Most bad AI output is not the model’s fault. It is a prompting failure. When ChatGPT keeps missing the mark, the problem is rarely the machine. It is the instructions you handed it. Learning how to prompt is the highest-leverage skill for getting real work out of any AI tool, and it is far more learnable than people assume.
That is why we built the C.R.I.S.P.Y. framework at Flux+Form. It turns a vague request into a precise instruction that points the model exactly where you want it to go. Six parts: Context, Research/Role, Instructions, Specifics, Parameters, and Yielding. We think of it as the prompt of all prompts.
Whether you are writing your first prompt or your thousandth, this guide breaks down how to prompt with C.R.I.S.P.Y., one piece at a time, with a real business example you can copy and use today.
How to Prompt Starts With What the AI Is Actually Doing
The model does not create. It patterns. It reads your words and predicts the most probable next words based on everything it has ever seen. It does not know things the way you know them. It makes a fast, confident guess toward the strongest probability and hands it to you.
That has a direct consequence for how to prompt. Your words decide which slice of everything it knows the model reaches into. Aim it at the right slice and the most probable answer is also the correct one. Leave the aim vague and it fills the blank with whatever is statistically average, then delivers it with total confidence. We call that slice the universe of data, and pointing at it is your job, not the model’s.
Here is the analogy that makes it stick. Wreck your knee and you book an orthopedist, not an eye doctor. Both are brilliant. Only one is pointed at your problem. Prompting works the same way. The framework below is how you book the right specialist every single time.
The C.R.I.S.P.Y. Framework for How to Prompt
Context (C)
Set the stage before you ask for anything. Context is the background that tells the model what situation it is walking into. Who the work is for, what the goal is, what is actually at stake. Skip it and even a capable model fills the gaps with assumptions, and its assumptions are just averages scraped off the internet.
Research / Role (R)
This is where most prompting advice is half wrong. The standard move is “act as a nuclear physicist” or “you are an expert marketer.” A role does real work, but not the work people think it does. It sets the lens: tone, vocabulary, depth, point of view. It does not make the model know more.
The research is blunt about this. A 2025 Wharton study titled “Playing Pretend” found that expert personas did not improve factual accuracy. A separate study across 162 roles and thousands of questions found the same thing: adding a persona did not beat using no persona at all, and sometimes made the answers worse. Telling a model it is an expert can even push it to sound right instead of be right.
So split the R in two. Role sets the perspective. Research supplies the knowledge. Do not just say “act as a nuclear physicist.” Say “approach this as a nuclear physicist, and pull the most current information on these specific topics.” If you do not yet know which topics matter, that is your cue to do a little legwork first, or to send the model to find out.
And this half only works if the tools are on. “Research the latest” is an empty instruction to a model that has no way to look anything up. With web access or a connected source switched off, it cannot research a thing. It will pattern recent-sounding claims out of old training data and deliver them with a straight face, which is the exact failure you were trying to dodge. OpenAI’s own guidance puts giving the model real source material near the top of the list for cutting down made-up answers. So before you lean on the Research half, turn on browsing or connect your sources. Tools off, the letter is a lie.
Instructions (I)
Say exactly what you want done. Instructions are the verbs. Build, compare, summarize, rank, rewrite. Break a big ask into ordered steps so the model cannot wander off. Vague verbs get vague work.
Specifics (S)
Hand over the details only you have. Numbers, examples, constraints, reference material, the actual facts of your situation. This is the other half of aiming at the right universe of data. The more real detail you provide, the less the model has to invent.
Parameters (P)
Set the shape of the output. Length, format, tone, what to include, what to leave out. Parameters are how you keep a good answer from showing up in the wrong container.
Yielding (Y)
Tell it how to work with you, not just for you. Should it ask questions before it starts. Should it flag the assumptions it is making. Should it stop and check before producing a long answer. Yielding turns a one-shot guess into a back-and-forth, and the back-and-forth is where the good work happens.
How to Prompt With the C.R.I.S.P.Y. Template
Copy this, fill in the brackets, and you have a complete prompt. Use it with any AI model.
Please follow the CRISPY framework to help with this request:
A Real Example of How to Prompt
Meet Bernice. She owns The Daily Grind, a specialty coffee shop in downtown Portland, three years in, pulling about $30,000 a month. Her morning rush of office workers is loyal. Her afternoons are dead. She has 2,000 Instagram followers and posts whenever she remembers to. She wants 10 percent more revenue in six months on a marketing budget of $1,000 a month.
Here is how her situation becomes a structured prompt:
Please follow the CRISPY framework to help with this request:
Context: A specialty coffee shop in downtown Portland wants to grow revenue. It is profitable but afternoon sales are weak, and the marketing budget is small, so every dollar has to work.
Research / Role: Approach this as a small-business marketing strategist with food-service and local-retail experience. Browse for current, proven tactics that independent coffee shops are using to lift slow dayparts, and use what you find rather than generic advice.
Instructions: Build a six-month plan to grow total revenue by 10 percent. Include specific tactics to fix afternoon sales without hurting the morning business.
Specifics:
- Monthly revenue: $30,000
- Marketing budget: $1,000 per month
- Instagram following: 2,000
- Location: downtown business district
- Strong mornings, weak afternoons
- Core customers: office workers
- Three years in business
Parameters:
- Break the plan down month by month
- Give every tactic an estimated cost and expected impact
- Stay inside the $1,000 monthly budget
- Mix digital and traditional approaches
Yielding:
- Ask clarifying questions until you are 95 percent sure of the requirements
- Break complex strategies into actionable steps
- Flag any assumptions you make about the market or customers
- Stop and check understanding before the detailed recommendations
One thing to confirm before you run it. That Research/Role line tells the model to browse, so the tool has to be available. Live browsing is on by default in most current paid AI tools and in many free ones. If yours cannot browse, either turn the feature on or paste in your own reference material instead, so the model has something real to work from.
How to Get Better at Prompting
- Start simple. Run a basic prompt first, then add the C.R.I.S.P.Y. parts as you see where it drifts.
- Iterate out loud. Do not expect a perfect first answer. Use the response to sharpen the next prompt.
- Feed it real detail. Specificity beats cleverness every time. The facts you provide are the facts it will not have to guess.
- Check the output against your goal. If it missed, look at your prompt before you blame the model.
Your Next Prompt
C.R.I.S.P.Y. is not another acronym to memorize. It is a way to think. Once you internalize that the model patterns toward whatever you point it at, every part of the framework becomes obvious. You are not begging a machine for a good answer. You are removing every reason for it to give you a bad one.
Copy the template, aim it at the right universe of data, and turn your tools on. That by itself will change the quality of everything you get back.
Want to know where you really stand with AI?
The Maker Index is twenty-one uncomfortable questions about value, creativity, and what it means to make something in a world where AI is changing who gets credit for the work. It is not a comfortable assessment. Take it anyway.
Take The Maker Index → ~7 minutes
And follow Flux+Form for more thinking like this:
