COSTAR Framework

COSTAR is a structured prompt-writing framework whose six letters — Context, Objective, Style, Tone, Audience, Response format — form a checklist ensuring a prompt tells the model everything it needs to produce the output you actually want.

Most bad prompts fail by omission: the writer knew the background, the audience, and the desired shape of the answer, but never said so, leaving the model to guess. COSTAR attacks that failure mode directly. It gained prominence after winning Singapore's GPT-4 prompt engineering competition and has since become a standard teaching scaffold because it is memorable and covers the gaps people most often leave.

How it works

Walk the six letters, in any order, and fill each in:

• Context — the background the model cannot infer: what happened, what data it is looking at, what constraints exist.
• Objective — the single task, stated precisely. One prompt, one job.
• Style — the writing style: analytical, punchy, formal, like a runbook, like a tweet.
• Tone — the emotional register: neutral, urgent, warm, apologetic.
• Audience — who reads the output, and what they already know. "For a CFO" and "for a junior engineer" produce different correct answers to the same question.
• Response format — the exact output shape: bullet count, word limits, JSON fields, sections. This is the line that makes output parseable and reviewable.

Not every element earns a sentence in every prompt — for machine-to-machine prompts, Style and Tone often collapse to nothing while Response format does heavy lifting. The checklist's value is forcing the decision consciously instead of by accident.

Why it matters

COSTAR converts prompt-writing from inspiration into procedure — which is exactly what beginners and teams need. A checklist makes prompts reviewable ("you specified no audience — who is this for?"), makes quality reproducible across a team, and shortens the gap between a first draft and a production-grade prompt. It will not make anyone a great prompt engineer by itself; it reliably stops them from being a careless one.

A worked example

Vague prompt: "Write something about the outage."

Through COSTAR:

Context: Checkout was down 47 minutes on May 3 due to an expired

TLS certificate; fix was renewal plus automated rotation.

Objective: Write the customer-facing incident summary.

Style: Plain language, no jargon, no internal system names.

Tone: Accountable and calm — no defensiveness, no drama.

Audience: Merchants who lost sales during the window.

Response: Three short paragraphs — what happened, impact,

prevention — under 150 words total.

The vague version produces a coin-flip; the COSTAR version produces the document you would actually ship, on the first try, from any competent model.

How miatz teaches it

COSTAR is the first framework in the miatz prompting module — learners take one deliberately vague prompt and rebuild it letter by letter, comparing outputs at each step so every element's contribution is visible. It then becomes a habit enforced by repetition: prompting drills in the daily training loop are graded partly on whether all six decisions were made deliberately.