You already know how to do this

Prompt engineering has been turned into this whole mystical discipline with courses and certifications and LinkedIn influencers charging $500 for webinars. It's not that complicated. If you can explain what you want to a reasonably smart colleague, you can write a good prompt.

The problem isn't that prompting is hard. The problem is that most people write prompts the way they'd type a Google search. Short, vague, and hoping the computer figures out the rest. That worked for Google (sort of). It doesn't work for AI.

Here are six patterns that will make your prompts dramatically better. Each one takes about 30 seconds to apply. No PhD required.

Pattern 1: Role prompting

The vague prompt: "Write me a marketing email."

The better prompt: "You are a senior email copywriter who specialises in SaaS products. Your style is conversational but professional, you favour short sentences, and you always include a single clear call to action. Write a marketing email for a project management tool aimed at small agencies."

Why it works: When you give the AI a role, it constrains its output to match that role's expertise and style. Without a role, you get generic output. With a role, you get something that sounds like it was written by someone who actually does that job.

The trick most people miss: Be specific about the role's characteristics, not just the job title. "You are a copywriter" gives different results than "you are a copywriter who hates buzzwords and writes in an informal British tone." The adjectives matter more than the noun.

Works best with: claude and chatgpt. Both respond well to role prompts. Claude tends to adopt the role more subtly while ChatGPT goes harder on it, sometimes too hard.

Pattern 2: Chain of thought

The vague prompt: "Should I use React or Vue for my project?"

The better prompt: "I'm choosing between React and Vue for a team of 3 developers who are familiar with React but not Vue. The project is a customer dashboard with real-time data, charts, and complex forms. Budget is tight so onboarding time matters. Think through this step by step, considering team experience, project requirements, ecosystem support, and long-term maintenance."

Why it works: "Think step by step" or "think through this" forces the AI to show its reasoning rather than jumping to a conclusion. This produces more nuanced, accurate answers because the AI catches its own errors during the reasoning process.

When to use it: Any time you want analysis, not just an answer. Technical decisions, strategy questions, debugging. Whenever the question has multiple factors to weigh.

The surprising bit: Just adding "think step by step" to the end of almost any prompt improves the answer quality. It's one of the most reliably useful tricks in all of prompt engineering, and it takes four words.

Pattern 3: Few-shot examples

The vague prompt: "Write product descriptions for my shop."

The better prompt: "Write product descriptions for my online shop. Here's an example of the style I want:

Merino Wool Beanie - Slate Grey Knitted from 100% New Zealand merino wool. Warm enough for January, light enough that you won't overheat on the train. One size fits most heads, including the unreasonably large ones.

Now write similar descriptions for: (1) A leather card holder in tan, (2) A ceramic coffee mug in speckled blue, (3) A cotton tote bag in natural."

Why it works: Showing the AI what you want is always more effective than describing what you want. One example is good. Two or three examples establish a pattern that the AI follows reliably.

The important detail: Your examples should demonstrate every quality you care about. In the example above, the beanie description shows the format (bold title, short paragraph), the tone (casual, slightly funny), the structure (material, practical benefit, personality), and the length (about 40 words). The AI picks up on all of these.

Works especially well for: Social media posts, product descriptions, email templates. Anything where consistency of voice matters.

Pattern 4: Constraint setting

The vague prompt: "Explain how DNS works."

The better prompt: "Explain how DNS works in under 150 words, using language a 14-year-old would understand. No jargon. Use one analogy. Don't start with a definition."

Why it works: Without constraints, AI tends to over-explain. It gives you the Wikipedia version when you wanted the pub version. Constraints force conciseness and clarity.

My favourite constraints to add: - Word or sentence limits ("in under 100 words") - Audience level ("explain like I'm a junior developer" or "assume I'm an expert") - Tone ("casual," "formal," "funny," "blunt") - Format restrictions ("no bullet points" or "only bullet points") - Content restrictions ("don't mention competitors" or "focus only on the downsides")

The counterintuitive truth: More constraints produce better output. It feels backwards, like you're limiting the AI. But constraints are actually giving it more information about what you want. An unconstrained AI is guessing. A constrained AI is targeted.

Pattern 5: Output format control

The vague prompt: "Compare these three project management tools."

The better prompt: "Compare Notion, ClickUp, and Linear for a 5-person startup. Reply as a markdown table with these columns: Tool | Best For | Worst At | Monthly Cost (5 users) | Learning Curve (1-5). Below the table, give a one-sentence recommendation."

Why it works: Telling the AI exactly what format to use means you get exactly what you need without having to reformat it yourself. Tables, JSON, bullet points, numbered lists, headers. Specify it and you get it.

Particularly useful with: chatgpt and claude for generating structured data, comparison tables, and formatted documents. gemini is also good at this but occasionally ignores format instructions on complex requests.

Advanced trick: You can combine format control with few-shot examples. Show the AI one row of your desired table, then ask it to fill in the rest. This works brilliantly for consistent data entry tasks.

Pattern 6: The revision loop

This isn't a prompt pattern so much as a workflow pattern, but it's the most important thing on this list.

Most people: Write a prompt. Get an answer. Use the answer or give up.

What you should do: Write a prompt. Get an answer. Tell the AI specifically what to change. Repeat.

Example revision prompts: - "That's close but too formal. Rewrite with a more casual tone and add a joke in the opening." - "Good structure but the third section is too long. Cut it in half and make the key point more prominent." - "I like the content but the format isn't right. Convert this into a numbered list with bold headings for each point."

Why most people skip this: It feels slow. You've already got an answer, why not just use it? Because a two-minute revision conversation usually improves the output by 50% or more. The first response is a rough draft. The third response is usually quite good.

The best revision technique: Be specific about what's wrong. "This isn't good" gives the AI nothing to work with. "The opening is too generic and the second paragraph contradicts the conclusion" gives it everything it needs.

The before and after

Here's a real example showing how these patterns stack up.

Bad prompt: "Write me an about page."

Good prompt: "You are a web copywriter who writes in a warm, slightly irreverent tone. Think about what makes a great about page for a small design studio. Then write an about page for a 4-person design studio called Kindling, based in Bristol. We specialise in brand identity for food and drink companies. In under 250 words. No corporate jargon. Start with something unexpected, not 'Welcome to' or 'We are.' Reference that we started in a pub and still do most of our best thinking there."

The first prompt gives you something that sounds like every other about page on the internet. The second gives you something specific, charming, and actually useful.

The difference took about 45 seconds of extra thought. That's the real lesson. Prompt engineering isn't a skill. It's patience. The willingness to spend one minute thinking about what you want before asking for it.

Which AI to use for what

Quick practical note, since this comes up constantly.

claude is best for writing tasks, analysis, and anything where nuance matters. It follows instructions more precisely and produces more natural-sounding text.

chatgpt is best for quick questions, internet-connected research, and creative brainstorming. It's more willing to riff and explore weird ideas.

gemini is best for tasks involving Google's ecosystem (Gmail, Docs, Search) and for multimodal prompts where you're combining text with images.

perplexity is best for research questions where you need cited sources. Don't prompt-engineer Perplexity. Just ask it questions naturally. Its strength is retrieval, not generation.

None of these patterns are secrets. They're common sense, written down. The difference between people who get great results from AI and people who get mediocre results is almost always this: the great-results people spend 60 seconds writing a detailed prompt instead of 5 seconds writing a vague one.

That's it. That's the whole course. You can have your $500 back.