Role, Goal, Context, Format: The Logic System for Better AI Prompts

Why Vague Prompts Produce Vague Work

A vague prompt produces vague work because the model has to guess at everything you left out: who is reading, what a good answer looks like, what actually happened this week. It guesses using the most average version of your request that exists, which is why so much AI writing sounds like it was assembled from press releases.

Watch what changes when the same request carries its four slots. Nothing about the tool changed here — only what went in.

Most professionals respond to the first result by reprompting: shorter, longer, “make it less formal”, “more specific”. That is guesswork answering guesswork, and it is how a two-minute task turns into twenty. The second result took about sixty seconds to set up because the writer already knew which four things to say.

A common mistake here is assuming the difference is length. It isn’t. A long, rambling paragraph of a prompt performs no better than a short one if it never names the reader or the deliverable. A good Role Goal Context Format prompt is often shorter than the mess it replaces. Structure beats volume every time.

The Role Goal Context Format Framework

The Role Goal Context Format framework is a four-slot structure for any AI request: Role sets whose standard the output must meet, Goal names the exact finished item you want, Context supplies the facts the model could not possibly know, and Format fixes the length, structure and boundaries of the reply. Four slots, one input, no memorised phrasing.

Each slot has a distinct job, and each one fails in its own recognisable way when you skip it.

Role sets the standard

Role is not costume-play. Writing “you are a world-class expert” adds nothing, because every model already tries to be that. What Role actually does is pick the vocabulary, the assumed knowledge and the bar the draft is measured against. “You are an operations manager writing for a finance director who cares about cost variance” changes every sentence that follows.

The sharpest version of Role names two things: who is writing and who is reading. Most guides only mention the first, which is why their examples still read generically. If you want the mechanics in depth, the guide to writing a Role that lifts output quality takes it apart line by line.

Goal names the deliverable

There is a real difference between a topic and a deliverable. “Something about the budget overrun” is a topic; “a 150-word note to the steering group explaining the overrun and the two options for covering it” is a deliverable. The model can build the second and can only guess at the first.

Use a verb that describes the finished artefact — draft, summarise, compare, rank, rewrite — and then say what the artefact has to achieve. That combination is covered further in defining a Goal that produces usable output.

Context carries the facts

Context is the slot people skip, and it is the one doing most of the work. It holds the things the model has no way of knowing: your numbers, your notes, last week’s decision, the sensitivity you have to write around. In practice, pasting the raw source beats describing it.

Thin context is also where invented detail comes from — an empty slot gets filled with something plausible rather than something true, which is the whole mechanism behind AI hallucination in professional work. One instruction closes most of that gap: work only from the text above, and write “not supplied” instead of filling in the blank. There is more on layering background, constraints and specifics in adding Context that makes output specific to your situation.

Format decides the shape

Format is the slot that decides whether you paste the reply straight into an email or spend ten minutes reshaping it. Length as a number, the sections you want, the tone, and — the part most people forget — what to leave out. “No preamble, no closing summary” alone removes about two paragraphs of throat-clearing from most replies. Vague length instructions are the usual culprit: “keep it brief” means nothing, while “under 180 words” is checkable.

Current guidance from the major labs leans toward telling the model what you want rather than listing what you don’t — “write in flowing paragraphs” outperforms “don’t use bullet points”. Banned-word lists still earn their place as a second pass, which is the subject of using negative constraints properly, and the full set of output shapes is in specifying Format so output is immediately usable.

Build Your First Four-Slot Prompt

Here is the build sequence for a Role Goal Context Format prompt, using a weekly status update as the example. It takes about five minutes the first time and roughly a minute once the template lives in a note somewhere.

Write who is producing this and who receives it. Two clauses is enough. “You are a project lead writing for an operations director who reads on their phone between meetings” already rules out half of what the model would otherwise produce.

Say what you will actually send, not what the topic is. A 200-word email. A five-row risk table. A three-bullet summary for a steering pack. If you cannot describe the artefact in a sentence, the model has no chance either.

Drop in your raw notes, the numbers, the thread you are replying to. Messy is fine — sorting mess is the one thing these tools are genuinely good at. Then add the grounding line that stops invention: work only from the text above.

Length as a number, the sections you want in order, the tone, and the things you never want to see. Put this last so it is the most recent instruction before the model starts writing.

ROLE
You are a [your job title] writing for [who reads this], who mainly cares about [what they judge it on].

GOAL
Produce a [exact deliverable] that [what it has to achieve].

CONTEXT
Work from this and nothing else:
[paste your notes, data, or the source document]
Also true: [deadlines, sensitivities, decisions already made]
If something needed is missing, write "not supplied" rather than filling it in.

FORMAT
[number] words maximum.
Structure: [the sections you want, in order].
Plain professional English. No preamble, no closing summary.
Leave out: [whatever you never want to see].

Keep the filled version somewhere you can reach in two seconds — a pinned note, a document, a saved instruction inside the tool. If you would rather start from something pre-built, the free AI Work Templates library has slot-filled starters for the usual office tasks.

Does slot order matter?

Mostly no, and the acronym is a memory aid rather than a syntax rule. Models parse labelled blocks in any sequence. There is one documented exception worth knowing: when you are pasting something long, put the document first and your question after it. Anthropic’s prompting best practices documentation reports that placing the query at the end can improve response quality by up to 30 percent in testing, particularly with complex, multi-document inputs.

For shorter requests, OpenAI’s prompt engineering guide suggests a familiar running order — identity, then instructions, then examples, with additional context usually best positioned near the end. Both amount to the same practical advice: label your blocks clearly and stop worrying about the sequence.

Three Rebuilds From Real Work

Three Role Goal Context Format examples, taken from tasks that eat professional afternoons. Each one is the same four slots wearing different clothes.

The Monday status update

The task most likely to be written badly under time pressure, because the writer knows everything and the reader knows nothing. Role fixes the asymmetry.

ROLE
You are a project lead writing for a director who reads this on a phone and only wants risks and decisions.

GOAL
Draft a weekly status email that makes clear what moved, what is at risk, and what needs a decision from the reader.

CONTEXT
Work only from these notes:
[paste your week's rough notes]
Do not invent progress, dates or names that are not in the notes above.

FORMAT
Under 180 words. Subject line, then three short sections: Moved / At risk / Needs you.
No pleasantries, no closing summary. Plain sentences, no adjectives about how hard the team worked.

The feedback you dread writing

Vague feedback is a Role and Context failure at the same time — no standard to write to, and no evidence to write from. What changes is not the tone but the specificity.

ROLE
You are an experienced manager writing feedback that a colleague will read once and act on. Direct, specific, not harsh.

GOAL
Draft feedback covering one strength and one development area, ending with two concrete habits for next quarter.

CONTEXT
Work only from this evidence:
[paste dates, deliverables, and what actually happened]
Do not generalise beyond it. If a claim is not supported above, leave it out.

FORMAT
Under 200 words. Three short paragraphs: what went well, the pattern to change, what to do differently.
Describe behaviour and effect, never personality. No praise sandwich, no motivational closing line.

The spreadsheet nobody reads

Data work is where Format earns its keep. The analysis is rarely the hard part; turning it into something a non-financial reader will act on is.

ROLE
You are an analyst writing for board members who do not read spreadsheets and need the point in thirty seconds.

GOAL
Turn the figures below into a written summary that says what changed, why it matters, and what to do about it.

CONTEXT
Work only from this data:
[paste the rows, or attach the file]
Focus on lines with more than [X] percent variance. State "not supplied" for anything you cannot calculate from the data.

FORMAT
Three key findings as bullets, then a short table of the largest variances, then two recommended actions.
No jargon. Every number you cite must appear in the data above.

What these three share is that none of them required a clever phrase. They required knowing which four things to state — and the rest is swapping the bracketed parts.

The Blank-Slot Test for Weak Output

The Blank-Slot Test is the diagnostic half of this system: when a draft disappoints, do not rewrite the prompt — name the one Role Goal Context Format slot you left empty and fill only that. Weak output almost always traces back to a single missing input, and the symptom tells you which one.

Fixing one slot beats rewriting the whole prompt, because a rewrite changes several variables at once and you learn nothing about which one mattered. This is the same logic as the diagnostic method for refining AI output, applied at the input stage instead of the editing stage.

What most guides get wrong is treating a disappointing draft as evidence that the model is weak. In practice it is evidence that one slot was empty, and the slot is usually Context.

What Not to Put in Context

Context is the slot of a Role Goal Context Format prompt that carries real company information, which makes it the one with genuine risk attached. The line that matters is not free plan versus paid plan — it is a consumer account versus a business or enterprise account. Consumer plans generally train on your conversations by default with an opt-out you have to find and switch off; business and enterprise accounts generally do not train on your data by default.

Anonymising costs almost nothing here. Swap real names for “the client” and exact figures for rounded ones — the structure of the output barely changes, and the version you paste back in is yours to correct. For the fuller treatment of policies, retention and what your employer can see, read the practical privacy guide for using AI at work.

What Four Slots Cannot Fix

This system fixes input quality. It does not turn a language model into a colleague, and pretending otherwise is how people end up embarrassed in front of a director.

There is also an honest ceiling on time savings. A structured prompt takes longer to type than a lazy one; the saving arrives downstream, in the editing you no longer do and the reprompting rounds you skip. On a task you repeat weekly, that trade pays back quickly. On a one-off you will never do again, sometimes it does not, and writing it yourself is the right call.

Start with one task — the recurring one you resent most — and build its four slots properly this week. Once that prompt exists, next week is a swap rather than a rebuild, and the same Role Goal Context Format slots port to the next task without any relearning.