Stop making AI guess what you mean

You ask AI for something.

A rewritten email, blunt feedback on a pitch, a summary of a messy doc.

The old way?

Type a fuzzy sentence, get a fuzzy answer, then re-prompt, paste in more context, and re-prompt again.

Now there’s a structure that flips this.

It’s called RAISE.

You give the model 5 things up front, in order, and it stops guessing what you mean from the getgo.

If you’ve used ChatGPT or Claude before, you can start doing this on your very next prompt.

You stop hoping the model reads your mind.

You start telling it exactly what “good” looks like.

Five letters, five things you hand the model before it writes anything.

• R: Role. The lens to answer from. Not fake credentials, just a lens. “You’re a senior engineer reviewing this for correctness” gets a different answer than “you’re explaining this to a non-technical client,” even with identical content underneath.
• A: Ask. The actual task, stated plainly. Say the verb: summarize, critique, rewrite, decide, compare. This is the step people skip most. They set up all the background, then leave the real instruction vague, like “what do you think?” instead of “find the three weakest points in this argument.”
• I: Input. The real material. Paste the actual draft, code, or transcript. Don’t describe it in your own words, paraphrasing loses the specifics the model needs.
• S: Spec. What “good” looks like before it writes a word. Length, tone, format, what to include, what to leave out. This removes the guesswork up front instead of correcting it after a first draft misses.
• E: Example. The shape of the output you want, if one exists. A sample line, a comparable doc, a rough template. One concrete example usually closes more ambiguity than another paragraph of instructions.

Read together: who’s answering, what you actually want, what they’re working with, what counts as done, and what it should look like.

Start with the first two letters. That alone beats most prompts people write.

Role sets who’s answering. Ask states the task, not implied by context.

You’re a senior engineer reviewing this for correctness. Find the bug in this function and explain why it breaks.

Notice what’s missing: no pasted code yet, no format rules. That’s fine. R + A alone already outperforms “can you check this?”

A vague prompt gets you a vague answer, even with a great Role and Ask. The model still needs the real material.

Don’t summarize your draft. Paste it.

You’re a senior engineer reviewing this for correctness. Find the bug in this function and explain why it breaks. Here’s the code: [paste the actual function]

That’s R + A + I, three parts, and already a meaningfully better prompt than most people send in a day.

Now define what “done” looks like before the model starts writing. This is the part that saves you a second and third round of edits.

Length, tone, format, what to include, what to cut.

You’re a direct-response copywriter who cares about ROI over polish. Rewrite this outreach email so it leads with a result, not a service list. Here’s the current draft: [paste email]. Keep it under 100 words, casual tone, end with a specific question instead of a generic CTA.

That’s R + A + I + S. At this point you’re already ahead of almost every prompt written today.

This is the deep end, and it’s the part that closes the gap between “pretty good” and “exactly what I meant.”

Show the model the shape of the output. One real example beats another paragraph of instructions.

Here are two full RAISE prompts, all five parts, back to back:

[Role] You’re a direct-response copywriter who cares about ROI over polish. [Ask] Rewrite this outreach email so it leads with a result, not a service list. [Input] Here’s the current draft: [paste email]. [Spec] Keep it under 100 words, casual tone, end with a specific question instead of a generic CTA. [Example] Here’s an email in the tone I’m going for: [paste example].

[Role] You’re a skeptical investor who’s seen hundreds of pitches and doesn’t care about buzzwords. [Ask] Read this pitch and tell me the three weakest points, not just what’s good about it. [Input] Here’s the deck: [paste text]. [Spec] Keep your answer to five bullet points max, blunt tone, no sugar-coating. [Example] Here’s the kind of blunt feedback I mean: “Your TAM slide assumes 100% capture, nobody believes that number.”

You don’t need to label the parts like that in your actual prompt i.e. you can leave out [Role], [Ask], etc. What matters is that all five things are in there, in that order, before the model starts generating.

Not every prompt needs all five letters. Here’s the rule for each one.

• Ask is the only one you can never drop. It’s the actual task. Without it, there’s nothing to work with.
• Role is optional. If you’re just fixing a bug or converting a file, skip it, there’s no perspective to set. But the moment you want feedback, persuasion, or writing that needs to sound a certain way, bring it back. It’s doing real work there.
• Input, skip it if there’s genuinely nothing to hand over. Asking for a fresh tagline from scratch? Fine without it. But if you’ve actually got a draft, some data, or code sitting in front of you, use it. This is honestly the one people forget most, and it’s usually the one that would’ve made the biggest difference.
• Spec, don’t bother for anything quick and low-stakes. But the second length, tone, or format actually matters, spell it out. Otherwise you’re leaving room for the AI to guess wrong.
• Example is the most skippable of all. Think of it as the last 10%, there to clear up whatever’s still fuzzy after you’ve nailed the Ask and the Spec. You don’t need it every time.

Now hold up, let’s be realistic…

RAISE is NOT a magic spell that gets you a perfect answer on prompt one, every time.

What it does is remove the guesswork that causes most bad AI answers in the first place. You still read the output, still catch what’s off, still iterate. RAISE just means your first draft starts a lot closer to done, because the model wasn’t filling in five blanks on its own.

If you catch yourself writing the same RAISE prompt over and over, same Role, same Spec, just swapping the Input each time, stop retyping it. Turn it into a reusable skill instead.

Check out the article below on how to create your own skills.

Ask never drops. Bring Role back for anything with tone or perspective. Bring Input back the moment you have real material. Bring Spec back once length or format matters. Add Example last, only when things are still fuzzy.

Once you’ve written a RAISE prompt a couple times for the same kind of task, save it as a skill and stop rebuilding it from scratch.

That means no more guessing, no more three rounds of re-prompting, and no more vague “what do you think?”

RAISE isn’t a trick to make AI smarter.

It’s a way to stop making it guess.