Better AI prompts come from seven repeatable patterns: give the model a role, add context, show an example, specify the output format, split complex jobs into steps, state your constraints, and iterate on the answer instead of starting over. None of this requires special keywords or prompt-engineering jargon — each pattern just removes one more thing the model has to guess. Applied together, they turn vague requests into drafts that need small fixes rather than rewrites. Here is how each pattern works, with a weak version and a better one you can test on any chatbot in a couple of minutes.
A chatbot has no idea who you are, what you already know, or what you plan to do with the answer. When your prompt leaves those things open, the model fills the gaps with its most generic output: a neutral summary, a polite list of options, or prose that sounds like a brochure. The seven patterns exist to close those gaps. If your task is simple, two of them are enough; if the answer comes back wrong, the missing piece is usually one of the others. Before you change tools, change the prompt — which assistant you use matters far less than how you ask, as the ChatGPT vs Claude vs Gemini comparison shows.
Weak: "Write an email about a late delivery." The model guesses the sender, the audience, and the tone, and usually lands on something stiff. Better: "You are a customer-service writer. Draft a 150-word email to a customer whose order shipped three days late. Apologize once, explain the shipping-fee refund, and end with their tracking link." The role sets the register, and the rest of the sentence sets the audience and the content. The same request phrased for a support lead, a founder, or a lawyer produces correctly different drafts.
Weak: "Suggest topics for our newsletter." Better: "We run a monthly newsletter for owners of small coffee shops — about 40% opened last time, mostly the equipment and maintenance items. Suggest six topics that are not in our last three issues (roasters, milk, staffing)." Context tells the model who the readers are, what has already been covered, and what kind of suggestion is useful. Without it you get generic listicle ideas; with it you get topics worth a planning meeting.
Weak: "Rewrite this paragraph more clearly." Clarity is in the eye of the reader, so the model improvises. Better: attach the paragraph plus a before-and-after pair: "Here is the tone I mean — before: 'We are pleased to inform you that...'; after: 'Your order shipped on Tuesday.' Rewrite my paragraph the same way." One concrete example pins down style, length, and structure faster than any number of adjectives such as "friendly" or "concise," which mean different things to different models.
Weak: "Summarize this article." You may get a paragraph, a list, or something in between. Better: "Summarize this article as five bullets, each with the main claim in the first sentence and one supporting detail after it. Under 200 words total." Telling the model the shape of the answer — bullets, table, JSON, steps, word count — makes the output usable the moment it arrives, which is the whole point when the summary is going into a report, a ticket, or a script.
Weak: "Write a complete onboarding plan for our new hires." A single giant request forces the model to divide its attention and usually returns a shallow outline stretched to look complete. Better: ask for the outline first, review it, then say "now expand section two." Each step gets the model's full attention, and you stay in control of direction between steps. The pattern is the same one behind the structured AI productivity workflows we recommend teams adopt: input, output, and a human check at each stage.
Weak: "Write a product description for our blender." Better: "Write a 60-word product description for our blender. No superlatives such as 'best' or 'revolutionary,' mention the 700-watt motor and the two-year warranty, and do not make health claims." Constraints do double duty: they shape the output and they act as guardrails, which matters for anything that will appear on your website or in customer-facing copy. A description drafted with explicit limits needs less editing and carries less legal risk.
The strongest prompt is often a follow-up. When the first answer is close but not right, say what is wrong in one sentence: "Shorter — under 80 words," "More casual, like a text to a colleague," or "Your third point is wrong; drop it and expand the second." Models handle refinement well because the conversation keeps the context. If the answer is far off, ask for two or three alternative versions and choose — that usually beats rewriting the original prompt from scratch.
In real use the patterns stack. A single prompt that works might read: "You are a technical writer (role). Our API users are developers who have read the quickstart but not the full docs (context). Here is the tone from our changelog (example). Give me five troubleshooting steps, numbered, each under 40 words (format), with no implementation details (constraint)." Notice what is absent: no incantations, no "act as an expert" theatrics, no demands to be "high quality." Every clause carries information the model would otherwise have to invent.
A quick way to improve your own prompts is to read them as a stranger: underline every noun and instruction that could mean two things, then replace the vaguest one with a concrete detail or an example. One pass of that edit usually moves an answer from generic to usable.
Assume the first draft is a draft. If a pattern clearly failed — no example was given, the format was unspecified — add the missing piece and resend. If the model is confidently wrong about a fact, stop prompting and check the source yourself, because no amount of phrasing fixes a model's uncertainty about specific numbers or events. And if a task keeps coming back wrong despite clean prompts, the fix is often a better tool or a tighter document for the model to ground itself in — the same escalation path covered in our guide to AI automation for small business.
| Pattern | When it helps most | Quick test |
|---|---|---|
| Role | Writing, customer service, tone-sensitive drafts | Add "You are a..." and compare tone |
| Context | Brainstorming, planning, anything about your situation | Add one sentence about the reader |
| Example | Rewrites, style matching, structured content | Include a before-and-after pair |
| Format | Anything you will reuse or paste elsewhere | Ask for bullets or a word count |
| Steps | Large documents and multi-part plans | Ask for the outline first |
| Constraints | Customer-facing copy, compliance-sensitive text | List what the output must not contain |
| Iteration | Every close-but-not-right answer | Point at the flaw in one sentence |
A good prompt gives the model enough that it does not have to guess: a clear goal, relevant context, an example of the output you want, and any constraints such as length or tone. You can judge a prompt only by its result — if the answer needs one small fix, the prompt was good; if you rewrite the whole thing, tighten the prompt.
Long enough to remove ambiguity, which is usually two to five sentences for everyday tasks. For complex jobs, add context and an example instead of one giant paragraph. When in doubt, start shorter and let a follow-up question carry the detail — most assistants handle a short prompt plus one clarification better than a wall of text.
No. Chat assistants are trained to follow plain instructions, so normal language with the seven patterns beats jargon or magic phrases. There is no keyword that forces better output; what changes results is structure, context, and examples.
Models imitate patterns, and one concrete example pins down style, structure, and level of detail faster than any description. A short before-and-after pair you supply tells the model exactly where the bar is, so its output lands closer to your target on the first try.
Treat it as an iteration, not a failure. Say what is wrong in one sentence — too long, too formal, the second point is wrong — or add the example you should have given first. You can also ask for two or three alternative versions and pick the best one.
References — official vendor prompt guidance (checked September 2026)
Last updated: 2026-09-07. Model behavior and documentation change frequently; verify current details on the official pages above before building production workflows around them.