Most people write prompts the way they'd leave a voice note: a quick thought, no real structure, and then they wonder why ChatGPT or Claude gave them something generic. The output isn't broken. The input was just missing pieces the model needed to do a good job.
CO-STAR is a simple framework that fixes this. It's six categories of information that, together, turn a vague ask into a prompt that actually gets you what you wanted the first time. Once you see the pattern, you can't unsee it, and you'll start noticing it in almost every well-written prompt you come across.
What Is CO-STAR?
CO-STAR is an acronym for the six things a good prompt should specify:
- C: Context
- O: Objective
- S: Style
- T: Tone
- A: Audience
- R: Response format
The idea started circulating in the prompt engineering community as a way to structure prompts for large language models, and it stuck around because it works with any model, ChatGPT, Claude, Gemini, or anything else. It doesn't require special syntax or plugins. It's just a checklist you run through before you hit send.
Here's the letter by letter breakdown, with a quick before and after for each one.
C: Context
Context is the background the model needs so it isn't guessing. Who you are, what the situation is, what's already been tried. Without it, the AI fills in the blanks with generic assumptions, and generic assumptions produce generic answers.
Before: "Write a product description for my shoes."
After: "I run a small brand that sells handmade leather sneakers to people who care about sustainability and craftsmanship. Write a product description for our new low-top design."
O: Objective
Objective is the actual goal of the prompt: what you want the output to accomplish, not just what topic it should cover. Two prompts on the same subject can need completely different answers depending on the objective behind them.
Before: "Tell me about email marketing."
After: "I want to increase the open rate on my weekly newsletter. Give me five subject line techniques I can test this month."
S: Style
Style is the writing approach: the format and voice you want the output modeled after. A listicle reads differently from a personal essay, and a legal summary reads differently from a casual explainer. Naming the style up front saves you a rewrite later.
Before: "Explain how compound interest works."
After: "Explain how compound interest works in the style of a friend explaining it over coffee, using one relatable analogy instead of formulas."
T: Tone
Tone is the emotional register: formal, playful, empathetic, urgent, and so on. Style and tone get mixed up a lot. Style is the shape of the writing, tone is how it feels to read. A prompt can ask for a listicle style in a warm, encouraging tone, or a listicle style in a blunt, no-nonsense tone, and get two very different results.
Before: "Write an email telling my client their project is delayed."
After: "Write an email telling my client their project is delayed by two weeks. Keep the tone apologetic but confident, not defensive."
A: Audience
Audience is who the output is actually for. A model has no idea if it's writing for a five year old, a room of engineers, or your CFO unless you tell it. The same information needs completely different vocabulary and depth depending on who's reading it.
Before: "Explain what an API is."
After: "Explain what an API is to a small business owner with no technical background, using an analogy from everyday life."
R: Response
Response is the format you want the answer delivered in: a table, a numbered list, a short paragraph, JSON, a script with dialogue and stage directions. This is the piece people skip most often, and it's the reason so many answers come back as a wall of text when a table would have taken five seconds to scan.
Before: "Give me ideas for improving our onboarding flow."
After: "Give me five ideas for improving our onboarding flow, formatted as a table with columns for the idea, expected impact, and effort to implement."
Putting It All Together
Here's what a full CO-STAR prompt looks like once all six pieces are in one place:
Context: I manage social media for a boutique coffee roastery that ships beans nationwide.
Objective: Write a launch post for our new single-origin Ethiopian roast.
Style: Short, punchy sentences, like a caption, not an article.
Tone: Warm and a little playful.
Audience: Coffee enthusiasts who already follow specialty roasters.
Response: A 60 to 80 word Instagram caption plus three relevant hashtags.
That's six sentences of setup, and it turns a vague request into something the model can actually execute well on the first try. Once you've written a few of these, the structure becomes second nature and you stop needing to type it all out from scratch every time.
Quick Tips for Using CO-STAR
- You don't need all six every time. Quick, low-stakes prompts can skip Style or Tone. Save the full framework for anything you'll actually use or publish.
- Write Objective first, even in your head. It shapes what Context is worth including.
- Keep a running note of CO-STAR prompts that worked well for recurring tasks (weekly reports, social captions, code reviews) so you're not rebuilding them from memory each time.
- If an output still misses the mark after all six are filled in, the fix is usually a fuzzy Objective, not a missing detail somewhere else.
Read more: CO-STAR Framework Explained, What Each Letter Means (With Examples)
Image Generation Prompt
Use this for the cover image (works well in Midjourney, DALL-E, or similar tools):
A clean, modern flat-design illustration representing the six letters C, O, S, T, A, R arranged as glowing building blocks stacked into a staircase shape, each block a different soft pastel color (blue, teal, purple, coral, yellow, green), set against a minimal light gray background, with a small icon floating above each block hinting at its meaning: a speech bubble for Context, a target for Objective, a paintbrush for Style, a mood face for Tone, a group of people for Audience, and a document for Response. Soft shadows, rounded corners, tech blog aesthetic, no text or letters rendered in the image itself, 16:9 aspect ratio.
Top comments (0)