The Five Prompt Patterns That Outperform Freeform Prompting
Most teams using AI tools at work are still typing into a blank box and hoping for the best. The results are uneven, the quality is hard to repeat, and nothing about how the team works gets better over time. The five prompt patterns below are the ones we teach in every workshop because they consistently outperform freeform prompting. Once a team uses them, prompting becomes a skill the whole team can share, not a personal craft.
Pattern 1: Role. Define who the AI is being
Most prompts skip the most important sentence: who is the AI in this conversation? 'Write me a summary' is a different request from 'You are a senior analyst writing for a busy executive. Summarise this report.' The model has different vocabulary, structure, and confidence depending on the role you assign. Adding a single Role sentence is the cheapest and most reliable improvement you can make to almost any prompt.
A worked example. Change 'Summarise this contract' to 'You are a commercial lawyer reviewing a vendor contract for a mid-sized Australian retailer. Summarise the contract for a CFO who needs to decide whether to escalate to legal. Focus on commercial terms, termination triggers, and any clauses that create ongoing exposure.' Use Role for any task where the audience or stance matters: drafting, review, analysis, explanation. If the output is going to a human with a specific job to do, name that human.
Pattern 2: Structure. Specify the shape of the answer
Models default to long-form prose. If you want bullet points, a table, an email with specific sections, or a decision recommendation with a clear yes-or-no at the top, you have to say so. Specifying the structure is the second cheapest and second highest-impact change you can make. It is also the pattern that turns AI output from 'something I have to rewrite' into 'something I can paste into the workflow.'
A worked example. Change 'Help me write a brief for our content team about the new product launch' to 'Write a content brief in this structure: a one-line product summary, the top three audience segments with the pain each one feels, three message pillars, five content angles per pillar, and the recommended channel and format for each angle.' Use Structure anywhere the output integrates into a downstream workflow: briefs, decision memos, code reviews, intake forms, project updates.
Pattern 3: Examples. Show what good looks like
Telling the model what you want is useful. Showing it is better. Two or three examples of the output you would accept, even if they are old, imperfect, or partial, calibrate the model's sense of voice, depth, and format faster than any amount of instruction. Examples are the pattern that most reliably solves the 'it sounds like AI' problem because the model has something concrete to match.
A worked example. Change 'Write a LinkedIn post about our new feature' to 'Write a LinkedIn post about our new feature. Below are three posts from our brand that performed well. Match the voice, length, and structure. Post one is 80 words, conversational, ends with a question. Post two is 110 words, includes one short bulleted list. Post three is 60 words, single pulled quote.' Use Examples for voice-sensitive output (marketing, sales, support), formats your team has standardised, and anywhere 'what we would write' is the key signal.
Pattern 4: Constraints. Say what to avoid
Models will happily produce output that breaks the rules of your brand, your sector, or your team's preferences if you do not name those rules. Constraints are the prompt's no-go list: jargon to avoid, claims you cannot make, formats you do not use, sources you cannot cite. In regulated industries, Constraints are the pattern that keeps AI output on the right side of policy. In every industry, Constraints prevent the worst version of an AI response from reaching a real person.
A worked example. Change 'Draft a response to this customer complaint' to 'Draft a response to this customer complaint with the following constraints: do not admit liability or use the words sorry or apologise (legal team policy); acknowledge the inconvenience in different language; do not commit to a specific resolution timeline. We will be in touch within five business days. Sign off as the support team, not a named individual.' Use Constraints in regulated industries, customer-facing communications, anything brand-sensitive, and anywhere 'what not to do' carries weight.
Pattern 5: Iteration. Refine without restarting
The biggest mistake in prompting is throwing away a near-miss and starting over. Iteration treats the model like a colleague: keep the parts that work, give targeted feedback on the parts that do not, ask for a specific revision. Two or three iterations is faster than the perfect first prompt and produces better output because you can see what is wrong before you correct it.
A worked example. The initial prompt produces a 400-word content brief. Iteration one: 'This is close. Keep the audience and message pillar sections. Cut the timeline section, it is not relevant. Add a success metric line under each content angle. Tighter language overall, assume the reader is in a hurry.' Iteration two: 'Better. The success metrics under angles 2 and 4 are too generic. Make them numbers we could track in analytics. Leave the rest.' Use Iteration on every output longer than a paragraph, especially anything that will be reviewed or edited downstream. Iteration upstream saves edit time downstream.
How to use the patterns together
The five patterns are not a checklist. They are a vocabulary. Most strong prompts use Role and Structure as the base, layer in Examples when voice matters, add Constraints when something can go wrong, and use Iteration to converge on the final output. The teams that get the most out of AI are not the ones with the longest prompt templates. They are the ones who can move between the five patterns naturally because they have practised them across enough real work.
If you are starting from scratch, add one pattern at a time. Pick a workflow you do weekly. Add Role to your existing prompt for a week. The next week, add Structure. By the end of the month you will have practised all five patterns on a workflow you do every week. That is the kind of practice that compounds.
- Most prompts can be improved by adding a single sentence of Role or Structure
- Examples are the highest-leverage pattern when voice or format matters
- Constraints are the pattern that prevents AI output from creating risk
- Two iterations beat one perfect first prompt almost every time
- Patterns are a vocabulary, not a checklist. Combine them as the task demands.
Do these patterns work across ChatGPT, Claude and Gemini?+
Yes. The patterns are model-agnostic. Wording varies (Claude responds well to XML-style structure tags; ChatGPT and Gemini are fine with plain markdown) but the five patterns translate across all major models.
How long does a good prompt need to be?+
As long as it needs to be, no longer. A two-line prompt with the right Role and Structure can beat a thirty-line prompt that buries the request. Length is not the signal. Specificity is.
Can I use these patterns for image and video tools?+
Mostly. Role and Constraints translate directly. Structure becomes format and aspect ratio. Examples become reference images. Iteration is essential because image and video tools are less consistent than text models. Expect three to five iterations before the output lands.
What is the most common mistake teams make?+
Skipping Role. People jump straight to the task without telling the model who it is writing as or who it is writing to. Adding a single Role sentence improves output more reliably than any other change.
Where do I save prompts that work?+
In a place your team can find them. A shared Notion page, a Confluence space, or a tool like PromptLayer. The patterns are the framework. A team prompt library is the artefact that lets the framework compound. We help teams build one on the day in the AI Foundations Workshop.
