Google's AI researchers use a specific framework for prompting. It's not secret. It's just buried in technical papers. Here's the practical version.
The Paper I Found
Deep in Google Research's publications, there's a framework they use internally for prompt engineering.
It's not marketed. It's not packaged as a course. It's just there, in technical language, for other researchers.
I translated it into something you can actually use.
The framework is called CO-STAR. Not officially. That's what I'm calling it because each letter represents a component.
And once you understand it, your prompts will never be the same.
Why This Matters
Most people write prompts like this:
"Write a blog post about productivity."
Google's researchers write prompts like this:
"You are a productivity coach with 15 years of experience working with remote teams. Write a 1,200-word blog post analyzing why traditional productivity advice fails for remote workers and offering three evidence-based alternatives. Your audience is managers who are frustrated with their team's output. Use a tone that's authoritative but empathetic. Include specific examples. End with one actionable framework they can implement this week."
See the difference?
The second prompt gives AI everything it needs to produce something useful.
The first leaves AI guessing.
CO-STAR is the system that creates prompts like the second one. Every time.
The CO-STAR Framework
C = Context
Where is this taking place? What's the broader situation?
Bad: "Write an email."
Good: "I'm a project manager at a software company. Our product launch is delayed by two weeks due to unexpected technical issues. I need to email our CEO with an update."
Context tells AI what world it's operating in.
O = Objective
What are you trying to accomplish? What does success look like?
Bad: Implied or unclear.
Good: "The objective is to inform the CEO of the delay, explain the cause without technical jargon, reassure them we're handling it, and set a new timeline. I want to maintain their confidence in the team."
Objective defines what winning looks like.
S = Style
How should it sound? What's the tone? What's the voice?
Bad: Not specified.
Good: "Style: Professional but not stiff. Confident without being defensive. Short sentences. No corporate jargon. Think 'competent leader providing an update' not 'employee making excuses.'"
Style controls how it reads.
T = Tone
What's the emotional temperature? How should it feel?
Bad: Not specified.
Good: "Tone: Calm and solution-focused. Acknowledging the setback but emphasizing the path forward. Reassuring without overpromising."
Tone controls how it lands.
A = Audience
Who will consume this? What do they know? What do they care about?
Bad: Assumed.
Good: "Audience: CEO with limited technical background who cares about customer impact and business outcomes, not technical details. They value transparency and hate surprises."
Audience shapes everything else.
R = Response Format
How should the output be structured? What's the format?
Bad: AI decides.
Good: "Format: Email, 200-250 words. Start with the bottom line (new timeline). One paragraph on the cause (high-level). One paragraph on the solution and prevention. End with a clear next step."
Response format ensures you get what you need.
The Template
Here's the plug-and-play version:
Context: [Describe the situation - who you are, what's happening, relevant background]
Objective: [What you're trying to accomplish - be specific about the goal]
Style: [How it should sound - tone, voice, writing style]
Tone: [The emotional quality - confident, urgent, friendly, etc.]
Audience: [Who will read/use this - their knowledge level, what they care about]
Response Format: [Structure, length, format requirements]
[Your specific request]Real Examples
Let me show you CO-STAR in action across different scenarios.
Example 1: Business Writing
Before CO-STAR:
"Write a product launch announcement."
After CO-STAR:
Context: I'm a product marketing manager at a B2B SaaS company. We're launching a major new feature that our customers have been requesting for two years. This is our biggest product update this year.
Objective: Announce the feature to existing customers via email in a way that drives them to try it within the first week. I want high engagement and feature adoption.
Style: Conversational but professional. Short paragraphs. Excitement without hype. Focus on benefits over features.
Tone: Enthusiastic and customer-focused. "We heard you" energy.
Audience: Existing B2B customers, mostly operations managers and team leads, who are busy and skim emails. They care about solving their workflow problems, not technical specs.
Response Format: Email, 250 words max. Subject line + preview text, brief intro acknowledging the request, 3 benefit bullets, clear CTA to try it now, sign-off.
Write the product launch announcement.The output? Actually usable. Minimal editing needed.
Example 2: Creative Work
Before CO-STAR:
"Give me ideas for a marketing campaign."
After CO-STAR:
Context: I run marketing for a local gym that's been losing members to boutique fitness studios. We have great equipment and lower prices, but we're seen as "generic." Our target members are 25-45, value-conscious but want community.
Objective: Generate 5 campaign ideas that reposition us as a community hub, not just a gym. I want ideas that are achievable with a $3,000 budget and can run for 3 months.
Style: Creative but practical. Each idea should be specific enough that I could start executing tomorrow.
Tone: Innovative but grounded. No pie-in-the-sky ideas.
Audience: This is for me (marketing manager) and the gym owner to review and pick from. We need to understand the strategy and see the execution path.
Response Format: For each of 5 ideas, provide:
- Campaign name
- Core concept (2 sentences)
- Execution steps (3-4 specific tactics)
- Estimated budget breakdown
- Success metrics
Generate the campaign ideas.The output? Five actionable campaigns. Not vague suggestions. Real plans.
Example 3: Technical Explanation
Before CO-STAR:
"Explain APIs."
After CO-STAR:
Context: I'm a sales rep at a software company. I'm talking to potential clients who need to integrate our software with their existing systems. They keep asking about "API capabilities" and I need to explain what that means and why it matters.
Objective: Understand APIs well enough to explain them confidently to non-technical clients and answer basic questions about integration. I need to sound knowledgeable without getting technical.
Style: Simple language. Use analogies. Build from what business people already understand.
Tone: Clear and confident. Educational, not condescending.
Audience: Me, a sales rep with no technical background. I understand business processes but not programming.
Response Format:
- One-paragraph explanation using a business analogy
- Three key benefits of having an API (in terms clients care about)
- Three common questions clients ask and simple answers
Explain APIs for my sales context.The output? Something you can actually use in sales calls. Not Wikipedia.
Why Each Component Matters
Let's break down what happens when you skip each one.
Skip Context → Generic Output
Without context, AI doesn't know what rules apply.
"Write a project update" could be for a tech startup or a construction project. The approach is totally different.
Context grounds the response in reality.
Skip Objective → Aimless Output
Without a clear objective, AI doesn't know what success looks like.
You get something that covers the topic but doesn't accomplish anything specific.
Objective creates direction.
Skip Style → Inconsistent Output
Without style guidance, AI uses its default style. Which is... generic professional.
Sometimes you need casual. Sometimes formal. Sometimes punchy. Sometimes flowing.
Style controls the feel.
Skip Tone → Flat Output
Tone is the emotional layer. Without it, everything reads the same.
"We're excited to announce" and "We regret to inform" need different tones. Obviously.
But AI won't know unless you specify.
Skip Audience → Mismatched Output
AI doesn't know who you're talking to unless you tell it.
Explaining the same concept to a 5-year-old, a college student, and an expert requires completely different approaches.
Audience shapes complexity, examples, and assumptions.
Skip Response Format → Wrong Structure
AI will pick a format. It might not be what you need.
Need bullet points? Might get paragraphs.
Need 100 words? Might get 500.
Need an email? Might get an essay.
Response format ensures you get usable output.
The Advanced Move
Once you're comfortable with CO-STAR, add this component:
E = Examples
Show AI what good looks like.
Context: [your context]
Objective: [your objective]
Style: [your style]
Tone: [your tone]
Audience: [your audience]
Response Format: [your format]
Examples of what I'm looking for:
[Example 1]
[Example 2]
Now create something similar for: [your specific request]Examples are the fastest way to get exactly what you want.
Show, don't just tell.
Common Mistakes
Mistake 1: Being Vague in Context
Bad: "I need to write something for work."
Good: "I'm a customer success manager at a SaaS company. I need to write a response to a frustrated customer who's experiencing repeated bugs in our mobile app. They're a paying customer threatening to churn."
Specific context = specific output.
Mistake 2: Multiple Objectives
Bad: "The objective is to inform them about the issue, apologize, offer a solution, get feedback, and upsell them on our premium plan."
That's five objectives. Pick one primary objective.
Good: "The objective is to retain this customer by acknowledging their frustration and presenting a specific fix plan."
Mistake 3: Confusing Style and Tone
Style is how it sounds. Tone is how it feels.
"Casual and friendly" describes both style and tone.
"Short sentences and reassuring" separates them.
Be specific about each.
Mistake 4: Assuming Audience
"Audience: Business people" is too broad.
"Audience: Small business owners with 5-20 employees who are tech-averse and budget-conscious" is specific.
The more specific your audience, the better AI can tailor the output.
When CO-STAR Is Overkill
You don't need the full framework for simple tasks.
"Summarize this article in 3 bullet points" doesn't need CO-STAR.
Use CO-STAR when:
- The output really matters
- You need something specific
- You're tired of editing generic AI output
- The task is complex
- You'll use this type of prompt repeatedly
For quick, simple requests? A basic prompt is fine.
Building Your Prompt Library
Here's what Google's researchers do: They save successful prompts.
Not just the output. The prompt itself.
Build a library:
- Email templates (various scenarios)
- Content creation (blog posts, social, etc.)
- Analysis frameworks (for different types of problems)
- Explanation templates (for different audiences)
When you need something similar? Adapt an existing prompt.
Don't start from scratch every time.
The One-Minute Version
If you only remember one thing, remember this:
Give AI six things:
- Where you are (Context)
- What you want (Objective)
- How it should sound (Style)
- How it should feel (Tone)
- Who it's for (Audience)
- What format you need (Response Format)
Do that, and your prompts will be better than 95% of people using AI.
The Challenge
This week, take one task you'd normally struggle with.
Write the prompt using CO-STAR. All six components. Be specific.
Run it.
Compare it to what you'd normally get with a simple prompt.
The difference will convince you more than I can.
Why This Works
Google's researchers figured out something important:
AI is incredibly capable. But it's not a mind reader.
The more precisely you communicate what you want, the more precisely AI can deliver.
CO-STAR is just a framework for precise communication.
You're not manipulating AI. You're not hacking it.
You're just being clear about what you need.
Turns out, that's the whole game.
Now go be clear.