The standard approach to giving an AI your brand voice is to describe it. Write a paragraph in your system prompt: conversational but authoritative, direct, avoids jargon, occasionally uses humor. Maybe add a few adjectives. Maybe include one or two example sentences.

This produces output that follows the description without quite sounding like you. The sentences are technically conversational. They're technically direct. There's something missing that's hard to identify — a flatness, a genericness that persists no matter how precisely you describe what you want.

The problem is that voice is not really describable. It's recognizable. You know it from examples.

What brand voice actually is

Think of a brand whose writing you find genuinely distinctive — a company whose emails you actually read, a newsletter that sounds unlike anything else, a campaign that stuck with you for years. If someone asked you to describe what makes their voice distinctive, you could attempt it. "Confident but not arrogant. Specific. Dry." The description captures something, but it doesn't capture it.

What captures it is the examples themselves. The specific subject line that earned your attention. The product description that was somehow funny without trying to be funny. The apology email that made you trust them more, not less.

Voice is the accumulation of specific choices — word-level, sentence-level, structural — made consistently over time. A description of those choices is a thin representation of the thing itself. Examples are the thing itself.

The gap between system prompts and swipe files

A system prompt tells your AI what to do. A swipe file shows your AI what good looks like.

In practice: a system prompt says "our tone is warm and direct." A swipe file contains the subject line that got a 41% open rate, the landing page section you keep coming back to as the reference for how you want to sound, the three ads that felt most true to the brand, and the email from two years ago that someone forwarded you with a note saying "this is exactly right."

Models learn from examples far more effectively than from descriptions. When an AI can reference your actual best work before writing — not a description of your best work, but the examples themselves — the output changes in ways that better prompting alone doesn't produce.

Building a brand voice swipe file

The starting point is your own past work. Go back through the emails, ads, and copy you feel best represents the brand. Save the ones that still feel right — not the ones that performed best (though those matter too), but the ones that sound most like you at your best.

Tag them consistently. brand-voice for anything that captures the tone well. headline, email-subject, cta for specific types. If you track performance, note what worked. A subject line with a 38% open rate is a better brand voice example than a technically polished one that nobody opened.

Add examples from outside your own work, too. Competitors who do something you admire. Brands in adjacent categories whose voice you'd want to borrow from. The more specific your collection is — not "good marketing writing in general" but "writing that sounds like the brand I'm trying to build" — the more useful it becomes as a reference.

How to use it with Claude

When Kaleidoscope is connected to Claude Code via MCP, your AI can search your swipe file before writing. A prompt like "write a product email for our new feature — search my swipe file for brand-voice examples first" gives Claude something to work from beyond its training average.

The result isn't automatically perfect. Brand voice in AI output still requires editing. But the gap between what the AI produces and what you want is significantly smaller when the model has seen your actual examples rather than just your description of them. You spend less time removing the genericness and more time refining something that's already close.

The discipline underlying this is the same one James Webb Young described in 1939: collect the raw material, keep it organized and accessible, and let the combinations work. The AI does the combining. Your swipe file is what makes the combinations worth reading.