How to Write Better AI Prompts: One Framework That Works

May 31, 2026

There's a moment everyone has with AI tools where the output is almost right but somehow useless — the email is too formal, the image is the wrong mood, the video moves in slow motion when you wanted energy. The reflex is to blame the model. Usually the model did exactly what you asked. The problem is what you asked.

Chat with SentX

Good prompting is not a secret list of magic words. It's a habit of saying out loud the things you were keeping in your head. This guide gives you one framework that works the same way whether you're chatting, generating an image, or making a video, plus copy-paste templates, fully written before/after examples, and a checklist you can keep next to your keyboard.

The universal prompt skeleton: role + context + task + constraints + format

The single most reliable upgrade to any prompt is to include five pieces of information: role, context, task, constraints, and format. Most disappointing outputs are missing two or three of them. The model fills the gaps with the most average possible guess, and "average" is rarely what you wanted.

Here's what each piece does:

Role — the perspective or expertise you want the model to write from. "A skeptical copy editor" produces different output than "an enthusiastic marketer." This is the cheapest, highest-leverage word in your prompt.
Context — who you are, what situation you're in, and what already happened. The model can't see your screen, your last meeting, or your audience. If it matters to the answer, say it.
Task — the actual thing you want done, stated as a verb. "Summarize," "rewrite," "brainstorm," "critique." Vague tasks get vague results.
Constraints — the boundaries: length, tone, reading level, what to avoid, what must be included. Constraints are where most of the quality lives.
Format — the literal shape of the output: a bulleted list, a table, three numbered options, a single paragraph, a code block.

Here is the copy-paste skeleton. Fill in the brackets and delete any line that genuinely doesn't apply:

Role: Act as a [specific expert/persona].
Context: I'm [who you are] working on [situation]. So far, [what's happened].
Task: [One clear verb] the following: [your material or request].
Constraints: Keep it [length]. Tone should be [tone]. Avoid [pitfalls].
  Must include [non-negotiables].
Format: Respond as [list / table / numbered options / paragraph / code].

You will not need all five every time. A quick factual question — "what year did the Voyager 1 probe launch?" — needs none of this. But the moment the answer matters, missing one of these five is almost always why it disappointed you. When you find yourself frustrated, run down the list and ask which piece you forgot to say.

The biggest fix: specificity over politeness

If you change only one habit, change this: stop being polite and start being specific. "Please write me a nice blog intro, thanks!" feels considerate, but "nice" tells the model nothing. Politeness adds words; specificity adds information. Only one of those changes the output.

Here's a real before-and-after. Watch how the vague version forces the model to guess, and the specific version removes every guess.

Vague prompt:

Write a product description for my coffee.

What that produces is something like this — generic, interchangeable, true of any coffee on earth:

Discover our premium coffee, crafted with care from the finest beans. Rich, smooth, and full of flavor, it's the perfect way to start your day. Experience quality in every cup.

That's not bad writing. It's just not about your coffee. The model had nothing to work with, so it wrote the average of all coffee descriptions.

Specific prompt:

Role: Act as a copywriter for small-batch specialty coffee.
Context: My product is a single-origin Ethiopian Yirgacheffe, light roast,
  with floral and citrus notes. Customers are home brewers who own a
  pour-over setup and care about origin and roast date.
Task: Write a product description for the bag.
Constraints: 60-80 words. Warm but not gimmicky — no "wake up and smell."
  Mention the brewing method it shines in. Don't claim health benefits.
Format: One short paragraph, then a single line of three tasting-note words.

What that produces:

This single-origin Yirgacheffe is picked from Ethiopia's highlands and roasted light to keep its character intact. Brew it as a pour-over and you'll find a cup that leans bright and floral, with a clean citrus lift that lingers without turning sharp. Roasted in small batches so it reaches you close to its peak.

Floral · Citrus · Clean.

Same model, same five seconds of compute. The only difference is that the second prompt did the thinking before hitting enter. Notice you didn't write a single "please" — and the result is dramatically better. The lesson generalizes: every adjective you leave out ("light roast," "pour-over," "60-80 words") is a decision you've handed to a machine that doesn't know your goals.

Iterating: fix a bad output in the NEXT turn instead of restarting

When an answer comes back wrong, your best move is almost never to delete it and rewrite the whole prompt from scratch. Tell the model what's wrong with the output it just gave you, in the next message. The model already has all the context from the conversation; you only need to nudge the one thing that missed.

This works because a conversation is cumulative. Starting over throws away everything the model already understood about your situation. Correcting in place keeps the good parts and surgically fixes the bad ones.

The pattern is: name what's wrong → say what you want instead → keep everything else. Concrete examples of good follow-up turns:

Good, but cut it to half the length and drop the last paragraph entirely.

The tone is too corporate. Rewrite it the way you'd explain it to a friend
over coffee — keep the same three points.

This is closer. Now make example #2 a real, specific scenario instead of
a placeholder, and leave the rest as-is.

Each of these is one sentence and produces a targeted revision. Compare that to retyping a 100-word prompt and hoping the model lands differently this time. Iteration is the actual skill — your first prompt is a hypothesis, and the follow-up turns are where you converge on what you actually wanted.

This is also where conversation memory matters. On tools that carry your context forward — including an AI that remembers your conversations like SentX — you don't have to re-explain your project, your audience, or your preferences in every session. The iteration loop gets shorter because the model already knows who you are and what you're working on. If you regularly explain complex material, the same iterative instinct powers a good workflow for getting AI to explain complex topics: get a first explanation, then say "now assume I already know calculus" or "use a kitchen analogy."

Cross-surface notes: chat vs image vs video

The same five-part skeleton works across chat, images, and video — but each surface weights the pieces differently. Knowing which piece carries the most load on each surface is what separates a usable result from a frustrating one.

Chat leans hardest on task and constraints. The model can produce nearly anything, so your job is to fence it in: length, tone, what to avoid, what format. Role and context sharpen it; constraints make it usable.
Image leans hardest on role and format — except here "role" means subject and style and "format" means composition. An image model can't ask follow-up questions, so a single well-built prompt has to carry subject, style, lighting, framing, and mood all at once. There's a whole craft to ordering those words, covered in the anatomy of an AI art prompt.
Video is image plus motion as a first-class constraint. Everything that matters for a still matters here, and then you add what moves, how fast, and how the camera behaves. Skipping the motion description is the single most common video-prompt mistake — the model will pick a default that's rarely what you imagined. If you're animating an existing picture, see how to add motion to a still image.

The throughline: on chat you constrain a model that can do too much; on image and video you fully specify a model that can't ask you to clarify.

Worked examples across all three surfaces

Here is the same skeleton applied to each surface, fully written out, so you can see how the five pieces translate. Adapt the brackets to your own project.

Chat: a working email rewrite

Role: Act as a direct, friendly product manager.
Context: I need to tell a teammate their deadline is slipping by a week.
  We have a good relationship; I don't want to sound cold.
Task: Rewrite my draft below to be honest but kind.
  Draft: "The timeline is no longer realistic. We need to move the date."
Constraints: Under 90 words. No corporate filler. Acknowledge their effort.
  Offer help, don't just announce the slip.
Format: A ready-to-send email with a subject line.

This produces a usable email because every decision — length, tone, what to acknowledge, the subject line — is already made. The model isn't guessing whether you want warm or curt; you told it.

Image: a specific scene

A weathered wooden lighthouse on a rocky Atlantic coast at golden hour,
warm low sunlight raking across the rocks, long shadows, a few gulls in
the distance, calm sea with gentle swell. Photographic, shot on a 35mm
lens, shallow depth of field, soft natural color grade. Composition:
lighthouse positioned on the left third, horizon low, lots of sky.

Notice how the subject (lighthouse, coast), style (photographic, 35mm), lighting (golden hour, raking sun), and composition (left third, low horizon) are all present. That's role and format in image clothing. You can build this kind of prompt in any image tool, including the SentX AI image generator. If your first result is close but the framing is off, iterate the same way you would in chat: regenerate with "move the lighthouse to the right third and lower the camera."

Video: motion as a first-class element

A weathered lighthouse on a rocky coast at golden hour. Camera slowly
pushes in toward the lighthouse over 4 seconds. Gentle waves roll in and
recede, the beam of the lighthouse rotates once, gulls drift across the
frame. Warm cinematic lighting, calm and contemplative mood, smooth steady
motion — no fast cuts, no shaky cam.

This is the image prompt plus a clear answer to what moves and how. The camera push, the wave rhythm, the rotating beam, the steady pacing — each is a motion constraint the model would otherwise guess. You can run prompts like this in the SentX AI video generator. When the motion is wrong, fix it in the next turn: "same shot, but make the camera static and let only the waves and beam move."

Three surfaces, one skeleton. Once you internalize role-context-task-constraints-format, switching between them stops feeling like learning three different tools.

A reusable prompt checklist

Before you hit enter on any prompt that matters, run this checklist. It takes ten seconds and catches the mistakes that cause the large majority of bad outputs.

Did I name a role or perspective? Even one word ("act as an editor") changes everything.
Did I give the context the model can't see? Who you are, your audience, what already happened.
Is my task a clear verb? Summarize, rewrite, critique, generate — not "help me with."
Did I set constraints? Length, tone, reading level, what to avoid, what must appear.
Did I specify the output format? List, table, paragraph, options, code block, composition.
For image/video: did I describe style, lighting, and (for video) motion? These are non-optional on visual surfaces.
Did I trade politeness for specificity? "Please make it nice" is not an instruction.
Do I have a follow-up plan? Know that you'll correct in the next turn, not restart.

Keep this list visible until the habit is automatic. Most people stop needing it within a week — the five pieces become how you naturally think about asking.

If your work is creative, the same discipline pays off in unexpected places; prompts for creative writing with AI use the exact same skeleton, just with "role" set to a voice and "constraints" set to form and meter.

FAQ

What is the best framework for writing AI prompts?

The most broadly useful framework is role + context + task + constraints + format. It's not an acronym to memorize so much as a checklist of the five things a model needs to know. Unlike narrower frameworks, it works the same across chat, image, and video — you just shift which pieces carry the most weight per surface. For chat, lean on task and constraints; for image and video, lean on style, composition, and motion.

Do I really need all five parts every time?

No. A simple factual question needs none of them. Use the full skeleton when the result matters — when you'd be annoyed to get a generic or wrong answer. The faster diagnostic: when an output disappoints you, run down the five parts and find which one you forgot to include. That missing piece is almost always the reason.

Why does being more specific work better than being polite?

Politeness adds words but no information; the model can't act on "nice" or "good." Specificity removes the model's need to guess. Every detail you leave out — the length, the tone, the audience, the format — becomes a decision the model makes on your behalf, using the most average option. Specific prompts simply make those decisions yourself, before hitting enter.

How do I fix a bad AI response without starting over?

Correct it in the next message instead of rewriting the whole prompt. The model retains the conversation's context, so you only need to point at the one thing that's wrong: "cut this in half," "make the tone warmer," "make example two specific." Restarting throws away everything the model already understood. Iterating keeps the good parts and surgically fixes the rest — and on tools that remember your conversations, the loop gets even shorter because you don't re-explain your context each time.

Are prompts different for images and video than for chat?

The framework is the same; the emphasis shifts. Chat models can do almost anything, so your job is to constrain them with task, length, and tone. Image and video models can't ask you clarifying questions, so a single prompt must fully specify subject, style, lighting, and composition up front — and video adds motion (what moves, how fast, how the camera behaves) as a required element. See the anatomy of an AI art prompt and the guide to adding motion to a still image for the surface-specific craft.

How do I get better at prompting over time?

Treat your first prompt as a hypothesis and the follow-up turns as the real work. Keep the checklist visible, notice which of the five parts you tend to skip, and pay attention to which corrections you make most often — those reveal the constraints you should have stated up front. Within a week or two, the five pieces become automatic, and you'll spend far less time fighting outputs and far more time getting what you actually wanted on the first try. You can practice the whole loop — chat, image, and video — in one place by creating a free account or reviewing the plans.