The CRISPE framework for prompting
CRISPE is a five-part prompt framework: Capacity and role, Insight, Statement, Personality, and Experiment. It is more involved than RACE, and that is the point. CRISPE earns its extra steps when the voice of the answer matters and when you want the model to offer you a few options rather than commit to one.
The acronym is a little awkward because the first two letters belong to one step (Capacity and Role), but the structure underneath it is sound.
Last reviewed
Capacity and Role: who is the model?
Capacity and Role is where you tell the model who it is, like the R in RACE. CRISPE splits it into capacity (the broad expertise) and role (the specific hat). "You are an expert in behavioural economics, acting as a pricing consultant" gives the model both the deep knowledge base and the immediate job.
You are an expert in behavioural economics acting as a pricing consultant for a subscription business.
Insight: what background does it need?
Insight is the background and context: the situation, the constraints, the things the model could not know without being told. This is where you brief it on your product, your audience, and whatever has already been tried. Treat it as the paragraph you would send a freelancer before they start.
We sell a 9 GBP per month note-taking app. Most churn happens in the first two weeks. We have already tried a discount offer and it did not move the numbers.
Statement: what is the task?
The Statement is the task itself, stated plainly with an action verb. It is the equivalent of the A in RACE. Keep it to the one thing you actually want done; the surrounding steps carry the nuance.
Suggest five pricing or packaging changes that might reduce first-fortnight churn.
Personality: what tone should it use?
Personality is where you set the voice, tone, and style of the response, the step RACE does not have. Formal or casual, blunt or diplomatic, plain or technical: name it. This is what makes CRISPE worth the extra effort for marketing copy, brand voice, and anything customer-facing.
Write in a direct, pragmatic tone. No hype, no hedging. Assume the reader is a busy founder.
Experiment: how many options should it give?
The Experiment step asks the model for more than one answer. Requesting several variations and then comparing them is one of the most reliable ways to get a good result, because the first draft a model produces is rarely its best. Ask for three to five options with a short rationale for each.
Give me five distinct options. For each, add one sentence on the behavioural principle it relies on and the main risk.
When should you use CRISPE instead of RACE?
Reach for CRISPE when tone is a first-class requirement or when you genuinely want choices. Brand copywriting, customer emails, naming, and positioning all benefit from the personality and experiment steps. For quick analytical tasks where the format matters more than the voice, RACE is faster and gets you there with less ceremony.
You can also mix them. A common pattern is RACE for the bones plus the experiment step bolted on the end: write the brief, then ask for several variations to compare.
Key takeaways
- CRISPE covers Capacity and role, Insight, Statement, Personality, and Experiment.
- Its two distinctive steps are Personality (voice and tone) and Experiment (ask for several variations).
- Use it when tone matters or when you want options to compare, rather than a single answer.
- For fast analytical tasks, RACE is lighter; the two frameworks mix well.
See how your prompt scores
Paste any prompt into the free Promptrace scorer and get an instant breakdown across all seven dimensions. No signup.
Open the prompt scorer →Frequently asked questions
What does CRISPE stand for?
CRISPE stands for Capacity and role, Insight, Statement, Personality, and Experiment. The first two letters cover one combined step (your expertise and the specific role), then you supply background, state the task, set the tone, and ask for several variations.
When should I use CRISPE instead of RACE?
Use CRISPE when the voice of the answer matters or when you want the model to generate several options to compare, thanks to its Personality and Experiment steps. RACE is quicker for straightforward analytical tasks where format matters more than tone.
Sources and further reading
Related guides
RACE
Role, Action, Context, Expectation. The four-part framework Promptrace is named after.
TechniquesFew-shot prompting
Show the model one or two worked examples of what good looks like. The most reliable steering technique.
TechniquesChain-of-thought
Ask the model to reason step by step. Why it helps on hard problems, and when to skip it.