What CRISPE Stands For
CRISPE is a prompt structure for getting Claude to produce expert-level output on any topic. It stands for:
- Capacity — what role or capacity Claude is adopting
- Role — the specific persona, including seniority and context
- Insight — the background and information Claude needs
- Statement — the specific task
- Personality — how the response should feel (tone, style)
- Experiment — request multiple variations or approaches
It's not magic. It's just a structured way to give Claude the specifics it needs to stop generating generic output.
Why Generic Prompts Fail
The most common "expert prompt" advice online is some version of: Act as an expert in [field]. Help me with [task].
This fails because "expert" is too vague. Claude averages across all experts in the field and gives you committee-average advice. Generic input produces generic output.
Example: Act as a sales expert. Write me a cold email.
Result: A bland cold email with "I hope this email finds you well" and "I wanted to reach out" — the exact patterns that get every cold email ignored.
Example using CRISPE:
CAPACITY: Senior sales consultant who has built 3 B2B SaaS sales teams
ROLE: You are Dana, a VP of Sales at a Series B fintech who sold her
last company to Stripe. Direct, hates fluff, thinks most sales advice
is bad. You've done cold outbound that converts at 18% reply rate.
INSIGHT: I'm a solo technical founder. My product: B2B API monitoring
tool, $99/mo per team. My target: heads of engineering at Series A-B
SaaS companies. I've written 50 cold emails and gotten 2 replies.
STATEMENT: Rewrite my cold email to get replies. Show me the rewrite
and 2 alternative approaches. Flag what was wrong with my original.
MY DRAFT:
[paste your current email]
PERSONALITY: Direct, specific, no marketing BS. Write like Dana would
talk — concise, opinionated, occasional dry humor.
EXPERIMENT: Give me 2 versions — one optimized for reply rate, one
optimized for getting a meeting. Explain which you'd send and why.
This produces a fundamentally different output than the generic version. Same Claude, same model, totally different quality.
The Six Parts — What Each One Does
Capacity
"Capacity" is about the function Claude is filling. Not the title — the job to be done.
- Weak: "A marketer"
- Better: "A growth marketer who specializes in B2B SaaS at the seed-to-Series-A stage"
- Best: "A growth marketer who has taken 3 SaaS companies from $100K to $1M ARR through SEO, content, and partnership deals — but not paid ads (she hates paid ads)"
The best version anchors Claude in specifics that Claude can reason from.
Role
Role is the persona layer on top of capacity. Add name, context, what they'd reject.
- Weak: "A senior engineer"
- Best: "You are Priya, a staff engineer at a Series C fintech who has shipped payments systems at 3 companies. She refuses to recommend microservices for companies under 20 engineers because she watched it destroy her last team's velocity."
The "what they'd reject" part is the secret. It gives Claude something to push back against, which produces sharper advice.
Insight
The context Claude needs. Don't make Claude ask you for information — provide it upfront.
- Weak: "I want to grow my SaaS"
- Best: "I'm at $30K MRR, solo founder, targeting SMBs at $49/mo, 200 customers, 8% monthly churn, 60% of signups come from content marketing, my biggest channel bottleneck is I only publish 1 post a week."
Give Claude enough context that the answer has to be specific.
Statement
The specific task. Use imperative verbs.
- Weak: "Help me with marketing"
- Best: "Rank the top 5 content topics I should write next based on search volume and conversion intent, given my stated context. Include the specific headline I'd use for each."
Specificity in the statement is the biggest single lever on output quality.
Personality
How should the response feel? What tone? What style?
- Terse and direct vs verbose and exploratory
- Formal business vs casual
- Opinionated vs balanced
- First person vs third person
Claude defaults to verbose + balanced + third person. If you want something different, name it explicitly.
Experiment
This is the least-used part and the biggest unlock. Ask for multiple variations or approaches.
- "Give me 3 alternative headlines, ranked by likely CTR, with reasoning"
- "Show me the aggressive version and the conservative version"
- "Draft 2 versions — one optimized for speed to close, one optimized for relationship"
- "Give me the first-principles answer AND the industry-standard answer. Tell me when to use which."
Multiple options force Claude to articulate tradeoffs instead of giving you "the" answer.
CRISPE + Prompt Codes
CRISPE works even better stacked with tested prompt codes. The framework gives Claude context; the codes shape how Claude reasons over it.
Example: CRISPE + L99
[CRISPE block for B2B sales VP persona]
L99 STATEMENT: Should I make my first sales hire now or wait until $50K MRR?
CRISPE gives you the expert perspective. L99 forces them to commit to one specific answer instead of enumerating tradeoffs.
Example: CRISPE + /blindspots
[CRISPE block for growth marketer persona]
/blindspots STATEMENT: Review my content strategy for gaps and assumptions I'm making.
CRISPE gives you the specialist's knowledge. /blindspots applies the knowledge to find your blind spots specifically.
See tested prompt codes that combine with CRISPE →
When NOT to Use CRISPE
CRISPE is overkill for simple questions. Don't use it for:
- Quick code lookups
- Simple factual questions
- Writing you'll edit heavily anyway
- Tasks where a 2-sentence answer is enough
CRISPE is for situations where the output quality matters enough to spend 2-3 minutes setting up the prompt.
Common CRISPE Mistakes
Mistake 1: Over-specifying personality
Don't write "Personality: confident, authoritative, expert, trustworthy, knowledgeable, experienced, wise." That's tone-vomit, not personality. Pick 2-3 specific adjectives and stop.
Mistake 2: Under-specifying role
"You are a marketing expert" is not a role. It's a job title. Add name, company context, and — critically — what this person would reject.
Mistake 3: Skipping Experiment
Most people skip the E in CRISPE. Don't. Asking for 2-3 alternatives is what separates a CRISPE prompt from a fancier version of "act as an expert."
Mistake 4: Using CRISPE without real specifics
CRISPE with made-up specifics produces made-up-specific output. If you don't know your user's actual context, CRISPE won't invent it. Give Claude real numbers and real constraints.
FAQ
Is CRISPE an official Anthropic framework?
No. It's a community-discovered prompt structure. Similar to L99, /skeptic, and other prefixes — not baked into the model, just a reliable pattern.
Does CRISPE work with ChatGPT or Gemini?
Partially. The structure is model-agnostic, but the specific phrasing that works with Claude doesn't always transfer. ChatGPT tends to need more explicit formatting; Gemini needs shorter context. Tune per model.
How long should a CRISPE prompt be?
300-800 words is the sweet spot. Under 200 and you're not providing enough context to matter. Over 1000 and the model starts losing focus on what you're asking for.
Can I use CRISPE in Claude Code?
Yes. Use it as a system-prompt-style header in your session or in your CLAUDE.md file. Works especially well for agent use cases where you want Claude to maintain a specific role across many turns.
Where can I see CRISPE in action?
The 40-page Complete Claude Guide ($5) has a full CRISPE chapter with 20+ real examples across different founder use cases (sales emails, strategic decisions, hiring, pricing, investor updates). The free prompts library includes a PERSONA code that's a simplified CRISPE.
What's the fastest way to get good at CRISPE?
Write 5 CRISPE prompts this week for tasks you actually care about. Show the first draft to Claude and say "critique my CRISPE structure — what's too vague, what's missing specifics." Claude is surprisingly good at evaluating its own prompt inputs.
The Shortcut
If you want CRISPE without having to write the structure every time, the Combo Generator has 20 pre-built CRISPE-style templates for common founder tasks. Pick your task, paste your specifics, run.
The full framework is in the Cheat Sheet ($10) with 10 worked CRISPE examples across different industries.