Have you ever typed a prompt into ChatGPT or into your favorite AI tool and get back something generic. That's sometimes not quite up to your standards or not quite for your business or industry? So you try again, maybe tweak a word or two to get something much better and then a few minutes later you get something usable but it took you many attempts to get it right.
Well, the thing is I did. Many times.
Through this, I realized that prompting is a communication skill, not a technical skill. The people who actually get remarkable results are not engineers but people who know how to give good direction and have the ability to communicate that.
The CARE framework (Context, Ask, Rules, Examples) made by a few researchers and writers at the Nielsen Norman group is simply a structure for giving that direction consistently. Once you understand it, you'll stop blaming the tool and start getting outputs you can actually use.
Asking it just doesn't work.
Have you heard of garbage in and garbage out?
It's just like that.
Garbage inputs produce garbage output.
Not because AI is unintelligent, but because it's trying to fill in what you didn't say with the most statistically average answer it can find. And that answer is almost never what you wanted.
When you prompt the model "write me a marketing email", the model has to guess a few things:
Who's the audience?
What's the tone?
What's the offer?
What's the desired action?
It guesses. And its guess is basically something that looks like an average marketing email on the internet.
What is Context?
Context tells the AI who it is, who it's talking to, and what situation it's operating in. Without proper context, the model writes output based on whatever generic data it finds.
Context can include the role you want AI to play "You are a senior web developer good in architecture and writing scalable and maintainable code".
It can also include the audience ("They are non-technical founders who have no clue how software is maintained"), and the situation ("This is for an authentication feature").
Each of these constraints dramatically narrows the output in a good way.
What is the Ask?
The ask is your deliverable, but specificity is what really separates a useful output from a generic one. Most prompts are too broad or too short at this stage.
A weak ask example would be: "Write a blog post about AI."
A strong ask: "Write a 600-word blog post introduction that explains why founders without technical backgrounds struggle to evaluate AI vendors, and why that matters for their product roadmap."
The second version tells the model the length, the format, the audience's problem, and the purpose, and it leaves almost no room for the model to guess wrong.
What is a Rule?
Rules are actually the most underused part of prompting. They prevent the model's default behaviors from getting into your output.
Stuff like corpo words, excessive length, etc.
A rule I usually use day to day is to tell AI not to give me output without a source, so I can avoid getting answers with no basis.
What is an Example?
Examples are the highest-leverage element of a prompt, and almost no one I know outside of work uses them. When you show the model a sample of the output you want or don't want. You cut through ambiguity faster.
You can provide a writing sample in your voice ("Here's a paragraph from a previous email — match this tone").
A structural example ("Format it like this: headline, one-sentence summary, three bullet points"), or a negative example ("Here's something I don't want — do not return an output that follows this structure").
Even one short example can shift the output quality dramatically.
This is actually called "few-shot prompting" in technical circles but the concept is just: show, don't only tell.
As I've said, getting great AI output is not about being technical. It's about being clear on what you want. The CARE framework is just a way to be clear on purpose.
You already know your business, your audience, and what good work looks like. The framework is just a container for that knowledge so that AI can use it.
So, maybe next time you sit down with an AI tool, and before you type anything.
Spend 60 seconds writing out one sentence for each letter of CARE.
You'll be surprised how much the output shifts.
This was originally sent as a newsletter issue. If you want one backend lesson per week, subscribe here.