Prompt Frameworks Explained: RTF, CO-STAR and Chain-of-Thought

Search "prompt framework" and you'll find dozens of acronyms, each presented as the secret to AI. The honest version: frameworks are just checklists that stop you forgetting parts of a good prompt. Three of them are genuinely worth knowing. The rest are the same ideas wearing new letters.

Here's what each one is, when it earns its keep, and when a plain sentence beats all of them.

RTF: Role, Task, Format

The minimal framework, and the one to reach for daily.

• Role: "You are a customer support lead…"
• Task: "…draft a reply to this complaint about a late delivery…"
• Format: "…under 100 words, apologetic but not grovelling, with one concrete next step."

Use it for: everyday work prompts — emails, summaries, rewrites, quick analyses. It covers 80% of situations at the cost of ten extra seconds.

Its gap: no slot for context or audience, which is exactly what's missing when RTF output feels generic. When that happens, upgrade to CO-STAR.

CO-STAR: The Full Checklist

CO-STAR won Singapore's government-run prompt engineering competition and became the corporate standard framework. The letters:

• C — Context: the background the model can't know
• O — Objective: what you want done
• S — Style: how it should be written (analytical, punchy, formal)
• T — Tone: the emotional register (warm, neutral, urgent)
• A — Audience: who will read it
• R — Response format: the shape of the output

Use it for: anything high-stakes or reusable — client deliverables, campaign copy, templates your team will run weekly. The Style/Tone/Audience trio is what separates it from RTF, and it's why CO-STAR output needs less editing.

Its gap: it's heavy. Filling six fields to ask for a paragraph rewrite is procrastination dressed as rigour.

Chain-of-Thought: Ask for the Reasoning

Different animal — not a template but a technique: ask the model to reason step by step before answering. "Work through this step by step, showing your reasoning, then give the recommendation."

Use it for: anything with logic in it — calculations, comparing options against criteria, finding the flaw in a plan, checking a contract clause. Asking for steps measurably reduces confident-nonsense answers, and it makes errors visible: you can see where the reasoning went wrong instead of just distrusting the conclusion.

Worth knowing: modern "reasoning" models do some of this internally, but explicitly requesting visible steps is still the best way to audit an answer you're going to act on.

Which One, When

• Quick daily task → RTF
• High-stakes or reusable output → CO-STAR
• Logic, analysis, decisions → Chain-of-Thought (combine it with either of the above)
• Simple factual question → no framework. "What does this error message mean?" needs no role play — frameworks add value in proportion to how much the output matters.

The Real Lesson Behind All Frameworks

Strip the acronyms and every framework is nudging you toward the same six ingredients: role, task, context, format, examples, constraints. We've unpacked them with before/after examples in The Anatomy of a Perfect AI Prompt — read that once and you'll never need to memorise an acronym again.

Or skip the memorising entirely: Cocoon's free Prompt Builder walks you through the ingredients interactively and assembles the prompt for you — RTF-fast for simple jobs, CO-STAR-thorough when it matters. And if your prompts still underdeliver, the five usual causes are in Why Your AI Prompts Don't Work.