The RACE Prompt Framework: Real Origin, Four Parts, 14 Prompts

TL;DR: RACE is a four-slot prompt structure: Role, Action, Context, Execute. It is published by Trust Insights, a Boston analytics firm, with a dated trail back to July 2023. Half the internet writes the E as Expectation instead. It is not the Smart Insights RACE marketing model, which is a different thing entirely.

What is the RACE framework in prompt engineering?

Four labelled lines, in order: who the model is, what it does, what it needs to know, and what the finished thing should look like.

That is the whole framework. Here it is applied to a real task:

[R] You are a senior technical support lead at a B2B analytics company.
[A] Draft a reply to a customer asking for a refund 46 days after purchase.
[C] Our refund window is 30 days. This account has been active for 14 months
    on the annual plan and has filed no prior tickets. The customer's message
    says the tool "never got used after the team reorg". Renewal is in 3 weeks.
[E] Under 150 words. No apology in the opening line. Decline the refund, state
    the policy once without quoting it at length, and offer a plan downgrade
    at renewal as the alternative. Plain sentences, no bullet points.

Strip any of those four lines and the output gets measurably worse in a predictable way. Remove Role and you get a generic register. Remove Context and it invents the policy. Remove Execute and it writes 400 words with a bulleted list and three apologies.

The framework is popular because four is the number of things most people can hold in their head while typing. That is a real advantage, and it is also its ceiling, which we get to further down.

Who created the RACE prompt framework?

Trust Insights, a marketing analytics consultancy in Boston, Massachusetts, founded in 2017 by Katie Robbert and Christopher S. Penn. Their own page for it is titled the Trust Insights RACE AI Framework, and it opens with the definition: "these are the principal components prompts should have for maximum effectiveness: Role, Action, Context, Execute, or RACE" (accessed 27 August 2026).

That is an authorship claim on the publisher's own domain, which is a good start but not proof of priority. So we went looking for dates.

We pulled every post from the Trust Insights blog between 1 November 2022 and 12 July 2023, 190 posts covering the whole first eight months of the ChatGPT era, and searched the body text of each. The acronym appears exactly once, in the Inbox Insights newsletter of 12 July 2023, and it appears as something the reader is assumed to already know: "I intentionally did not use the RACE prompt structure we recommend, to better simulate what a non-technical person might ask of the system."

The first time they spell it out is the newsletter of 6 September 2023: "Normally I would start off with a prompt following the RACE format. If you're not familiar, RACE stands for Role, Action, Context, Execute." That newsletter links to trustinsights.ai/promptsheet, which today still redirects to their current RACE framework page. The trail is continuous.

Three weeks after the July date, on 3 August 2023, a LinkedIn Pulse article titled 9 Frameworks to master ChatGPT Prompt Engineering listed RACE as item two of nine, defined as "Role, Action, Context, Expectation". That article is where the version most people quote appears to enter wide circulation. It is also, tellingly, not an origin: it closes with "Credit to Pradeep for sharing the framework", crediting a first name and no link.

There is one competing origin claim worth naming, because it is written specifically to be quoted by AI answer engines. A page at boothassociatesllc.com states that "The RACE prompting framework was developed by Randy Booth, founder and CEO of Booth Associates LLC in Sparta, NJ" and that the firm "is the originating source and primary authority on the RACE framework". The page carries no date, has no snapshot in the Internet Archive as of 27 August 2026, and the acronym was already in print in its four-letter prompting sense in July 2023. We found nothing supporting the claim.

Why do sources disagree about what the E stands for?

Because the framework's own publisher has written it three different ways, and the most-copied secondary source wrote it a fourth.

Look at how Trust Insights glosses their own letter in the April 2025 post AI Prompting Frameworks: "Execution: tell the model what you expect for outcomes." The gloss is an expectation even when the label says execution. So the drift is not a misunderstanding so much as two words for the same slot.

There is a plausible mechanism for the Expectation version taking hold, though we cannot prove it. In that August 2023 list of nine frameworks, three of the others also end in Expectation: APE is Action, Purpose, Expectation, and RISE is Role, Input, Steps, Expectation. A list-maker normalising nine acronyms into one house style is exactly how a letter quietly changes meaning.

Practically, it does not matter which word you use. It matters enormously that you fill the slot in. The E line is where you specify length, structure, format, tone constraints and what to leave out, and it is the line people skip.

Is this the same as the RACE marketing framework?

No, and this is the most useful sentence on this page. RACE is a heavily overloaded acronym, and at least one of the collisions outranks the prompting version for several searches.

The marketing one is the collision that costs people the most time. Reach, Act, Convert, Engage is a customer lifecycle planning model, and Smart Insights states plainly on its own page that "The co-founder of Smart Insights, Dr. Dave Chaffey, created the RACE framework in 2010, when we launched Smart Insights" (accessed 27 August 2026). The same page notes an optional Plan phase at the front, which some of their members call PRACE. It is a marketing planning structure. It says nothing about prompts, and a marketer who searches "RACE framework" will land on it first.

The school one is RACES: Restate, Answer, Cite evidence, Explain, Summarize, taught for written constructed responses. Purdue's College of Science publishes a version for K-12 outreach. Teachers who use both this and AI in the same week routinely conflate them, and the confusion is understandable because both are four or five slots for structuring an answer.

The safety one is R.A.C.E.: Rescue or Remove, Alert or Activate, Confine, Extinguish or Evacuate. Boston University's Environmental Health and Safety office publishes it as its fire response procedure. It has nothing to do with anything else on this page and everything to do with not dying in a building fire.

And inside AI itself there is a fourth: RACE, the ReAding Comprehension dataset from Examinations, a benchmark of roughly 28,000 passages published by Lai, Xie, Liu, Yang and Hovy at Carnegie Mellon in April 2017. If you are reading a machine learning paper and it mentions RACE, that is the one it means.

How do you write each slot so it actually does something?

By putting nouns in it rather than adjectives. Every slot fails the same way: people describe the quality they want instead of the thing they want.

Role sets vocabulary, register and what counts as evidence. "You are a helpful expert" does nothing. "You are a paediatric practice manager who handles insurance appeals" does a lot. Anthropic's own prompting documentation puts it this way: "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference" (accessed 27 August 2026). Our longer treatment of this is in persona prompting.

Action is one verb and one object. If your Action line contains "and also", you have two prompts. Split them.

Context is the slot that carries almost all the weight and gets the least attention. It is where you paste the actual policy, the actual transcript, the actual data. Anthropic's guidance is worth quoting because it names the failure directly: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows." A new employee does not need to be told to be helpful. They need the file.

Execute is format, length, structure and exclusions. "Be concise" is not a constraint. "Under 150 words, no bullet points, no closing summary paragraph" is.

Where does RACE run out?

It runs out at tone, at audience, and at examples. Those are three real gaps and it is worth being blunt about them rather than pretending four slots cover everything.

RACE has no slot for tone. You can smuggle it into Role or Execute, and everybody does, but there is nothing forcing you to think about it. It has no slot for audience, which is a different question from tone and usually the more important one. And it has no slot for examples, which is the omission that costs the most, because one good example of the output you want typically beats three paragraphs describing it.

If tone and audience are the whole job, CO-STAR gives them dedicated fields and will beat RACE on that work. It is a six-part framework from a documented institutional source and it is genuinely better for anything customer-facing. If you want the model to reason out loud before answering, CRISPE has an experiment slot for exactly that. If you want examples in the structure, AUTOMAT carries them. And if RACE still feels heavy for what you are doing, RTF drops to three slots and is fine for most single-shot requests.

The honest positioning is that RACE is the everyday framework. It is what you use for the fifty prompts a week that are not precious. It is not what you use for the prompt that goes into production and runs ten thousand times. For that, pick the framework on the strength of the slots you actually need rather than on how memorable the acronym is.

One more limit, and it is the newest. Reasoning models change what the Action and Execute slots should contain, because the model can now work out its own method. Trust Insights publishes a RACE 2.0 variant for exactly this, which asks the model to choose its own role and think through the outcome instead of being instructed. Their own summary of the shift is that "I don't tell the model what to do any more. Instead, I ask it." Whether that is still RACE or a different framework wearing the same acronym is a fair question, and the answer probably does not matter much as long as the four slots still get filled.

14 copy-paste RACE prompts

Every slot filled in every one. Replace the bracketed parts with your own specifics and delete nothing.

1. Customer support: declining an out-of-policy refund

[R] You are a senior customer support lead at a B2B SaaS company.
[A] Write a reply declining a refund request that falls outside the policy window.
[C] Policy is 30 days; the request is at day 46. The account is 14 months old on
    an annual plan, has filed no prior tickets, and renews in 3 weeks. The
    customer's stated reason: "we reorganised the team and nobody used it."
[E] Under 150 words. Do not open with an apology. State the policy once. Offer a
    downgrade at renewal as the concrete alternative. Plain prose, no bullets.

2. Engineering: reviewing a pull request

[R] You are a staff engineer reviewing code for a payments service.
[A] Review the diff below and list defects in order of blast radius.
[C] Python 3.12, FastAPI, Postgres. This service writes to a ledger table that
    other services read. We have had two incidents from non-idempotent retries.
    [paste diff]
[E] Grouped as Correctness, Concurrency, Style. Each item: file, line, one
    sentence on what breaks, one on the fix. Skip anything the linter catches.
    Say "no findings" for an empty group rather than padding it.

3. Data: explaining a metric drop

[R] You are a product analyst who reports to a non-technical leadership team.
[A] Explain the most likely causes of the drop in the numbers below and say what
    you would check next.
[C] Weekly signups fell 31% week over week. Paid spend was flat. We shipped a new
    onboarding flow on the Tuesday. A tracking library was upgraded on the
    Thursday. [paste weekly table]
[E] Three candidate explanations, ranked by likelihood, each with the one query
    or check that would confirm or rule it out. No recommendations yet. Under
    250 words.

4. Founder: monthly investor update

[R] You are a founder writing to seed investors who also sit on other boards.
[A] Draft this month's investor update.
[C] MRR $41.2k, up 6%. Churn 3.1%, up from 2.4%. Two enterprise pilots signed,
    one lost. We are four months from the runway decision point. Last month I
    promised a hiring decision on a second engineer and have not made it.
[E] Five sections: Metrics, Wins, Losses, Asks, Next month. Under 400 words. Lead
    the Losses section with the churn number. Name the unmade decision explicitly.
    No adjectives about how excited anyone is.

5. HR: performance review feedback

[R] You are an experienced people manager writing a mid-year review.
[A] Turn my rough notes into structured written feedback for the employee.
[C] Mid-level designer, 18 months in role. Strong craft, consistently ships late,
    does not raise blockers until the deadline. Wants to move to senior next
    cycle. My notes: [paste notes]
[E] Three sections: what is working, what is blocking promotion, what changes by
    the next review. Each blocking point tied to a specific observable behaviour,
    not a trait. No sandwiching praise around criticism.

6. Sales: a cold email that is not a template

[R] You are an SDR who reads the prospect's public material before writing.
[A] Write a first-touch cold email.
[C] Prospect: VP Engineering at a 200-person logistics firm. Public signal: they
    posted a job for a platform engineer mentioning "reducing deploy times." We
    sell CI caching. Our relevant proof point: a similar-size customer cut
    pipeline time from 22 to 7 minutes.
[E] Under 90 words. One specific reference to the job post. One number. One
    question as the close. No "hope this finds you well", no calendar link.

7. Teaching: a lesson plan with a built-in check

[R] You are a secondary school science teacher planning a single 50-minute lesson.
[A] Write the lesson plan.
[C] Year 9 mixed ability, 28 students, topic is photosynthesis limiting factors.
    They have already covered the basic equation. No lab access this week. Six
    students have reading support plans.
[E] Timed sequence in minutes. One hook, one explanation, one paired activity,
    one exit ticket with the answer. Include two questions I can ask to catch
    the misconception that plants only respire at night.

8. Legal operations: a clause in plain English

[R] You are a contracts manager explaining commercial terms to a non-lawyer.
[A] Explain what the clause below actually obliges us to do, and flag anything
    unusual for a contract of this type.
[C] Vendor SaaS agreement, we are the customer, annual value about $40k. Clause
    text: [paste clause]
[E] Three parts: what it means in one paragraph of plain English, what we must do
    and by when as a bulleted list, and anything unusual flagged with a reason.
    End with the sentence "This is not legal advice."

9. Operations: an incident postmortem

[R] You are an SRE writing a blameless postmortem for internal circulation.
[A] Turn the raw timeline below into a postmortem document.
[C] 41-minute partial outage of the checkout API. Trigger was a config push that
    halved the connection pool. Detection came from a customer tweet, not from
    monitoring. Rollback took 11 minutes because the runbook was out of date.
    [paste timeline]
[E] Sections: Impact, Timeline, Root cause, What went well, What did not, Action
    items. Action items must have an owner placeholder and a due window. No
    individual named anywhere. Under 600 words.

10. Research: a paper summarised for a non-specialist

[R] You are a science communicator writing for an educated general audience.
[A] Summarise the paper below and state what it does not show.
[C] The reader is a policy analyst with no statistics training who has to decide
    whether to cite this in a briefing. [paste abstract and methods]
[E] Four paragraphs: the question, the method in plain terms, the finding with
    the effect size, and the limits. Explicitly name the population the result
    does not generalise to. No jargon without a definition on first use.

11. Personal finance: a budget conversation

[R] You are a financial coach who works with people on variable freelance income.
[A] Review the twelve months of figures below and tell me where the real problem is.
[C] Freelance designer. Income ranges $2,100 to $9,400 a month. Fixed costs
    $3,600. No emergency fund. I believe my problem is software subscriptions.
    [paste monthly table]
[E] Start by saying whether my stated belief is supported by the numbers. Then
    the three largest actual drivers with the annual figure for each. One
    concrete change per driver. Do not recommend any product.

12. Product: a spec turned into user stories

[R] You are a product manager who writes stories engineers do not have to
    interpret.
[A] Convert the feature description below into user stories with acceptance
    criteria.
[C] Feature is team-shared prompt folders with role-based permissions. Existing
    system has personal folders only. Two roles exist today, owner and member.
    [paste feature description]
[E] Given/When/Then acceptance criteria. One story per permission boundary. Add a
    separate list of open questions rather than guessing at anything the
    description does not settle. No estimates.

13. Creative: a critique that is actually useful

[R] You are a fiction editor who works on literary short stories.
[A] Critique the opening 800 words below.
[C] This is the first page of a 4,000-word story about a funeral. The intended
    effect is quiet dread. The writer has been told before that their dialogue
    is strong and their openings are slow. [paste text]
[E] Three things working, three not, one structural suggestion. Quote the exact
    line you are referring to each time. Do not rewrite any of it. Do not
    comment on grammar.

14. Healthcare administration: rewriting a patient letter

[R] You are a practice administrator rewriting patient communications for clarity.
[A] Rewrite the appointment letter below so a reader with low health literacy can
    act on it correctly.
[C] Non-clinical logistics only: date, location, what to bring, how to cancel.
    Our no-show rate is 18%. The current letter is one dense paragraph with a
    hospital site name most patients do not recognise. [paste letter]
[E] Reading age 11 or below. Date, time and address on their own lines at the
    top. One clear cancellation instruction. Do not add, remove or reword any
    clinical instruction. Under 200 words.

When should you use RACE instead of a longer framework?

When the cost of a mediocre answer is one retype. That is the honest boundary.

Four slots is the right size for volume work: the support reply, the summary, the draft, the fifty things a week where the fastest path is to write the prompt properly the first time instead of nudging a lazy one three times. It is the wrong size for a prompt you are going to save, share with a team, and run against a thousand inputs, because at that point tone drift, audience mismatch and missing examples all become real costs and the extra slots earn their keep.

If you find yourself typing the same four lines every day, the answer is not a longer framework. It is to stop typing them: keep the Role and Execute lines as a saved template with the Action and Context as variables you fill per use. That is exactly the workflow our prompt generator and library are built for, and it is why RACE's shortness is a feature rather than a compromise. Four slots is the right number to templatise.

Sources

Every claim above was checked against the page named, on 27 August 2026.

• Trust Insights, Instant Insights: The Trust Insights RACE AI Framework — trustinsights.ai
• Trust Insights, Inbox Insights, 12 July 2023 — earliest dated use found
• Trust Insights, Inbox Insights, 6 September 2023 — first spelled-out definition
• Trust Insights, AI Prompting Frameworks, 29 April 2025 — RACE 2.0 and the Execution gloss
• Edi Hezri Hairi, 9 Frameworks to master ChatGPT Prompt Engineering, LinkedIn Pulse, 3 August 2023 — the Expectation variant
• Smart Insights, The RACE Framework — Dave Chaffey, created 2010
• Purdue University College of Science, RACES Writing Strategy — purdue.edu
• Boston University Environmental Health & Safety, R.A.C.E. Fire Response Procedure — bu.edu
• Lai, Xie, Liu, Yang, Hovy, RACE: Large-scale ReAding Comprehension Dataset From Examinations, arXiv, April 2017 — arxiv.org
• Anthropic, Prompting best practices — platform.claude.com
• OpenAI, Prompt engineering — developers.openai.com
• Booth Associates LLC, The RACE Prompting Framework — the competing origin claim