Brand voice system

As the sole copywriter at M1, I write almost everything that clients see. Marketing emails, product copy, in-app messages, push notifications, App Store screens… the list goes on. All of it needs to adhere to two big things: 1) the brand voice and 2) FINRA 2210. Every retail communication has to be fair, balanced, and not misleading, with specific disclosures depending on what it's about. I helped shape messaging, voice, and tone amid a brand evolution. There's documentation, but much of the nuance lives in my head.

There are two ways that can end. I become the bottleneck for copy review, or I let people inside my head.

M1 is a tech company bullish on AI. If you were me, maybe you'd ask: can AI write copy? Is my job obsolete?

Yes, AI can write copy. That's obvious. I'm not too precious about my role to admit it.

The real question is: can AI write copy like a senior brand copywriter with over 10 years of experience and a particular set of pet peeves? And can it avoid triggering too many compliance edits?

Turns out it can. If I teach it to.

Enter: emilybot, a Claude skill that takes everything I think about while writing copy for the company and uses it to generate solid first drafts.

For me, it's not a go-to step in my drafting process. I still have a lot of fun stringing words together from my brain.

For people at the company who aren't writers, it gets them a lot closer to how I'd approach a copy draft. That way, they can draft sample copy that starts a lot closer to what I'd say, instead of waiting on me to come up with something.

It's modular. It always pulls up Philosophy, Voice & tone, and Compliance for every task. Then it pulls up the specific channel based on the task.

FINRA 2210 is referenced for do-and-don't patterns: the difference between a guarantee and a possibility, between creating urgency with a real deadline versus manufactured fear, between a factual comparison and an unsubstantiated one. Alongside that, a coded disclosure library organized by product, so the correct language attaches to the correct claim automatically.

Disclosures are scoped by channel. Requirements that apply to a web page don't apply to in-product copy. To get it right, I had to understand the rules and collab with the compliance team.

The system includes an explicit prohibition on AI writing patterns: the “it's not X, it's Y” construction, em-dashes. Basically, if it reads like a language model wrote it, rewrite it.

I put that in for a good reason. I didn't want the copy to sound machine-generated. Technically on-brand but… somehow soulless. And let's be honest, as soon as you send an email with an em-dash, people start yelling “AI! AI!” (Such a shame. It used to be my favorite punctuation.)

Nine versions so far, each with a changelog noting what changed and why. One named owner.

It is built to absorb new products, audience updates, and whatever else arrives. I can add copy phrases that get banned by compliance to the “do not say” list. Oh, and that means it hasn't stolen my job quite yet.