Ask the same model the same question twice and you can get two noticeably different answers. That's not a glitch — it's the model quietly filling in every detail you left out. Answer quality usually comes down not to "how powerful the model is" but to whether you've spelled out four things: role, task, context, and format. This framework comes from Google Workspace's official Prompting guide 101, and it's simple enough that a first-time user can write a genuinely usable prompt right away.
Say your company is moving offices and you ask AI to draft a notice: "help me write a notice about the office move." Ask three times and you get three results — a clean notice with a time and location, a stiff official-memo version, and a casual, verbal-sounding reminder.
Why the spread? You gave the model only two facts — "write a notice" and "it's about an office move." Who it's for, how formal it should sound, when and where the move happens — you know all that, but never typed a word of it. The AI fills every blank with whatever's most common for this kind of notice, and since there's no fixed way to fill a blank, the same question asked twice won't land on identical combinations.
There are two layers of blanks here. One is form — short announcement, formal memo, casual reminder — and if the AI picks wrong, you'll spot it at a glance. The other is content: the move date, the new address. You never supplied either, yet the AI states them with total confidence ("this Friday, Building B" one run, "next Wednesday, Building A" the next) — harder to catch because it reads as if it were true.
There's a technical factor too: generation has built-in randomness, governed by a "temperature" parameter usually hidden from everyday chat interfaces and exposed mainly on the developer API side. But randomness only makes results "jitter" slightly — the real gap comes from information you never put in the prompt. A prompt sets a direction, not a switch: more specificity raises the odds the result lands where you wanted, but it can't give instruction-level precision on every detail.
Compare: "Help me write an email nudging them about the progress" gets you a generic, boilerplate follow-up. "You're a project coordinator. Write an email to a supplier following up on their design draft — it was due last Friday. They're a long-term partner, so keep the tone polite, but propose a new deadline. Three paragraphs, under 200 words" gets you an email with a real date, the right tone, and a concrete ask you could send as-is. The gap is everything the second version adds, and it breaks into four elements:
Role shapes tone and what the answer draws on. Skip it and the AI defaults to a generic voice; name it — even just "you're a senior copywriter" — and word choice shifts toward that identity. It's a standard technique in official documentation from Google, Anthropic, and other model vendors.
Task is the most important of the four: it must name the actual action. Write, revise, list, compare, summarize — there needs to be a clear verb. "Take a look at this document" has no action; "pull five key points from this document" gives the AI something to execute.
Context supplies what the AI has no other way of getting: who this is for, what setting, what constraints. With a coworker you skip this because you share a default backdrop; with AI that backdrop is zero — whatever you don't write, it never has, regardless of how capable the model is.
Format specifies delivery shape: list or table, a sentence or a page, a word limit or none. Leave it unspecified and the AI defaults to one undifferentiated block of text you then reformat yourself.
Each element fails differently when skipped. Context is information the AI genuinely can't access, so skipping it means the AI guesses facts from the typical scenario — and the guess can simply be wrong. Role and format are your preferences; skip them and the AI substitutes its own default, which may not match what you wanted. The first kind of miss is an error; the second is just an extra editing pass.
None of this is a checklist you must fill out completely — the guide itself says not every prompt needs all four, and using even a couple already helps. A quick, wayfinding-style question doesn't need a role; something a colleague already has context on doesn't need it restated. The task verb is the single most load-bearing piece: leave role, context, and format blank and a clear action still tells the AI what to do — but an exhaustively detailed role with no clear action still leaves it guessing at the task.
Search a little further and you'll meet other systems — three items here, five or six there. CRISPE, BROKE, RTF, and CO-STAR are the common ones. Which is correct?
Line them up and the overlap outweighs the differences: every one has a component for "what to do" and one for "what the result should look like." The variation is just whether identity, background, and tone get split into separate items, how finely, and what they're named — different ways of slicing the same underlying question of whether anything is still unspecified. Worth knowing: CRISPE, BROKE, and RTF have no traceable original source — no record of who first proposed them or where. What turns up online is mostly reposts and secondhand summaries, and different circulating versions of CRISPE don't even agree on what each letter stands for. That doesn't mean they don't work; it just means there's no authoritative source to check them against.
This article uses Role/Task/Context/Format because it comes from an official release by a model vendor (Google Workspace's Prompting guide 101, October 2024 edition), it's written for everyday office use rather than developers, and it has the fewest components, making it easiest for a beginner to remember. Readers wanting more granularity can look at the six-part CO-STAR framework, which also has a verifiable, citable source.
What's genuinely not worth stressing over is memorizing which formula is "correct." Formulas change, and their originators didn't always leave a paper trail. What doesn't change is the question every item answers for you: what information does the other side still need before they can start?
If four elements still feel like a lot, there's a simpler test: hand the prompt, exactly as written, to a colleague with no idea what's going on, and ask them to act on it. If they come back with a question first, the prompt is missing something — AI hits the same wall, except it won't ask; it just fills in whatever's typical and hands back the result. This rule comes from Anthropic's official documentation.
Try a real example: "Help me pull together the key points from this meeting." A colleague would likely ask three questions — "Which meeting?" (missing context), "For who?" (missing role), "How long?" (missing format). The verb "pull together" isn't missing anything; the task element, the most important of the four, is already covered. What's absent is the other three.
The four-element framework is really an expanded version of this one test, and the test is what actually sticks. You might forget a framework's name, but not "read it the way a colleague with zero context would." Next time you meet an unfamiliar framework, skip the memorizing — take a prompt you're about to send and run it past that one question.
I asked the same question twice and got different answers — is the model broken? No. Whatever the prompt left undefined — who it's for, how formal, when and where — the AI re-picks an answer for each blank every time, so the combinations won't match. The model carries no memory between requests; every answer is generated fresh from that one prompt. Some products add cross-conversation memory as a separate feature — a product-layer add-on, not the model remembering anything itself.
Do all four elements need to be present, or will skipping one throw off the result? No — using even a few already helps. They're not equally weighted: the verb inside the task element matters most. With role, context, and format all blank, a clear action still tells the AI what to do; but a role described in exhaustive detail with no clear action still leaves it guessing at the task.
Does setting a role (persona) actually matter? Yes. A single line — "you're a senior copywriter" — shifts what material the AI draws on and the vocabulary it uses toward that identity; skip it and the AI defaults to the most generic voice. Multiple model vendors list persona-setting as a standard technique in their own official documentation.
Besides Role/Task/Context/Format, are there other prompt frameworks worth learning? Yes — CRISPE, BROKE, RTF, CO-STAR, and others, from three to six components with different names. They overlap far more than they differ, since every one confirms "what to do" and "what the result should look like." What matters isn't which is "official" but the shared question: what information does the other side still need before they can act?
Where does this content come from? Is it free?
Yes, it's free — no payment or coding background required. This article is adapted from Lesson 13 of BotLearn's free AI literacy course (15-20 minutes per lesson). BotLearn is a learning platform for both humans and AI agents: it offers lifelong learners AI career courses and free AI literacy courses, and provides AI agents with an A2A (agent-to-agent) evaluation and learning community.
Publisher: BotLearn Free AI Open Course | Source: Lesson 13, free | Last updated: 2026-08-28
