FREE LESSON · 5 MINUTES

COSTAR: write AI prompts that get a useful answer the first time

Most weak AI answers come from weak questions. COSTAR is a short checklist of six things to tell the AI before it starts. Learn it once, use it every day.

The idea in one minute

An AI tool does not know your situation. It does not know who will read the text, what you want to happen next, or how you like to sound. If you leave that out, it guesses, and you get a generic answer.

COSTAR gives you six prompts for yourself. Each letter is one thing to cover. You do not need all six every time, but running through them takes ten seconds and catches what is missing.

CContextThe background the AI cannot know: your situation, what has happened so far, any facts it must use.
OObjectiveThe one job you want done. Start with a verb: write, summarise, compare, list.
SStyleThe kind of writing you want: a short business email, a plain-language explainer, a bullet briefing.
TToneThe feeling it should carry: warm, firm, calm, upbeat, formal.
AAudienceWho will read it, and what they already know. A new customer and a colleague need different words.
RResponseThe shape of the answer: length, layout, language, anything to include or leave out.

A worked example

You work in customer service. A customer's order is late and you do not yet have a confirmed new date. You ask the AI for help.

QUICK PROMPT

Write a reply to a customer whose order is late.

The AI has to invent everything: the product, the reason, the date, the mood. You will rewrite most of it.

WITH COSTAR

Context: A customer ordered a standing desk on 3 October. It was due on 15 October and has not arrived. The warehouse has not given us a new delivery date yet. We do not know the reason for the delay.
Objective: Write a reply that apologises, says plainly what we know and what we do not, and promises a follow-up.
Style: Short, clear customer email. No jargon.
Tone: Sincere and calm. Take responsibility without over-apologising.
Audience: A private customer who is a little annoyed and wants a straight answer.
Response: Under 120 words, with a subject line. Do not promise a delivery date or offer compensation.

Same task, but now the AI knows the facts, the limits and the reader. The last line matters most: it stops the AI from inventing a delivery promise you cannot keep.

Always check before you send

COSTAR makes the first draft better. It does not make it correct. You still read the result, check the facts and decide what goes out. The person who sends it is responsible for it, not the AI.

Try it today

Pick one task you did with AI this week.
Write it again with the six letters, even if some are one short line.
Compare the two answers. Note which letter changed the result most for you.

WANT MORE THAN THE BASICS?

Go deeper with your own tasks.

The AI Act Academy is a self-paced app that explains the EU AI Act level by level, with quizzes. The first two lessons are free. For hands-on training with your team, working on your real tasks, get in touch and we will talk it through.

Try the first two lessons free Email hello@drihp.eu

What has changed since 2023

COSTAR dates from 2023, which is old in AI years. Here is what the newer sources say, in short.

The basics still hold. Anthropic's current prompting guide still starts from the same place: be explicit and clear, give the background, say what you want. That is what COSTAR asks you to do. Anthropic, prompt engineering best practices
People still study it, with mixed results. A 2025 study with university students found that after learning CO-STAR and two other frameworks, their prompts and literature reviews improved, but they often skipped parts of the framework. A 2025 paper tested a COSTAR variant with an extra "Answer" step and reported that plain COSTAR helps clarity with larger models but is less consistent with small local models. Islam et al., 2025 · Ohalete et al., 2025 (COSTAR-A)
The bigger idea is now "context engineering". In September 2025 Anthropic described it as the natural next step after prompt engineering: less about finding the right words, more about deciding everything the model gets to see, such as instructions, documents, tools and earlier messages. For everyday use this means COSTAR is still a good checklist for one message, but also paste in the real documents and facts the AI needs. Anthropic, effective context engineering for AI agents

Our take: treat COSTAR as a habit for the first draft, not a magic formula. The skill that lasts is knowing what the AI needs to know, and checking what it gives back.

Where COSTAR comes from

COSTAR (written CO-STAR by its authors) comes from the Data Science & AI team at GovTech Singapore, the Singapore government's technology agency. It appears in GovTech's Prompt Engineering Playbook and was made widely known by Sheila Teo, who won GovTech's first GPT-4 prompt engineering competition in 2023 and described the framework in her article. This lesson is our own explanation and example. DrIHP is not affiliated with GovTech.

Back to drihp.eu