ChatGPT Prompt Frameworks: 10 That Hold Up in 2026 (And Why I Deleted Most of My Old List)

In March 2024 I published a list of 13 ChatGPT prompt frameworks on this blog. Search Console shows 182 clicks for it over the last 16 months, spread across five old and new URLs, so people found it and presumably used it.

For this rewrite, every framework on that list got checked against its source. Only three of the 13 trace back to a named author, and I had all three wrong. The real RISEN is Role, Instructions, Steps, End goal, Narrowing. Mine pushed those into brackets behind words like "Requirement" and "Negotiation" and lost End goal completely. The real CREATE starts with Character. Mine started with "Conceptualize" and turned into a project plan. CLEAR had the same problem. For most of the others no author turned up at all.

My favorite is GATHER, which told you to run a SWOT analysis and write a timeline with responsibilities before you typed anything. I doubt anyone ever did that to get a blog post out of ChatGPT.

So I deleted most of it. This version keeps the ChatGPT prompt frameworks that have a traceable source or a clear job. It says when to use each one and what changed since 2024. One-person agency here, and Claude is the AI tool I work in every day, so the ChatGPT-specific details below come from OpenAI's documentation, and I've linked them.

What a Prompt Framework Does

This was the example in the 2024 post, and it still describes most prompts:

"Hello, please create a 1000 word blog post about digital marketing for me, thanks!"

That gives ChatGPT a length and a topic. It says nothing about the reader, the point of the post, what you know that the rest of the internet doesn't, or what to avoid. The model fills those gaps with the most likely text, which is how you end up with "in a world of..." and "immerse yourself". The same averaging is why AI-built websites all look alike.

A prompt framework is a checklist for those gaps. Each letter of the acronym reminds you to supply something you would otherwise skip. The model never sees the acronym and gets no boost from it, so the only question is which checklist makes you fill in the right things.

Anthropic's documentation has a better test than any acronym: show your prompt to a colleague with little context on the task and ask them to follow it. "If they'd be confused, Claude will be too." Swap in ChatGPT and the advice holds.

The 10 Prompt Frameworks Worth Knowing

Ten frameworks, roughly from shortest to longest, with the source where I could find one. CLEAR sits at the bottom because it works differently from the rest.

Framework	Stands for	Source	Use it for
RTF	Role, Task, Format	No clear author	Quick one-off requests
Google's four parts	Persona, Task, Context, Format	Google Workspace	Everyday office tasks
CARE	Context, Ask, Rules, Examples	Nielsen Norman Group	Tasks with firm rules
RACE	Role, Action, Context, Expectation	No clear author	Tasks where you know what good looks like
RODES	Role, Objective, Details, Examples, Sense check	Credited to Sebo	Briefs you want confirmed first
RISEN	Role, Instructions, Steps, End goal, Narrowing	Kyle Balmer	Work with a fixed order
CRISPE	Capacity and Role, Insight, Statement, Personality, Experiment	Credited to Matt Nigh	Getting several options
CO-STAR	Context, Objective, Style, Tone, Audience, Response	GovTech Singapore	Writing for a specific reader
CREATE	Character, Request, Examples, Adjustments, Type of output, Extras	Dave Birss	Longer creative briefs
CLEAR	Concise, Logical, Explicit, Adaptive, Reflective	Leo Lo, 2023	Reviewing a prompt you wrote

Quick tasks: RTF and Google's four parts

RTF is three slots, and for a quick request that's usually enough:

You're a copywriter at a small web agency.
Write a 150-word reply to a client who asked for a discount.
Plain email, no bullet points.

Google teaches almost the same thing for Gemini and adds context as a fourth part. Its guide says not every prompt has to follow that structure, which is the most useful sentence in it.

Firm rules: CARE and RODES

CARE comes from Kate Moran at Nielsen Norman Group, and the letter doing the most work is E. Anthropic calls examples "one of the most reliable ways" to steer format and tone, and suggests three to five. OpenAI's guide for reasoning models says those models often don't need examples and recommends trying without them first. Those two fit together: start without, and add one when the format keeps coming out wrong.

RODES ends with a sense check, where you ask the model to confirm it understood the task before it writes. That mattered more in 2024. OpenAI says its GPT-6 Astra is more likely to ask for clarification on its own.

Work with a fixed order: RACE and RISEN

RACE is RTF with one useful addition: Expectation, where you say what a good result looks like. RISEN goes further with Steps, an End goal and Narrowing, which is where your constraints go.

Steps is the slot that has aged worst, for reasons I'll get to below. Use it when the order matters to you, like "check the facts first, then rewrite", and leave it out when you're only guessing how the model should work.

Options: CRISPE

CRISPE ends with Experiment, which means asking for several answers. I'd keep that one idea and forget the rest: "give me four headline angles and tell me which is weakest" is a better prompt than most acronyms will get you.

Writing for a specific reader: CO-STAR and CREATE

GovTech Singapore says its officers use CO-STAR, and it's the one that fits writing best because Audience and Response get their own slots. Here is the 2024 blog post request again:

Context: I run a one-person web agency. My readers own small
businesses and have been burned by a marketing agency before.
Objective: Write a 1,000-word post on how to tell if an agency's
monthly report means anything.
Style: Plain explanation with one worked example.
Tone: Direct and a bit dry.
Audience: Business owners who don't know SEO terms.
Response: Markdown with H2 headings. Start with the problem.

I didn't run this one through ChatGPT, it's here to show the shape. The output still depends on what you feed it, and your own numbers and stories matter more than any slot. As I wrote in my AI copywriting trends post, the brief matters more than the model.

CREATE covers similar ground with more room for examples and follow-up tweaks. It's also the one my old post mangled the worst.

Checking your own prompt: CLEAR

Leo Lo's CLEAR comes from a 2023 paper in the Journal of Academic Librarianship. Its letters are qualities of a good prompt, so you use it after you've written something: cut what isn't needed, then adapt and reflect once you've seen the answer. It pairs with any of the others.

Which ChatGPT Prompt Framework Should You Use?

I looked for a controlled test comparing these acronyms on current models and didn't find one. The closest I got was a 2025 paper on a CO-STAR extension, tested on small models of up to 8 billion parameters, where the effect varied by model and use case. The ranking lists I opened that say they tested frameworks showed me no scores and no method.

So "best" is an opinion, including mine. The Wharton team behind the Prompting Science Reports found that popular tactics help on some questions and hurt on others, and that's the most defensible thing anyone has said on this.

My pick: RTF for anything quick, CO-STAR for anything a customer will read. For anything you'll run every week, skip the frameworks and write a saved brief, which I'll come to.

What Changed Since 2024

In April 2026, The Decoder reported that OpenAI's GPT-5.5 guide told developers to stop carrying over old prompts and start from the smallest one that does the job. Many framework lists were written before that.

Process-heavy prompts lost ground. The same guide says over-specified process adds noise and can make answers sound mechanical. It recommends stating the outcome and what success looks like, with any constraints. Its suggested structure runs role, personality, goal, success criteria, constraints, output and stop rules, which is a framework with no acronym.

"Think step by step" stopped paying. OpenAI's reasoning guide advises against it, because those models reason internally. Wharton's second report found only marginal gains on reasoning models, at a real cost in time and tokens.

"You are a physics expert" doesn't add knowledge. Wharton's fourth report tested expert personas across six models on graduate-level questions, and they generally didn't improve accuracy. Personas may still change tone, and OpenAI put a short role statement back at the top of its structure. So keep the Role slot for voice and framing, and skip the flattery.

Conflicting instructions became the bigger risk. OpenAI's GPT-6 documentation says the model follows long instructions better but reacts more to what's in its context, and unclear or conflicting guidance can make it stop early. A framework you fill in on autopilot is a good way to produce that conflict.

Put It in a File

The complaint that opened my 2024 post, that AI had ruined phrases like "immerse yourself", is now a file. It's called anti-slop.md, it runs to about 1,000 words of banned phrases and sentence shapes, and it applies to everything written in my voice, including this post.

I'm in decent company. OpenAI's GPT-6 guide ships a sample prompt telling the model to avoid "slop words" like "foster" and "importantly", plus the "This isn't about X. It's about Y." move.

A framework helps you write one good prompt. A saved brief means you stop rewriting it: who you are, who reads your work, what good looks like, what's banned, with two real examples. In ChatGPT that belongs in custom instructions or a Project. In my setup it's a folder of plain files. None of them uses an acronym.

What Happened to the Other Frameworks From 2024

If you came here looking for one of the old entries:

SMART is a goal-setting acronym that I bent into a prompt checklist. PACT, GUIDE, QUEST, FOCUS and IDEA have no source I can find, and they ask for what the table above already covers. SPARK circulates with different letters depending on who you read. RHODES, CLIMB and GATHER were project plans with a prompt label on them, and the version of RHODES people use is RODES, which is in the table.

I can't tell you which acronym will work best on your task, and neither can the lists ranking for this keyword. Take one real job, write the plain version and the framework version, and compare what comes back.

If your team is past single prompts and wants this built into how it works, that's what my AI consulting is for. And if you're not sure you need that, I wrote a flowchart for exactly that question.