How to Write AI Prompts: A Framework That Actually Works

Ask two people to get help from the same AI model, and you can end up with two completely different results — one gets a generic, half-useful answer, and the other gets something they can actually use. The model didn't change. The prompt did.

That gap isn't magic. Google, OpenAI, and Anthropic all publish official guidance on how to structure a prompt, and once you see the pattern across their documentation, it's less "art" and more a checklist you can reuse for almost anything — writing, research, images, even reviewing a contract.

What Do Google, OpenAI, and Anthropic Actually Recommend?

Despite being three separate companies, their prompting guides converge on the same handful of building blocks.

Google Workspace's official prompting guide describes a prompt as Persona – Task – Context – Format, and gives this exact example: "You are a program manager in [industry]. Draft an executive summary email to [persona] based on [details about relevant program docs]. Limit to bullet points."

OpenAI's official API documentation structures prompts as Identity – Instructions – Examples – Context, and recommends giving a model a role as plainly as: "You are coding assistant that helps enforce the use of snake case variables in JavaScript code."

Anthropic's documentation for Claude frames the whole exercise with a useful mental model: treat the model "like a brilliant but new employee" who has no context on your organization's norms. Its "golden rule" is that if a colleague with minimal context would be confused by your prompt, the model probably will be too. Anthropic's own recommended techniques include assigning a role in a single sentence, providing three to five relevant and varied examples, wrapping instructions and context in tags like <instructions> and <context>, and explaining why a rule matters rather than just stating it.

Separately, CO-STAR (Context – Objective – Style – Tone – Audience – Response) has a clearly documented origin: it was published by GovTech Singapore's Data Science and AI Division in its 2023 "Prompt Engineering Playbook," and later popularized after a GovTech staffer used it to win a GPT-4 prompting competition. Other shorthand frameworks that circulate online — RTF (Role–Task–Format), RACE (Role–Action–Context–Execute/Expectation), and TAG (Task–Action–Goal) — are genuinely useful, but no single inventor or official publication can be pinned down for them. They're best described as informal industry conventions rather than the product of any one company or institution.

Line all of this up, and it maps cleanly onto a simple four-part structure: who the AI should act as, what it should accomplish, what background or constraints apply, and what shape the answer should take.

Simple structure	Where it shows up officially
Role / Persona	Google Workspace's Persona, OpenAI's Identity, Anthropic's role prompting
Goal / Objective	Google Workspace's Task, CO-STAR's Objective
Context / Conditions	CO-STAR's Style/Tone/Response, OpenAI's Instructions
Situation	Google Workspace's and CO-STAR's Context, RACE's Context

Here's a general-purpose template built on that structure. It's generic enough to adapt for almost any task:

You are [persona/role]. Your goal is [the specific outcome you want]. Context: [relevant background, constraints, or audience]. Situation: [the specific scenario you're dealing with right now]. Format the response as [desired structure — e.g., bullet points, a table, a short paragraph].

The rest of this guide adapts that same skeleton to four very different jobs: generating images, running deep research, reviewing legal documents, and looking into a health question.

How Do You Get Better Images Out of Gemini?

This is the area where a well-built prompt makes the biggest visible difference, so it's worth going into more depth here than anywhere else.

As of September 2026, Google's current image models are commonly nicknamed "Nano Banana" — the standard Gemini 3.1 Flash Image ("Nano Banana 2") and the higher-end Gemini 3 Pro Image ("Nano Banana Pro"). According to Google's official documentation, they support aspect ratios including 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9, with resolutions up to 1K, 2K, or 4K depending on the model. Google's own examples show that naming a style up front — "a photorealistic wide-angle shot" versus "an Impressionist-style oil painting" — changes the entire character of the output before anything else in the prompt matters.

Google's official prompting formula for a text-only image request is: Subject + Action + Location/context + Composition + Style. On top of that, Google's developer and Cloud documentation supply specific vocabulary worth borrowing directly:

• Camera angle / composition: "wide-angle shot," "extreme close-up," "low angle shot," "top-down drone perspective," "eye-level shot"
• Lens / film texture: "shot on medium-format analog film, pronounced grain, high saturation," "shallow depth of field (f/1.8)"
• Lighting: "golden hour," "dramatic teal and magenta rim lighting," "three-point softbox setup"
• Color / background: "cinematic color grading with muted teal tones," "thick, teal-tinted fog"

A Template You Can Reuse for Any Image Prompt

Style: [photorealistic / illustration / sticker / painting style — name it first]
Subject: [main subject and what it's doing]
Composition: [camera angle or shot type, e.g., wide-angle shot, extreme close-up, top-down]
Lighting: [e.g., golden hour, three-point softbox, dramatic rim lighting]
Color and background: [palette, setting, backdrop]
Aspect ratio: [e.g., 1:1, 16:9, 9:16]
Resolution: [1K / 2K / 4K]

Example: A Blog Thumbnail Sized for Actual Use

Here's that template filled in for a real, practical case — a blog post thumbnail, with the sizing and layout constraints spelled out inside the prompt itself:

A photorealistic, cinematic wide-angle shot of a minimalist home workspace at dusk: a laptop screen glowing with soft abstract lines of code, gently blurred in the background. Golden hour side-lighting falls across the desk, casting long, soft shadows. Color palette: warm amber tones against a deep charcoal-teal backdrop. Composition: center-weighted, with generous open space across the middle third of the frame so a headline can be overlaid afterward. Aspect ratio 16:9, resolution 2K. Keep the central area clean and uncluttered so it still reads clearly when cropped down to a small thumbnail or social preview card.

Naming the aspect ratio, resolution, and a deliberately empty "safe area" for text is what turns a nice picture into something that actually works as a usable blog thumbnail rather than a random image that happens to look good.

How Do You Use a Real Photo as a Reference Image?

This is the part most people don't realize Gemini can do. According to Google's official documentation, Gemini accepts up to 14 reference images in a single prompt, and can use them to add elements, edit only one specific part of a photo (a form of inpainting), transfer an existing photo into a different art style, or combine several unrelated reference images where each one is assigned a distinct role — for example, one image for pose, another for art style, another for background. Google's own documentation gives this exact instruction as an example of element editing: "Using the provided image of my cat, please add a small, knitted wizard hat on its head."

That means you don't have to generate everything from scratch — you can hand Gemini an actual photo and tell it precisely what to keep and what to change:

Using the attached photo of [your product/object] as the structural reference, regenerate it in a clean, minimalist studio setting. Keep the object's exact shape, proportions, and any logos or markings unchanged. Replace the background with a seamless, soft gray studio backdrop. Light it with a three-point softbox setup so surface reflections look natural. Do not alter the object itself — only the background, lighting, and shadows should change.

This directly reflects Google's documented approach to inpainting and element-preserving edits: naming exactly what must stay fixed (the product) and what's allowed to change (the background and lighting) is what keeps the model from redesigning the whole subject by accident.

Keeping a Character or Style Consistent Across Several Images

Google's official prompt guide also recommends assigning each reference image a specific job when you're combining more than one — useful for a short image series that needs to feel visually consistent:

Reference image A [attach: character-reference.jpg]: use this as the fixed reference for a specific character — keep the face, hairstyle, and outfit identical in every image.
Reference image B [attach: style-reference.jpg]: use this only for the overall art style (soft watercolor texture, muted pastel palette) — ignore its subject matter entirely.
New scene: the character described in Reference A is standing on a train platform at sunrise, watching a train depart.
Render the new scene in the style from Reference B, while keeping the character's appearance from Reference A completely consistent with previous images.

Google's guidance also flags a smaller but easy-to-apply habit: describe what you want rather than what you don't want. "Empty street" works better than "no cars."

How Do You Prompt Gemini for Deep Research?

Deep research tools work best when they're treated less like a search box and more like a well-briefed research assistant. Cross-checked guidance on this points to five elements a good research prompt should include: a role for the AI to adopt, a specific task, the scope/context (industry, timeframe, audience), the desired output format, and — always — an explicit instruction to cite sources.

You are a [type of analyst/researcher]. Research [specific topic or question] for [purpose/audience].
Scope: focus on [market/category/time range], prioritizing [primary sources — official filings, peer-reviewed studies, government data] and using industry reports only to fill gaps.
Output format: [e.g., a structured brief with headers, a comparison table, a bulleted list of findings].
If sources conflict, present both and note which one appears more reliable and why.
Cite all sources.

Example:

You are a market research analyst. Research the current state of solid-state battery commercialization for electric vehicles, for an internal briefing aimed at a product strategy team with no deep battery-chemistry background.
Scope: focus on developments from 2025-2026, prioritizing company filings, peer-reviewed papers, and industry-body data over general news coverage.
Output format: a one-page brief with headers for "Current State," "Key Players," "Timeline," and "Open Risks," plus a short comparison table for at least three companies.
If sources disagree on timelines, present both estimates and flag which source appears more authoritative.
Cite all sources.

Anthropic's own documentation recommends a similar structured approach for complex research tasks, suggesting the model be told to "develop several competing hypotheses" and "track confidence levels" as it works — worth adding to the prompt above if the topic is genuinely uncertain or contested.

Can AI Help You Review a Legal Document?

Used carefully, yes — as a first pass, not a final answer. Thomson Reuters' legal-tech guidance recommends a simple structure for legal prompts: Intent + Context + Instruction, expanding for more complex tasks into role/context, task, reference source, output format, and constraints (jurisdiction, tone, what not to do).

You are a legal-assistant tool helping review a [contract type / document type]. Using the attached [document name], identify clauses that are commonly flagged for review under [jurisdiction] practice, and explain in plain language why each one might warrant a closer look.
Output format: a table with columns for "Clause," "Why It's Flagged," and "Suggested Question for Counsel."
Constraints: do not state whether any clause is enforceable or unenforceable. End the response by noting that a licensed attorney should confirm all findings before any decision is made.

Example — reviewing a standard NDA:

You are a legal-assistant tool helping review a standard mutual non-disclosure agreement (NDA). Using the attached NDA draft, identify clauses that commonly get flagged in general U.S. business practice, and explain in plain language why each one might need a closer look.
Output format: a table with columns for "Clause," "Why It's Flagged," and "Suggested Question for Counsel."
Constraints: do not offer a definitive legal conclusion on any clause. End the response by noting that a licensed attorney should confirm all findings before this NDA is signed.

This does not replace a lawyer's judgment, and it isn't legal advice. The American Bar Association's Formal Opinion 512, issued July 29, 2024, states that a lawyer's obligations under the Model Rules of Professional Conduct — competence, confidentiality, communication with clients, candor toward the court, and supervisory responsibility — apply in full when generative AI is used in legal work. The opinion places the responsibility for verifying anything AI produces squarely on the lawyer, noting that the necessary depth of verification "will vary depending on the task and the AI tool being used." Thomson Reuters echoes this directly: lawyers "still always need to use their professional skills and judgment to decide how to use the information and insights AI provides," because "AI assists with legal reasoning but doesn't replace it." Multiple lawyers have already faced professional discipline for submitting court filings containing AI-fabricated case citations — a reminder that anything AI produces in this space needs to be checked against the actual source documents before anyone relies on it.

Can AI Help You Understand a Health Concern?

The American Medical Association's official guidance offers five specific prompt patterns patients can use when researching a health question with an AI chatbot:

I have these symptoms: [description]. What are common causes, how do they differ, and what details would help narrow them down?

Explain this to me in simple words: [medical term, diagnosis, or instructions from a doctor].

Given my situation [age, gender, lifestyle, goals, etc.], what do I need to know about [condition/topic]?

What are the different ways to treat [diagnosis/condition], and what are the general benefits, risks, and side effects of each, based on publicly available information?

What questions should I ask my doctor to better understand [situation/diagnosis]?

Example, filling in the first pattern:

I've been getting a dull headache behind my eyes most afternoons for the past two weeks. What are common causes, how do they differ, and what details would help narrow them down?

None of this replaces a real medical consultation, and it isn't a diagnosis. The AMA's CEO, Dr. John Whyte, has stated plainly that "as technology evolves, it's vital that patients use AI to complement — not replace — the advice of their doctors," and the AMA recommends contacting emergency services directly in any urgent situation rather than querying a chatbot. The World Health Organization's 2024 guidance on large multimodal models in health lays out more than 40 recommendations precisely because the risks differ across use cases — diagnosis, patient-facing tools, administration, education, and research all carry distinct ethical concerns. The patient-safety research organization ECRI went further, naming AI chatbot misuse the top health-technology hazard on its 2026 list, warning that these tools are neither regulated as medical devices nor validated for medical use, and that users reduce their risk mainly by "always verifying information obtained from a chatbot with a knowledgeable source." ECRI also flags a subtler risk worth knowing about: chatbots can create an "echo chamber effect," reinforcing a user's existing belief even when that belief is wrong.

A Quick Note on Coding and Writing Prompts

The same building blocks carry over to smaller, everyday tasks. GitHub's official Copilot documentation recommends stating the overall goal before diving into details, breaking complex coding requests into smaller pieces, and supplying example input/output where possible — sometimes organized informally as Persona–Context–Task–Format. For writing and summarizing, the same logic applies: specify the tone, give a target length, state what the summary should actually center on, and — where possible — supply one example of the style you want, so the model has something concrete to match rather than guessing.

So Is Prompt Engineering Really a Skill Anyone Can Learn?

Looking across all of this, the pattern holds up: whether it's Google, OpenAI, Anthropic, a legal-tech vendor, or a medical association, "good prompting" keeps reducing to the same short list — say who the AI is acting as, say exactly what you want, give it the background it needs, and tell it what shape the answer should take. None of that requires special access or technical training.

What changes by use case isn't the underlying structure — it's how much specificity actually matters. A vague image prompt produces a vague image; naming the camera angle, the lighting, and the aspect ratio produces something usable. A vague research prompt produces a shallow summary; specifying scope, sources, and output format produces something closer to a real brief. And in the two areas where getting it wrong actually matters — legal and health questions — the official guidance converges on the same caution: AI can organize information and speed up a first pass, but it doesn't carry the judgment or accountability that a licensed professional does. Building that caveat into how you use these tools, rather than treating it as a footnote, is arguably the most transferable "prompt engineering" skill of all.