Prompt Frameworks, Compared

Eight acronyms promise better prompts. Underneath they are the same seven ingredients in different orders. Here is the matrix, the one that fits each job, and the rules that keep work data safe.

8popular prompt acronyms
7elements they are all built from
0of them cover all seven
5rules for prompting safely at work

Prompt frameworks are memory aids, not magic words. A model does not recognise “RACE” or “SOLVE”; it responds to whether the prompt tells it who it is acting as, what situation it is in, what to do, within which limits, in what shape, with what example, and how success will be judged. Every acronym below picks a subset of those seven elements. Knowing the elements lets you use any framework — or none — and notice what a given acronym leaves out.

1 · The Eight, Decoded

FrameworkStands forBest forWeak at
RTFRole · Task · FormatQuick, well-understood requestsNo context, so answers stay generic
TAGTask · Action · GoalOutcome-driven work with a measurable targetNo role or output format
RACERole · Action · Context · ExpectationMost everyday professional tasksNo explicit constraints
CAREContext · Action · Result · ExampleOutput that must match a style or templateNo role or limits
RISERole · Input · Steps · ExpectationProcessing supplied material step by stepAssumes the input is trustworthy
PACTProblem · Approach · Compromise · TestDecisions with trade-offsNo output format
SOLVESituation · Objective · Limitations · Vision · ExecutionPlanning under constraintsLong to write for small tasks
DREAMDefine · Research · Execute · Analyse · MeasureMulti-stage projects and investigationsA project plan more than a prompt

Several acronyms circulate with more than one expansion (RISE is also given as Role · Input · Steps · Example). The expansions above are the most common ones.

2 · Seven Elements, Eight Subsets

Read down a column to see which frameworks remember an element; read across a row to see what each one forgets.

FrameworkLettersRoleContextTaskLimitsFormatExampleSuccess test
RTFRole · Task · Format✓·✓·✓··
TAGTask · Action · Goal··✓···✓
RACERole · Action · Context · Expectation✓✓✓·✓··
CAREContext · Action · Result · Example·✓✓·✓✓·
RISERole · Input · Steps · Expectation✓✓✓·✓··
PACTProblem · Approach · Compromise · Test·✓✓✓··✓
SOLVESituation · Objective · Limitations · Vision · Execution·✓✓✓··✓
DREAMDefine · Research · Execute · Analyse · Measure·✓✓···✓

▸ Two elements are the most often dropped and the most valuable: limits (what not to do, what to leave out, how long) and an example of good output. Adding either one to any framework usually improves the answer more than switching frameworks.

3 · One Security Task, Written Three Ways

The task: turn last quarter’s internal audit findings into a summary for the leadership team.

StylePromptWhat you get
Bare“Summarise these audit findings.”A generic list in the model’s own order and tone
RTF“You are a GRC lead. Summarise these audit findings for the leadership team as a table: finding, risk rating, owner, due date.”The right shape, but no priorities or business framing
All seven“You are a GRC lead (role). Our ISO 27001 surveillance audit is in March and the board meets next week (context). Summarise the attached findings (task) in under 200 words, without naming individuals, and flag anything you are unsure of rather than guessing (limits). Use a table: finding, risk, owner, due date, then three sentences on what needs a decision (format). Match the tone of this example: … (example). A good answer lets a director decide in two minutes which items need budget (success test).”A ranked, decision-ready summary in house style — and an explicit invitation to admit uncertainty

4 · Which One to Reach For

If the job is…Use
A quick, routine requestRTF
Most professional writing and analysisRACE, plus limits
Matching an existing style or templateCARE
Working through a document you supplyRISE — and treat the input as untrusted
A decision with trade-offsPACT
A plan under tight constraintsSOLVE
A multi-week investigationDREAM, one stage per prompt

5 · Five Rules for Prompting Safely at Work

  1. Use the approved tool. Confidential or personal data goes only into an enterprise AI service with no-training terms — never a personal account. See Shadow AI.
  2. Never paste secrets. Keys, passwords, tokens and connection strings do not belong in any prompt, even an approved one.
  3. Treat pasted content as untrusted. Emails, web pages and documents can carry hidden instructions (indirect prompt injection). Ask the model to summarise them, not to act on them.
  4. Ask for uncertainty. Tell the model to flag what it is unsure of and to cite its sources, then check the citations exist.
  5. A human signs it. AI output is a draft. Anything sent, published or acted on needs a person who has read it and is accountable for it.

Bottom line

  • Learn the seven elements, not the eight acronyms — role, context, task, limits, format, example, success test.
  • No framework covers everything; add limits and an example to whichever you use.
  • RACE is a sensible default for professional work.
  • Supplied input is untrusted input — summarise it, do not obey it.
  • Prompts are covered by policy — approved tools only, no secrets, a human signs the output.

Sources & method. Framework expansions as commonly published in prompt-engineering guides; variants exist. The seven-element matrix, worked example and recommendations are the author’s analysis, consistent with model providers’ public prompting guidance. Original analysis; no third-party graphics or text reproduced.

← Back to Cyber security