Six Prompt Patterns That Actually Work

👁 115 reads

The difference between a prompt that wastes your time and one that delivers something usable on the first try is almost never the model. It is almost always the prompt. Once you internalize a handful of structural patterns, writing a good prompt takes the same thirty seconds as writing a bad one - you just know what to include.

Why most prompts fail before the model even reads them

When you type a vague instruction, the model does not guess what you wanted - it fills the gaps with statistically likely content. That is usually generic content, because generic is, by definition, the most common. The fix is not to ask more politely or use magic words. The fix is to reduce ambiguity: give the model a role, a context, a clearly scoped task, a format, and optionally an example of what good looks like.

Think of it as a checklist. Not every prompt needs every element, but knowing the elements means you can diagnose exactly what is missing when output disappoints you.

Pattern 1 - Assign a role

Telling the model who to be sets a tone and a knowledge frame for everything that follows.

Weak: Write a bio for my website.

Strong: You are a copywriter who specialises in personal brands for independent consultants. Write a 90-word third-person bio for my website. I am a UX researcher with eight years of experience, mostly in fintech, and I work with Series A and B startups.

Why it wins: The role primes the model to use appropriate register and vocabulary. The context - fintech, startup stage, research specialty - rules out irrelevant angles. The word count creates a concrete deliverable instead of an essay.

Pattern 2 - Load the context, not the question

The most common prompting mistake is leading with the ask before the model has any background. Front-load context first, then make the request.

Weak: Summarise this email thread for my boss.

Strong: Below is an email thread between a client, a project manager, and a developer discussing a delayed feature launch. My boss needs a three-bullet summary: what caused the delay, what the agreed next step is, and who owns it. Write in plain, direct language - no jargon. [paste thread]

Why it wins: The model now knows the audience (boss), the purpose (decision-making, not archiving), the structure (three bullets, specific topics), and the register (plain). Without that frame, a summary might be chronological, verbose, and developer-focused - technically accurate but useless to a non-technical executive.

Pattern 3 - Scope the task with constraints

Constraints are not limitations - they are instructions. Word count, audience reading level, what to exclude, what tone to avoid: these are all constraint information the model cannot infer.

Weak: Explain machine learning to me.

Strong: Explain the concept of machine learning to a small-business owner who runs an online clothing store and has no technical background. Use one analogy drawn from retail. Keep the explanation under 150 words. Do not use the words ‘algorithm’ or ‘data science’.

Why it wins: Every constraint here eliminates a failure mode. Without the audience specification, the explanation might be too abstract. Without the word cap, it might run to 600 words. The banned-words rule forces simpler vocabulary. Constraints are how you pre-empt the most predictable problems.

Pattern 4 - Give an example of the output you want

If you already have an example of good output - even an imperfect one - paste it in. This is sometimes called few-shot prompting, and it works because showing is more precise than describing.

Weak: Write a subject line for my newsletter.

Strong: Write five subject lines for a newsletter about productivity for remote workers. Here are three subject lines from previous issues that performed well: ‘The two-minute rule you are probably misusing’, ‘Your calendar is lying to you’, ‘Why async is not always the answer’. Match that voice - direct, slightly contrarian, no exclamation marks, no questions.

Why it wins: Three examples communicate voice, pattern, and tone faster and more reliably than a paragraph of adjectives ever could. ‘Slightly contrarian’ means something different to everyone; the examples make it concrete and specific to your brand.

Pattern 5 - Specify the format explicitly

Format and content are separate decisions, and the model will make a default format choice unless you tell it otherwise. That default is rarely wrong, but it is rarely optimal either.

Weak: Give me ideas for my product launch.

Strong: I am launching a browser extension that blocks distracting websites on a schedule. Give me ten launch ideas targeting indie developers. Present them as a table with two columns: the tactic name, and a one-sentence description of the action I need to take. Focus on low-cost, high-effort tactics, not paid advertising.

Why it wins: The table format makes the output immediately scannable and actionable. The one-sentence constraint stops each idea from turning into a mini-essay. The low-cost qualifier rules out suggestions that require a marketing budget - which is genuinely useful filtering, not just aesthetic preference.

Pattern 6 - Treat the first draft as a starting point, not a result

Most people accept the first response or abandon the task if it disappoints them. Iteration is where the real value is.

After a first draft, try targeted revision prompts:

• The second paragraph is too formal. Rewrite it in a more conversational tone.
• Cut this by 30% without losing the three main points.
• The opening sentence is weak. Give me five alternative opening sentences with more specificity.
• Assume the reader is sceptical of this claim. Add one sentence that pre-empts that objection.

Targeted revision prompts work far better than make it better, which gives the model no useful direction. Point to the specific problem, describe what better looks like, and keep the rest of the draft stable.

Iteration is not a sign of failure. A well-structured first prompt usually gets you 70-80% of the way there; two or three targeted follow-ups close the gap. That is still dramatically faster than editing a blank document from scratch.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Put the framework to work today

The next time a prompt produces a mediocre result, run through the checklist before re-prompting: is there a role? Is the context front-loaded? Are the constraints specific enough? Is there an example? Is the format defined? Nine times out of ten, one of those five elements is missing or vague, and adding it is a ten-second fix. Bookmark the six patterns - OpenAI’s prompt engineering guide and Anthropic’s equivalent both expand on this further - but start using them in the prompts you are writing right now, not after you have read everything there is to read about prompting.

Kinetyra