How I Actually Write Prompts (The Framework I Use Every Day)

Where This Comes From

I've been using AI to write things for close to two years now. My very first prompt was basically "help me write something about..." -- and I honestly expected magic. What I got back was vague, generic, and barely usable. Sound familiar?

The journey between that and where I am now -- where I can get something usable on the first try more often than not -- involved a lot of frustration and wasted time. I'd rewrite the same response five times, trying to force the AI into giving me what I actually needed. Eventually I stopped blaming the AI and started looking at what I was feeding it.

Here's the thing I landed on: writing a prompt is basically doing project management for AI. Your prompt is the job brief. The vaguer the brief, the more the AI "creatively interprets" -- and the more random your results. Everyone knows this in theory. But sitting down and actually doing it? Most people, myself included for a long time, just don't have a system.

I want to share the framework I settled on. Nothing here is borrowed from a textbook. Every piece of it came from trial and error.

Before the Framework, a Mindset Thing

A lot of people hear "prompt framework" and think, "I'm not a prompt engineer. Why do I need this?"

I used to think the same thing. But here's what I've realized: whether you're asking AI to draft a weekly report, put together a project proposal, or debug code, if you want it right the first time, you need to be clear about what you want. That has nothing to do with your job title. It has everything to do with how much time you're willing to waste on bad outputs.

The framework isn't supposed to turn you into someone who writes mini-essays to the AI. Think of it more as a checklist -- something you run through before you hit enter. Once it becomes habit, it takes a few seconds in your head. You won't even realize you're doing it.

The Six Questions I Answer Every Single Time

I boiled everything I've learned down to six questions. Every time I write a prompt, even a quick one, these are the things I think through.

1. What identity should the AI adopt?

Giving the AI a role isn't a gimmick. It directly controls what knowledge depth it pulls from and how it communicates.

Try this experiment. Ask an AI "explain quantum mechanics" and then ask it "explain quantum mechanics to a high schooler who loves jokes." Completely different answers. The identity changes which parts of its training it draws on.

The catch is that the role has to be specific. "You are an expert" tells the AI absolutely nothing. "You are a supply chain manager who's spent over a decade running community group-buying operations and talks to suppliers every single day" -- now you've given it something to work with.

What I actually do: I pick the role based on the task. Writing a marketing brief? "Senior copy headline specialist." Crunching sales numbers? "Data analyst who's helped retail stores optimize inventory for years." What genuinely surprised me is that the role itself changes how the AI frames the answer -- it thinks from that person's perspective and often surfaces angles I hadn't considered at all.

Pitfall I hit early on: I used to think role-setting was overkill for simple questions. It's not. Even for something like "help me write an email," adding "you're a direct communicator who respects people's time" instead of leaving it open changes the whole output from wordy to concise.

2. What exactly do I want the AI to do?

This sounds painfully obvious. But most bad prompts fail right here.

"Help me write a plan" -- what kind of plan? For whom? How detailed? Will you present it to a boss, or is it just for your own reference?

The pattern that works for me: verb + deliverable + goal.

For example: "Help me write a Dragon Boat Festival in-store event plan (deliverable) with the goal of increasing foot traffic by over 30% (goal)."

Leave out the goal and the AI gives you a stock template. Leave out the deliverable and you get a list of suggestions instead of an actual plan.

A tip that saved me a lot of time: Be specific about the verb. "Write" is different from "outline," which is different from "critique each section of." The verb tells the AI what kind of work to do. If you're vague about the verb, you'll get whatever the AI thinks is most likely -- which may not be what you had in mind.

3. What background does the AI not have?

This is the most commonly skipped step, and honestly, it might be the most important one.

You have all this context swimming around in your head. Your business situation, your audience, what resources you actually have, what's worked before and what hasn't. The AI knows none of this. Think about it like asking a colleague for help. If you say nothing, they guess. But if you tell them "my storefront is 80 square meters, five employees, most of my customers are moms from the neighborhood," their advice becomes dramatically more practical instead of generic.

Same with AI. The more relevant context you give -- your role, your audience, your constraints -- the less generic the response. It shifts from "here's what anyone could do" to "here's what makes sense for your exact situation."

The dumbest mistake I ever made: I once asked AI to help me write a community operations plan without giving any background whatsoever. What came back was a whole list of influencer collaborations and KOL campaigns as recommendations. I run a small neighborhood grocery shop. What budget do I have for influencers? If I'd just said upfront "I own a community fresh grocery store, about 80 sqm, daily revenue around 8000," every single thing downstream would have been completely different.

I still catch myself sometimes starting a prompt before I've fully given context. I've gotten better, but it's a habit worth fighting for.

4. What are the hard requirements?

State what is absolutely non-negotiable. Budget caps, time limits, resources available, things that must be included -- spell these out.

I like using the word "must" because the AI treats it as a hard constraint rather than a soft suggestion.

For example:

• Total budget must stay under 5000
• The activity period cannot exceed 7 days
• Only 3 employees to execute, so the plan can't be complex

The more specific your requirements, the less likely you are to get a fantasy plan full of good ideas that nobody could ever execute.

Where I see people trip up: They think hard requirements are only for big, formal projects. But even a simple "response must be under 200 words" or "must use real product names, not placeholders" saves you from having to re-prompt.

5. What do I NOT want?

This step is the one people skip most often, and I think it's the one that separates okay prompts from good ones.

AI has a strong tendency to give you the "standard answer." Ask it how to write a WeChat article and it gives you the safe, textbook structure. Ask how to run a campaign and it gives you the playbook everyone and their dog has seen a hundred times.

If you don't tell the AI what you don't want, it will give you exactly that.

I keep a "don't" list alongside my prompt:

• No big tech company case studies (our team has five people -- learning from Google doesn't help us)
• No internet slang or buzzwords (our audience doesn't connect with that)
• No theory or principles (I need things I can act on this afternoon)

Declaring what you don't want cleans the space so the AI can fill it with what you actually need.

This one surprised me the first time: I was getting tired of every AI response opening with "In today's rapidly evolving digital landscape..." Adding "don't use any opening clichés or filler paragraphs" to my prompt felt almost too simple. But it completely eliminated that problem. Sometimes the most impactful line in your prompt is a negative constraint.

6. What should the output format be?

If you want the AI to produce something you can actually use -- instead of a draft that needs heavy rewriting -- you need to specify the format.

You don't need a rigid template with every field defined. Just indicate: how many sections, whether to use tables, how to handle heading levels. A little effort here saves you two or three rounds of follow-up prompts.

What I learned the hard way: I once asked for a "marketing plan" without specifying format. What I got was one giant block of text. It had good ideas buried in it, but I spent more time reformatting it than I would have spent writing it from scratch from the start. Now I always say something like "use numbered sections with bold headings" or "put the budget breakdown in a table." The structure might seem like a minor detail -- it isn't.

A Quick Note on Style

People often forget to mention style when writing prompts. Then the result comes back too formal, or too casual, or with that unmistakable "AI tone."

I got into the habit of tacking a line or two onto the end of my prompts: "like chatting with a friend," "no jargon," "short paragraphs with plenty of line breaks so it doesn't feel dense." A few words, huge impact on the final result.

I also give AI reference samples a lot now. If I have an example of something written the way I like, I'll paste in a snippet or describe the structure and tell the AI to follow it. This is especially useful when you need a specific tone -- rather than describing what you want in the abstract, just show it an example. The AI picks up structure and voice from a real sample far better than from a verbal description.

My biggest style pitfall: Not specifying this at all is the pitfall. If I don't say anything about tone, the AI defaults to something that sounds like a Wikipedia article crossed with a corporate memo. Fine for some uses, terrible for most of the things I actually need.

A Full Example: Writing a Blog Post Using All Six

Here's what a finished prompt looks like when I put everything together:

You're a senior editor at a tech publication with over eight years
of content experience. You specialize in making complex technology
concepts accessible to regular people, especially younger readers.

Task: help me write a blog post about AI prompt engineering. The
goal is to make readers who use AI daily but always feel it's
"not that great" realize the problem isn't the AI -- it's how
they talk to it.

Audience: mostly working professionals aged 20-35 who use AI
for work every day but basically just say whatever comes to mind
without systematically thinking about prompts.

Hard requirements:
- Around 1500 words -- not too long
- Have a clear point of view -- don't hedge
- Must include actionable tips people can try right now

Don't include:
- Theory and background explanations
- Inspirational fluff
- Big company case studies

Style: like chatting with a friend, throw in the occasional
rhetorical question, keep paragraphs short so it reads easily.

This is more natural than a rigid "Module 1 / Module 2" format, and it works just as well -- maybe better, because the AI can follow a narrative flow instead of filling in boxes.

Things I've Picked Up Along the Way

You don't need the full framework every time. Quick things -- looking up a definition, translating a sentence, checking grammar -- a couple sentences are fine. The six-question version is for those moments when you want something you can use directly or with minimal edits. Not everything needs that treatment. Learning when to use the full framework versus when to keep it light is a skill in itself.

The order isn't sacred. Start with the role or start with the task -- genuinely doesn't matter. I tend to start with the role because once that's set, the task and requirements flow naturally from it. But I've seen people start with constraints or the "don't" list. Whatever gets your thoughts organized is the right order.

Don't chase the perfect prompt. I've seen people -- and I've been that person -- spend half an hour sculpting a prompt word by word. In practice, it's faster to spend two minutes on a decent first prompt, see where the AI goes wrong, and adjust. Iterating three times is almost always quicker and produces better results than agonizing over one "perfect" draft. The AI is patient. Use that.

The habit matters more than memorizing the framework. If before every important AI interaction you spend two minutes thinking: Did I give a role? Is the background enough? Are constraints clear? Did I say what I don't want? -- once that becomes automatic, your prompt quality goes up dramatically. You're not going to think about "the six questions" forever. Eventually it just becomes how you think about asking for help.

There's nothing mysterious about any of this. It's about making the things that feel obvious to you explicit to the AI. The things you consider self-evident? Not a single one of them is obvious to the AI. But once you spell them out, the difference in output is like night and day.