The CRISPE Prompt Framework, Explained (2026)

A full breakdown of CRISPE, built to skim. Sections, in order:

1. What CRISPE is and when it shines · 2. C — Capacity / Role · 3. R — Insight (Relevant context) · 4. I — Statement (Instruction) · 5. S — Personality · 6. E — Experiment · 7. A full assembled CRISPE prompt · 8. CRISPE vs plain prompting (table) · 9. Common mistakes · 10. FAQs · 11. Sources & further reading.

A note on the acronym: CRISPE is mapped slightly differently across the community. We use the common reading — Capacity/Role, Insight (the relevant context/background), Statement (the instruction), Personality (voice/style), Experiment (ask for multiple options). The labels matter less than the slots.

CRISPE is built for open-ended, analytical work: strategy questions, brainstorming, problem framing, comparing approaches. Its defining move is the Experiment slot, where you ask the model for several distinct responses so you can compare angles rather than commit to its first guess.

Compared with RTF (a single locked deliverable) or CO-STAR (tone-controlled copy), CRISPE assumes the answer isn't obvious yet and you want the model to think and offer choices. That makes it a poor fit for mechanical tasks — using Experiment on a simple extraction just burns tokens — but a strong fit when divergence is the goal.

If you find yourself running the same single prompt five times to 'see other options,' CRISPE bakes that into one prompt via Experiment.

Capacity (also read as Role) sets who the model is acting as and at what level of expertise. For analytical work, make it a role that would plausibly hold a strong, defensible opinion.

``` Act as a seasoned growth strategist who has launched pricing changes at several B2B SaaS companies. ```

**Why it works:** a specific, senior role shifts the model toward considered analysis rather than surface-level summary — the same reliable role-prompting effect documented by DAIR.ai. The expertise level is part of the instruction: 'seasoned' and 'has launched pricing changes' signal the depth you expect.

Insight is the relevant background and context the model should reason from: your situation, constraints, what's been tried, what success looks like. In CRISPE this slot does the heavy lifting, because analysis is only as good as the facts behind it.

``` Insight: We are a B2B SaaS at $2M ARR, 40% gross margin on a flat per-seat plan. Churn is low but expansion revenue is flat. Sales says big accounts want usage-based options; finance fears revenue volatility. ```

**Why it works:** rich context lets the model ground its options in your reality instead of generic playbooks. State your constraints explicitly — the more it knows about what you can't do, the more useful the options it proposes.

Statement is the actual instruction — the question or task you want addressed. In analytical prompts this is usually a 'how should we / what are the ways to' question rather than a 'write me X' command.

``` Statement: Propose pricing-model changes that could grow expansion revenue without scaring off finance on volatility. ```

**Why it works:** a clear instruction keeps the model's analysis pointed at your actual decision. Pair it with the Experiment slot below so the instruction yields several candidate answers rather than one. Keep the instruction to a single decision; multiple questions dilute every option.

Personality is the voice and disposition of the response — direct and opinionated, balanced and cautious, devil's-advocate, plainspoken. For strategy work, an opinionated personality often surfaces sharper trade-offs than a neutral one.

``` Personality: Direct and opinionated. Name the trade-offs plainly and say which option you'd bet on and why. No hedging, no 'it depends' without a follow-through. ```

**Why it works:** specifying personality stops the model from defaulting to balanced-but-useless 'on one hand / on the other' output. For reusable brand or analyst voices, the Brand Voice Generator can capture the personality once and reuse it.

Experiment is CRISPE's signature: you ask for multiple distinct responses so you can compare. Specify how many and how they should differ (by strategy, by risk level, by time horizon).

``` Experiment: Give me 3 distinct options, ordered from lowest to highest risk. For each: the model, the upside, the main downside, and one sentence on when it's the right call. ```

**Why it works:** requesting several explicitly-differentiated options turns one call into a comparison, which is far more useful for a decision than a single 'best' answer the model can't actually validate. Constrain how they differ, or you'll get three near-duplicates. This is the divergent counterpart to the refine-loop in our 12 prompt patterns guide.

All five elements stacked into one prompt you can paste and adapt:

``` Capacity: Act as a seasoned growth strategist who has launched pricing changes at several B2B SaaS companies. Insight: We are a B2B SaaS at $2M ARR, 40% gross margin on a flat per-seat plan. Churn is low, expansion revenue is flat. Big accounts want usage-based options; finance fears revenue volatility. Statement: Propose pricing-model changes that grow expansion revenue without creating volatility finance can't live with. Personality: Direct and opinionated. Name trade-offs plainly and say which you'd bet on and why. Experiment: Give me 3 distinct options, ordered low to high risk. For each: the model, the upside, the main downside, and when it's right. ```

Assembled this way, CRISPE gets you a comparable set of analyzed options with a clear recommendation — a much better input to a decision than a single generic answer. Treat the model's options as a starting set to pressure-test, not a verdict.

---

As with every framework, drop slots that don't apply. If voice doesn't matter, skip Personality. The non-negotiable parts for CRISPE are Insight (context) and Experiment (multiple options) — those are what make it CRISPE rather than RTF.

Use CRISPE when: the answer isn't obvious yet — strategy, brainstorming, problem framing, comparing approaches — and you want the model to analyze and hand you several differentiated options to choose from.
Use RTF / CO-STAR instead when: you already know the single deliverable you want. CRISPE's Experiment slot wastes tokens on tasks with one correct output shape, like extraction or a fixed piece of copy.

It helps to see CRISPE as one point on a spectrum of how much you let the model think. The frameworks differ mainly in how convergent or divergent they are.

RTF and RACE are convergent: you know the single deliverable you want and you're steering the model toward it. One output, judged against a format or a success bar.

CO-STAR is also convergent but voice-focused: one piece of copy, tuned for a specific reader and register. Still one answer.

CRISPE is divergent: the Experiment slot deliberately produces several differentiated options because the right answer isn't settled yet. You're using the model to widen the option set before you narrow it. That makes CRISPE the natural front end of a decision — generate options with CRISPE, then converge on the winner with an RTF or RACE prompt that turns the chosen option into a finished artifact.

A useful workflow chains them: CRISPE to brainstorm three pricing strategies, pick one, then a RACE prompt to write the rollout brief for it. That CRISPE-then-RACE handoff is a small prompt chain in itself — divergent step feeding a convergent step.

Unconstrained Experiment. 'Give me some options' yields near-duplicates. Specify the count and the axis they vary on (risk, horizon, strategy) so the options are genuinely distinct.

Thin Insight. Analytical output is only as good as its context. A vague situation produces a generic SWOT. Spend your effort here.

Neutral Personality on a strategy task. Letting the model default to balanced 'it depends' output wastes the framework. Ask it to commit to a recommendation and defend it.

Treating options as answers. CRISPE surfaces candidates; it can't validate them against your real numbers. Use the options to structure your thinking, then verify — don't paste a model's risk assessment into a board deck unchecked.

CRISPE is a community framework; the techniques it bundles are documented in established sources (current as of June 2026):

DAIR.ai Prompt Engineering Guide: https://www.promptingguide.ai/

Learn Prompting: https://learnprompting.org/

OpenAI prompting guide: https://platform.openai.com/docs/guides/prompt-engineering

Claude prompt-engineering overview: https://docs.claude.com/en/docs/build-with-claude/prompt-engineering/overview

Google Gemini prompting strategies: https://ai.google.dev/gemini-api/docs/prompting-strategies

Related on this site: Complete Guide to Prompt Engineering, 12 Prompt Patterns That Convert, and the simpler RTF and CO-STAR frameworks.