Why prompts matter so much
The quality of what Claude gives you is almost entirely determined by the quality of what you give Claude. A vague prompt gets a vague answer. A specific, well-structured prompt gets something you can actually use.
Most Optimists who feel like AI doesn't work for them are writing prompts like "write a blog post about [topic]." That's like sending a new hire an email that says "write content." It's technically a request, but it's not a brief.
The good news: this is a learnable skill. And you're about to learn it. 🎓
The CRISP framework
We built this framework specifically for how Optimists work. It gives you a repeatable checklist for writing prompts that consistently get good results.
Context
Tell Claude who you are, what you're working on, and what the situation is. The more specific, the better.
Role
Tell Claude what role to play. This improves style and tone significantly. One important caveat: persona prompting helps with creative work but can reduce factual accuracy. Use it for copy and creative tasks, not for fact-finding.
Instructions
Be specific about what you want. List what to include, what to avoid, and what format to use. Don't just say "write me a caption." Tell Claude exactly what good looks like.
Samples
Give examples of what good output looks like. Paste in past approved captions, copy from the client's website, or a link to their Communication Codex. Examples do more work than descriptions.
Parameters
Specify format, length, number of variations, and any hard constraints. Tell Claude exactly what the output should look like before it writes a single word.
Before and after: a CRISP in action
Prompt comparison
Write me a blog post about SEO.
You are a content writer for Online Optimism, a digital marketing agency in New Orleans. Write a 600-word blog post for our website about how small businesses can improve their local SEO in 2026. Use a conversational, witty tone. Write in first person ("In my experience..."). Include specific, actionable tips. Do not use any of these words: leverage, robust, seamless, delve, landscape, harness, elevate, journey. Do not use em dashes. Do not use the structure "Not only X, but Y." Vary your list lengths (never exactly three items in a row). End with a CTA that invites the reader to get in touch with an Optimist.
The second prompt will give you output you can actually use. The first prompt will give you... something.
Prompt refinement techniques
You don't have to get it perfect on the first try. Think of prompting as a conversation, not a one-shot request.
- Start broad, then narrow. Ask for a first draft, then follow up: "Now make the opening hook more surprising." Or: "Remove any word from this list: [paste]." Or: "Rewrite the third paragraph in a more casual tone."
- Paste the Communication Codex directly. If there's no Claude Project set up for a client, paste their Codex text into the chat. It's not as elegant as a project, but it works.
- Use the "Continue" prompt. If Claude gives you 3 options and you want more, just type "Give me 3 more." It will build on the same context.
- Tell Claude what's wrong, not just what you want. "This is too formal" is less useful than "Replace the word 'utilize' with 'use' throughout and shorten every sentence over 20 words."
💡 Reddit content tip: Never start a Reddit prompt with "Write a Reddit post about..." Instead, give Claude the subreddit, the question being answered, the persona of the account, and examples of what good Reddit responses look like in that community. Reddit has its own voice. You have to brief Claude on it.
Prompt Lab: build your own
Use the CRISP framework fields below to build a prompt for something you're actually working on. The assembled prompt shows up at the bottom, ready to copy.
Common prompt mistakes to avoid
- Being too vague about the output format. "Write something" gives Claude too much freedom. Tell it exactly what you need.
- Forgetting to include constraints. "Don't use em dashes" and "avoid these words" matter. Claude defaults to its natural patterns unless you redirect it.
- Using role prompting for fact-based tasks. Telling Claude to "act as an expert" can make it more confident and less accurate on factual claims. Save role prompting for creative and stylistic work.
- Starting over from scratch when output isn't perfect. Refine with follow-up messages. It's faster and you keep the good context.
- Not pasting the Communication Codex when there's no project. Brand voice matters. If there's no Claude Project, put the context in the prompt.