Four prompt frameworks do most of the work, and AI Central's How To Craft The Perfect Prompt Every Time lays them out in ascending order of control. RTF sets role, task and format. RACE adds the circumstances and the expected outcome. RISEN hands the model an ordered plan and a measurable end goal. RODES adds benchmarks and a sense check. Pick by how much of the thinking you are willing to leave to the model.
Reviewed August 2026. Source material: How To Craft The Perfect Prompt Every Time, an AI Central guide built around GPT-5.
The gap these frameworks close
A weak prompt is rarely a badly written sentence. It is an under-specified one. You ask for a LinkedIn post, and the model has to guess the vantage point, the shape of the output, the audience, and what finished looks like. It guesses reasonably. That is the trap, because reasonable and usable are not the same thing.
AI Central treats this as a structure problem rather than a wording problem. The guide opens by saying that writing effective prompts does not have to be complicated, then spends its length on four named frameworks, each a set of slots you fill in. The slots are the point. They force you to say out loud what you were otherwise leaving to chance.
RTF, the one to default to
RTF stands for Role, Task and Format, and AI Central calls it a classic. Three slots, and the guide is specific about what belongs in each.
- Role, which AI Central says should be your current role.
- Task, what you want the model to accomplish.
- Format, the final output you want.
There is a tension inside that, and it is instructive. The written guidance says the role should be your current role, while AI Central's own worked example assigns the model the role of a seasoned LinkedIn social media expert. Those are not the same instruction. Naming your own role tells the model who it is writing for and what you already know, so it stops explaining the basics. Naming an expert role tells it what standard to write to. The slot is doing double duty, and the strongest prompts fill it with both.
The format slot is where that example earns its keep. AI Central does not ask for a good post. It specifies a punchy hook of at most nine words, a counter-hook of at most nine words that has to start with the word But, a pivot line beginning Instead of, exactly five steps, and a closing practical tip. Word ceilings, a mandatory opening word, a fixed number of steps. That is a template rather than a request, and it is the difference between output you publish and output you rewrite.
RACE, when the answer depends on circumstances
RACE keeps the role, renames the task as an action, and adds two slots. AI Central sums up the upgrade in a line.
RACE adds depth to RTF by including context and expectations. Ideal for detailed, scenario-specific prompts.
The worked example is a market research brief. The action, analyse emerging trends in renewable energy, would on its own produce an encyclopedia entry. The context slot is what rescues it: solar and wind specifically, the last five years specifically, with technological advancement and regulatory change named as the forces that matter. The expectation slot then names the artefact and its reader, a detailed report carrying market forecasts, competitor analysis and investment recommendations, written to guide strategic decisions at a renewable energy investment firm.
Expectation is the slot most people skip, and it changes the output more than any other. Telling the model that an investment firm will act on the result rules out hedged, survey-style prose. Context narrows what gets considered. Expectation decides who the writing is answerable to.
RISEN, when the order of the work matters
RISEN is for jobs with a sequence: Role, Instructions, Steps, End Goal, with a narrowing clause at the close. AI Central's summary is blunt about who is in charge.
RISEN ensures AI follows your roadmap.
Note whose roadmap. In the worked example, a social media manager at a health and wellness brand does not ask the model how to grow the account. The steps are supplied: audit current performance and audience insights, build a strategy around educational and motivational posts, schedule daily posts plus bi-weekly stories or live question and answer sessions, run influencer cross-promotion, launch a monthly hashtag challenge for user-generated content, then refine the whole thing against engagement metrics.
Two details carry the whole framework. The end goal is numeric, thirty percent follower growth and a doubled engagement rate within three months, which makes the result checkable rather than merely plausible. The narrowing clause then cuts the surface area, restricting the work to the platforms where the target demographic actually is, named in the example as Instagram and Pinterest. Skip the narrowing and the model spreads effort evenly across every channel it can think of, which is how you end up with a plan nobody can execute.
RODES, when you want it benchmarked and stress tested
RODES is the longest of the four: Role, Objective, Details, Examples, Sense Check. AI Central's verdict on it is one line.
RODES leverages benchmarks for actionable strategies.
AI Central's example puts a user experience designer on an e-commerce site with a conversion objective, then names the changes in play rather than leaving them open: streamline the checkout process, optimise product search, improve mobile responsiveness. The examples slot points at Amazon, and not at Amazon in general, at its checkout flow and its product recommendation system. That is what a benchmark is for, importing a standard the work can be measured against.
The sense check is the slot none of the others have, and the most interesting idea in the set. It tells the model to test its own proposals against user behaviour patterns visible in site analytics, and against implementation timeline and budget. You are asking for the plan and the objection to the plan in a single pass. Most teams run that second pass a week later, in a meeting where somebody finally mentions the budget.
The pattern underneath all four
Read them in sequence and they are one idea at four settings. Every framework opens with a role and closes with a description of the finished thing. What changes in between is how much of the reasoning you hand over.
- RTF fills three slots and leaves the circumstances to inference.
- RACE adds the circumstances and names who the output is answerable to.
- RISEN adds your sequence and a number the result has to hit.
- RODES adds an external benchmark and a built-in critique.
None of the four contains any subject matter. They are containers, which is why the same set survives market research, social media management and interface design unmodified. That gives you a working rule: escalate one step whenever the model hands back something reasonable that you cannot actually use.
What to do with this
- Default to RTF for routine work, and spend your effort on the format slot, word limits and section counts included.
- Move to RACE the moment the right answer depends on circumstances, and always name who will read the output.
- Use RISEN when you already know the sequence, and give the end goal a number and a deadline.
- Use RODES for high-stakes work, name a specific benchmark rather than an admired company, and keep the sense check even when the plan looks finished.
- Save your filled-in versions. A framework you rebuild from scratch every time is a framework you will quietly stop using.
AI Central closes on the division of labour, which is the honest version of every prompt guide ever written.
GPT5 is a Tool.
The rest of that line finishes the thought, the prompts are the engine. The model supplies capability. Structure decides what comes out of it, and structure is the part you control.
Which framework should I use if I only learn one?
RTF. It covers most everyday requests, takes seconds to fill in, and trains the habit the other three depend on, stating the finished shape before you ask for anything. When it stops being enough, the other three are extensions of it rather than replacements.
Do these frameworks only work with GPT-5?
AI Central built the guide around GPT-5, but nothing in the four structures depends on a GPT-5 feature. They are ordinary sentences arranged under labelled headings, so they paste into any assistant unchanged. What varies between models is how much you get away with leaving out.
What is the actual difference between RACE and RISEN?
RACE describes the situation and the outcome and leaves the method to the model. RISEN supplies the method as ordered steps and adds a measurable target. Choose RACE when you want the model's approach, choose RISEN when you already have one and need it executed in order.
My prompts are getting long. Is that a problem?
Length is not the metric, specificity is. A long prompt full of vague slots is worse than a short precise one. The test is simple: if the output is already usable, stop escalating. If it is generic, the missing piece is almost always context, expectation or a number.