer of system prompt.

How long should an AI prompt be?

Research shows quality rises with detail and peaks at moderate detail - a well-briefed paragraph beats both a one-line request and a 2,000-word info-dump. Include role, audience, goal, format, and key constraints.

Is it safe to paste client data into AI tools?

Not by default. Pasting confidential client information into consumer AI tools is one of the most common documented small-business AI mistakes. Check your plan's data-handling and training settings, anonymize where you can, and use business-tier plans for sensitive work.

---

Prompting well is a practice skill, and it's faster with people to practice alongside. Our events have featured speakers whose whole craft is prompt-adjacent - a sales copywriter, a content creator with 8.8 million followers - and the room is full of Okanagan founders testing what actually works. Join the Kelowna Founders Club free and build the skill with company.