The CREATE Prompt Framework

TL;DR: CREATE is Character, Request, Examples, Adjustments, Type of Output, and Extras — six fields built for a longer creative brief, not a quick question. It traces to Dave Birss's 2023 LinkedIn Learning course, and his own companion guide names the fourth field two different ways on two different pages. The framework's real edge over its size-matched rivals is a dedicated Examples field neither AUTOMAT nor CO-STAR has.

What Is the CREATE Prompt Framework?

CREATE is a six-part prompt structure: Character, Request, Examples, Adjustments, Type of Output, and Extras. You set the persona, state the task, show a sample of what good looks like, fine-tune the result, specify the format, and add any special phrasing the task needs.

Character: [the persona or expertise you want the model to answer from]
Request: [the specific task, stated as an instruction]
Examples: [one or two samples of the style or content you want matched]
Adjustments: [tweaks that steer away from a likely wrong turn]
Type of Output: [the exact format — length, structure, medium]
Extras: [special phrases: ask it to explain its thinking, ask it to ask you questions first]

That's a genuinely larger commitment than most frameworks in this cluster. TAG, the companion post to this one, covers the opposite end of the same spectrum: three fields, no persona, no format instruction, built for a request you'll throw away in five minutes. CREATE is built for the opposite case: a brief detailed enough to matter, usually because someone other than you will read the output, or because you're going to run something close to this same request again.

Who Created the CREATE Framework, and Does Its Own Source Agree With Itself?

Dave Birss is the earliest verifiable source. His companion guide, The Prompt Collection, is distributed as a free PDF from his 2023 LinkedIn Learning course, titled How to research and write using Generative AI tools. The PDF's own cover names its source directly: "From the LinkedIn Learning course", right before opening into the CREATE breakdown as its first chapter. NYU's SPS Nexus blog, in a post dated 27 September 2023, independently cites the same origin when it walks through the framework: "(Birss, 2023)", with a reference list entry reading "Birss, D. (2023)." and naming the same LinkedIn Learning course by title. That's two independent confirmations of the same year and the same source, which is about as solid as attribution gets in this cluster.

Here's where it gets genuinely interesting, and where checking the primary source paid off. Birss's own guide is not internally consistent about what the fourth letter stands for. The overview page lists the six components as:

"Character Request Examples Additions Type of output Extras"

Two pages later, the section devoted to that same fourth component is headed:

"ADJUSTMENTS"

Both phrases come from the same 14-page section of the same PDF. We're not picking a winner between the overview's "Additions" and the section header's "ADJUSTMENTS". They mean close to the same thing in context, roughly, small course-corrections to an output, but it's worth knowing that the framework's own primary source doesn't spell its own name the same way twice, and neither of the two secondary sources we checked mentions the discrepancy.

How Do You Build a CREATE Prompt, Field by Field?

Here's a request built with all six fields, followed by what happens when you swap just one of them.

Full CREATE prompt:

Character: You are a B2B sales leader with 15 years of experience, writing
in plain language with no corporate jargon.
Request: Write a LinkedIn post arguing that cold email still works in 2026.
Examples: [paste one of your own past posts that nails your voice]
Adjustments: Avoid the phrase "game changer" and anything that sounds like
a listicle headline. Open with a specific number, not a question.
Type of Output: LinkedIn post format — short paragraphs, no bullet lists,
under 200 words.
Extras: End with one specific, non-generic call to action, not "thoughts?"

Birss's guide describes the Request field this way: "This is the task you want ChatGPT to do for you. You clearer you are, the higher your chance of getting a great response" — a sentiment every framework in this cluster shares in some form, even if the wording above reads a little rough in the original.

Now swap only the Examples field to a completely different sample post, one written in a punchier, joke-heavy voice instead of the plain-spoken original. Character, Request, Adjustments, Type of Output, and Extras all stay identical. The output shifts noticeably: shorter sentences, more rhetorical questions, a different rhythm entirely, because the model is now pattern-matching against a different anchor. That's Examples doing the thing none of the other five fields can: describing a voice by showing it instead of naming adjectives for it, which is the same principle behind few-shot prompting generally.

Extras deserves a closer look too, because it's the field people skip most confidently, assuming it's decorative filler. Birss's own guide devotes an entire chapter to phrases like these, and all four still work on current models. "IGNORE EVERYTHING BEFORE THIS PROMPT." resets context mid-conversation instead of forcing you to open a new chat. "ASK ME QUESTIONS BEFORE YOU ANSWER." works when you're not sure what information the model actually needs, and lets it ask rather than guess wrong. "EXPLAIN YOUR THINKING." surfaces the reasoning behind an answer, which is useful for anything you'll need to justify to someone else afterward. "ACT UNLIKE A TYPICAL AI." pushes the model past its most obvious, most average response toward something less generic. None of these four phrases is unique to CREATE. You can drop any one of them into a bare prompt with no framework at all. What CREATE actually does is give them a named home in the structure, so you remember to reach for one instead of quietly accepting the model's first, most predictable answer.

When Does CREATE Actually Earn Its Six Fields?

When the brief is detailed enough that the setup cost is worth paying once, usually one of two situations:

• Someone other than you reads the output, so tone, voice, and format all genuinely matter and a wrong guess on any of them is expensive to fix after the fact.
• You'll run something close to this same request again, so the fixed cost of six fields amortizes across many uses instead of one — a recurring content template, a support-reply generator, a report you regenerate monthly.

If neither is true, CREATE is very likely more setup than the task deserves. Birss's own guide backs this: it markets the framework for prompts you "brief ChatGPT properly", which is a phrase that implies stakes, not a quick lookup.

Worth being concrete about what "amortizes" actually means here. A support team that writes CREATE prompts fresh for every ticket type is paying the six-field setup cost dozens of times a week for marginal benefit, because most tickets don't need a persona or a style sample rebuilt from scratch. The same team writing one CREATE prompt per ticket category, saved once with the category's own Character and Examples filled in, pays the setup cost a handful of times total and reuses it indefinitely. The framework doesn't change; what changes is whether you're filling in six fields once per category or once per ticket. That distinction is the whole argument for CREATE over something smaller, and it disappears completely the moment you're only ever going to run the request once.

When Is CREATE Overkill?

For anything fast, low-stakes, or genuinely single-shot. Six fields cost real time to fill in, and on a throwaway question that cost buys you nothing — you'd get a comparably useful answer from TAG's three fields in a fraction of the setup.

The most commonly skipped field in practice is Examples, and it's the one that costs you the most when skipped. Leave it blank because you don't have a sample handy, and CREATE quietly becomes a five-field framework mostly duplicating what RTF's Role and Format already cover, minus CREATE's actual differentiator.

How Does CREATE Compare to AUTOMAT, CO-STAR, and TAG?

CREATE, AUTOMAT, and CO-STAR are the three largest frameworks in this cluster, and each spends its size on something different.

The honest read: pick CREATE over AUTOMAT when you have a real sample to anchor the voice and don't particularly care about naming edge cases. Pick AUTOMAT over CREATE when the prompt is going into production and needs to handle malformed input or out-of-scope requests gracefully — CREATE has no field for that at all. Pick CO-STAR over either when the audience reading the output is the variable that matters most; CREATE's Character field sets who's answering, not who's reading, and that's a real gap if the two are different people. For the full seven-framework lineup including CRAFT and Chain-of-Thought, see our roundup of ChatGPT prompt frameworks.

Can You Combine CREATE With Other Techniques?

Yes, and the cleanest pairing doesn't add a second framework at all — it nests a technique inside the Request field. For anything that needs reasoning before the creative work starts, add a chain-of-thought instruction inside Request: "Think through three possible angles for this post, pick the strongest, then write it." CREATE still owns Character, Examples, Adjustments, format, and Extras; the reasoning step just happens inside the task itself.

What doesn't work well is stacking CREATE with a second full framework — AUTOMAT's edge-case handling bolted onto CREATE's six fields, for instance. That's roughly thirteen instructions in one prompt, which is more than most models reliably track without dropping a few. If you genuinely need AUTOMAT's scope and edge-case coverage on top of CREATE's Examples field, the better move is picking the one framework that actually has both, or accepting that you'll iterate the prompt rather than front-load every requirement into a single first attempt.