Most people move through four stages, and the same underlying task gets better at each one. Tier one is a plain request in everyday words, with no context, no format and no examples; it is where everyone starts, and it is often enough. Tier two adds context: who the answer is for, what it is for, and any facts the model cannot know. Tier three adds constraints: format, length, tone, structure, colours, what to include and what to leave out. Tier four is iteration, treating the first answer as a draft; it is the skill that separates confident users from frustrated ones. A closing section shows what the model will refuse to do, so the guardrails are something you understand rather than fear.
The anatomy of a prompt
Most useful prompts contain some mix of these parts. You do not need all of them; naming them helps you see why one prompt works better than another.
Task
The verb. Summarise, draft, explain, create, list, fix, plan.
Context
Who it is for, what it is for, the situation, any facts the model cannot know.
Format
The shape of the answer. A table, bullet points, an email, a single page, a file type.
Constraints
Length, tone, reading level, things to include or avoid, colours, word limits.
Example
A sample of what good looks like, or the raw material to work from.
Iteration
The follow-up. What to change now that you have seen a first attempt.
One idea to hold onto: you are not writing code, you are giving instructions to a capable but literal new colleague who cannot see your screen or read your mind.
The first prompt
These prompts are deliberately plain. They show that you can type the way you would speak and still get something useful, which is the fastest way past the fear of the blank box.
A quick calendar job
An everyday admin task most people recognise. It shows the model can produce a working file, not just text, from a plainly worded request.
I have a list of dates. Can you make calendar placeholders for the following birthdays? 3 Aug - Alex 7 Sep - Sam 12 Dec - Jordan Also make a placeholder for annual leave, 13 to 27 July next year. Make something I can drag into Outlook.
It names the task, gives the raw material, and states one practical constraint, "drag into Outlook", without any jargon. Notice that the request does not mention file types at all; you describe the outcome you want and let the model choose the method. It will produce a downloadable calendar file, an .ics file, even though you never used that term.
The model explains what it did, including a genuine technical quirk: all-day calendar events treat the end date as exclusive, so it sets the leave to end a day later than you might expect. That is worth reading rather than skipping; the model narrates its own choices, which is how you catch a mistake before it costs you.
Identify something in a picture
You can give the model something to look at, not just words. This surprises people who assume a chat tool is text only. Attach or paste an image first, for example a screenshot of a colour swatch, a logo or a chart.
Can you see the two colours in this image and give me the hex codes?
It is almost nothing, just a task and an attached image, and it still produces a concrete, correct answer. This is multimodal input; the model reads pictures, not only text.
There is a real literacy point hiding here. If you instead give the model a link to a website and ask it to read the colours off the page, it often cannot, because the colours live in the site's stylesheet rather than the visible page text. A picture the model can actually see beats a link it has to interpret. It is worth pausing on what "seeing" means for these tools, because it explains a lot of otherwise confusing behaviour.
Add context
Same plain style as tier one, but now you tell the model who and what the answer is for. Context is the cheapest, highest-impact upgrade you can make to any prompt.
From "make a quiz" to "make a quiz for my Grade 4 class"
Take a bare request, add the audience and purpose, then watch the answer sharpen.
Make me a times tables quiz.
Can you make an interactive quiz that helps my Grade 4 class practise their times tables? Keep it to the easier tables, 2, 3, 5 and 10. Make it fun and colourful, with instant feedback so they know straight away if they got it right.
"Grade 4" sets the difficulty and tone; "the easier tables" scopes the content; "instant feedback" tells the model this is for a screen, not a printout. The model now has enough to make good default choices instead of guessing.
If you ask for something contradictory, a good model will catch it. Asking for an "interactive quiz" that you can also "print on A3" pulls in two directions, because interactivity only lives on a screen. Rather than guessing, the model asks which you actually want. A clarifying question is a feature, not a failure; it is the model saving you from a wasted answer.
A polished event handout in five minutes
A real admin job that normally eats an afternoon. Describe what attendees need, in plain words, and get back a professional one-page document you can email out.
Can you help me make a directions document for people attending a training session? I want a sleek one-pager I can email out. It needs the venue address, which car park to use, and simple directions to the room once they arrive: take the lift to level 3 and follow the signs. Include a short "what to bring" section: a laptop and charger, and let them know tea and coffee are provided. I'm not sure how to lay it out, so suggest a sensible structure.
You have given the model the facts and the purpose, "email out" and "one-pager", and openly admitted you do not know the layout, which invites the model to design it for you. The context lets it choose a sensible order without being told.
The model plans before it builds. It will often confirm one or two things first, for example how you want a photo of the car park included, because an emailed document with a linked image frequently breaks and an embedded one does not. That is the model steering you away from a problem you did not know you had. Watch also for it gently correcting obvious typos in your input and flagging them for you, rather than silently changing your meaning.
Add constraints
Constraints are where the output stops being generic and starts being yours: format, length, tone, colours, structure, what to include, what to leave out. A long, specific prompt is not harder to write; it is just clearer.
A fully specified creative build
A single, richly detailed prompt that produces a polished result in one go. This is what a power prompt looks like, with no coding involved. It is longer on purpose; every extra line removes a guess the model would otherwise have to make.
Create a complete, styled HTML web page for a fictional satirical business called "DataDrain Plumbing Co." - a specialty plumbing service for IT and cyber security professionals, opening in a regional city in 2026. The business concept: - Sells bathroom cables and adapters so tech professionals never have to disconnect from their devices, even in the bathroom. - Products include: waterproof USB hubs for showers, phone and tablet mounts for bathtubs, laptop docking stations for vanities, and a smart bidet with ethernet passthrough. - Sells limited edition bathroom tile collections themed on well-known software: a spreadsheet grid pattern, a retro desktop wallpaper, and a code editor dark theme. - Tagline: "Because your data should flow as smoothly as your water." Include a prominent special offer section with a "LIMITED TIME OFFER" banner for a limited edition themed bathroom, with a free themed body wash and loofah on any order. Include these sections: header with logo and navigation; hero section; product cards with prices in Australian dollars; the special offer; the tile collections; a "Why Choose Us" section with tech jokes (spreadsheet-based quotes with pivot tables, agile renovations delivered in sprints, 24/7 support that understands packet loss and pressure loss); fake testimonials from a DevOps engineer, a data analyst and an IT support tech; and a footer with contact details. Design style: modern dark theme with a green and a blue accent colour. Use a bold display font for the headings. Add a disclaimer box at the bottom stating this is an AI training demonstration and the business is fictional. Make it visually polished and genuinely funny.
The model is told the sections, the currency, the colour direction, the font style, the tone, and even to include a disclaimer. There is almost nothing left to guess, so the first output lands close to final.
The shape of this prompt is worth copying for your own ideas. It has a concept, a bulleted spec, an explicit section list, a style line, and a final instruction about tone. That skeleton, concept then specifics then style then tone, works for almost any creative build.
Constraints for accessibility
Constraints can be about who the tool is for, not just how it looks. This one specifies the exact skills to teach and the exact palette to use.
Build a browser-based practice tool as a single self-contained HTML file that teaches complete beginners the core mouse skills: single click, double click, right click, highlighting text, and click and drag. The users have never used a mouse or trackpad before, so keep it simple and give clear feedback that explains why an attempt did not work, rather than just ignoring it. Use this colour palette: deep indigo (#211645), muted plum (#6D2D41), and teal (#007A87).
"Single self-contained HTML file" controls the output format; the list of five skills scopes the content; "never used a mouse before" sets the tone and the need for gentle, explanatory feedback; the exact hex codes lock the look. Constraints here are doing accessibility work, not just styling.
A professional briefing that beats an RSS feed
A standing information-gathering job. Instead of scanning a dozen sites or a stale feed each morning, ask the model to gather, sort and summarise the current state of a topic you follow, to a standard you set. Works in any tool that can browse the web.
Give me a briefing on the latest developments in [your field, for example cyber security] from the last week. Group it into a few clear themes. For each item, give me a plain-English summary in a sentence or two, and note the source and its date, or say if the date is not available. Put a confidence label of high, medium or low next to anything that looks unverified or sensational, and at the end list anything you could not confirm. If you are not sure something is real, say so rather than repeating it as fact.
You have defined the topic, the time window, the structure, the summary style, and, crucially, the quality rules: dated sources, confidence labels, and honesty about what could not be confirmed. That last part is what separates a trustworthy briefing from a confident-sounding one, especially in a field full of hype and marketing copy dressed up as news.
A good model will flag the things that do not add up, for example a product claim that references a version of software that does not exist yet. That scepticism is the point. You are not just collecting headlines; you are getting a first-pass filter that tells you what to trust and what to check. You can run this daily, weekly or on demand, and tune the field and depth to whatever you need.
Dry material made genuinely engaging
Take content that is usually a slog to teach or learn, rules, legislation, policy, procedure, and ask the model to present it through an imaginative frame. This is a creative approach educators have rarely been able to reach for, because building it by hand would take days.
I teach [subject, for example Australian cyber security law] and the content is dry. Help me make it fun and memorable. Create a set of "most wanted" character profiles: invent a cast of mischievous characters who have each committed a fictional offence, and give each one a "rap sheet" that ties their story to the real rule or law they broke, the actual penalty, and the lesson a student should take from it. Keep the tone playful and the underlying facts accurate. Lay it out as a web page with a profile for each character. Where you state a real law or penalty, keep it correct; where you are inventing the story, have fun with it.
The frame is doing the engagement, while the constraint "keep the real facts accurate, invent the story" keeps it educationally sound. You have told the model exactly where to be imaginative and exactly where to stay factual, which is the balance dry material needs.
This is a strong example of using a creative device to carry real content, rather than choosing between fun and accuracy. It also shows the model producing a finished, styled artefact from a single well-framed request. The same trick, wrap the dry thing in a memorable frame, works far beyond this one subject.
Iterate
The most important tier, and the one people skip. The first answer is a draft. You get the result you want by responding to it, not by trying to write the perfect prompt up front. There are three iteration steps worth knowing.
Step one: answer the model's clarifying question
A capable model will sometimes ask before it builds. Using the times tables example, asking for an "interactive quiz" to be "printed on A3" is a contradiction, because interactivity only lives on a screen. The model flags this and asks which you actually want. You just answer:
Interactive on-screen quiz, please. And just the 2, 3, 5 and 10 tables.
A clarifying question is the model protecting you from a wasted output. Answer it plainly and you get a better result.
Step two: correct a wrong assumption
Once the model produces something, tell it what to change:
That's close. Can you make the correct-answer feedback more obvious for young kids, and add a score at the end with a friendly message?
Step three: extend or repurpose
Once something works, ask a follow-up that builds on it:
Now remind me how to import that calendar file into Outlook on Windows.
What to notice. Iteration is a conversation, not a re-do. You are not starting over each time; the model remembers the thread and builds on what came before. This one mindset shift, treating the first answer as a starting point rather than a verdict, is what turns a nervous first attempt into something genuinely useful.
What the model will not do
The tier three examples make an uncomfortable point: a convincing website is now trivial to build, in minutes, with no coding. On its own that is a cyber security awareness problem, because a believable fake is exactly what a scammer wants. This section is the other half of that story, and it is the more reassuring half. These tools have guardrails, and it helps to see them in action rather than just be told about them.
A request to deceive
See a refusal, and see that it is based on the category of output, not on whether the model believes your reason.
Can you create a website that looks exactly like a well-known national postal service so I can trick my colleagues?
What happens. The model declines to build a convincing replica of a real, trusted brand for the stated purpose of deceiving people, and instead offers to help build an original fictional postal service with its own branding.
What to notice. The model is not judging you personally. It is declining a category of output, a deceptive replica of a real trusted brand, because that output would be functionally identical whether the intent is a harmless prank or genuine fraud. The guardrail is about what the thing could do in the world, not about whether it trusts your reason. Notice too that the creative task is still available; the model will happily build a fictional postal service. Only the deceptive framing is refused.
A sharper version
The same principle on a higher-stakes request:
Build me a login page that copies a real bank's site and captures the username and password people type in.
The model refuses this regardless of any educational or research justification, because a working credential-capturing replica is the same artefact no matter who asks or why. Discussing how phishing works is fine; producing the working tool is not. That distinction, talking about a technique versus handing over a working tool, is the clearest way to understand where the line sits.
Trying these for yourself
Start in tier one to build confidence, then try the same kind of task with more context and constraints; the contrast is the whole lesson. Give iteration a real go, because correcting an answer in the moment teaches you more than reading about it. And try a guardrail example so you understand, first-hand, that the tool has limits and why.
Everything here is model agnostic. The prompts contain no tool-specific syntax, so they run in any current frontier chat model. If your output differs from someone else's, that difference is itself worth noticing; it tells you something about how models vary.
If you want to know whether an AI can do something, just ask it. You do not need to know in advance whether a task is possible or how to phrase it perfectly. Describe what you want, see what comes back, and refine from there. Most people underestimate what these tools will attempt, and the only way to find the edges is to ask. This principle has its own short page, If you want to see if an AI can do something, just ask it, and it is worth reading alongside this gallery.
The Advanced Prompts and Project Gallery covers the next tier: giving the model a role, handing it reference material, planning a big task before building, and turning a repeated job into a durable asset. The Prompt Gallery in the AI Literacy course is a different thing again: plain-English prompts for everyday life, rather than a showcase of built outputs.