Summit Rx emails sounded like five different people wrote them — because five different people did. I built a Custom GPT that encoded our voice so every marketer drafts in the same confident tone, every time.
Marketing at Summit Rx isn't a single copywriter behind a curtain — it's four or five people writing emails, LinkedIn posts, ad copy, and landing pages under real deadlines. Every one of them is talented. Every one of them writes slightly differently.
The result: a subscriber who got two emails a week from us would notice something was off. One week's email sounded warm and specific. The next felt corporate and hedged. A third used phrases like "seamless solution" that nobody on the team would ever actually say out loud. The brand voice wasn't broken. There were just too many versions of it.
The team had been using vanilla ChatGPT for draft assistance for months. It writes fine business English but it writes it in ChatGPT voice — the hedgy, "certainly!" sales-rep tone that every LLM user recognizes within two sentences. Copy went out with that tone baked in. Made everything feel slightly generic. Subscribers can smell it.
First attempt: a shared Google Doc titled "Summit Rx Brand Voice" with bullet points. Be direct. Be specific. Avoid jargon. Use short sentences. The usual brand-voice advice.
Nobody read it. And when they did, they couldn't actually use it — "be direct" is a principle, not an instruction. You can't tell ChatGPT to "be direct" and expect it to know what that means in our company's voice. The doc sat in a Drive folder untouched for three months.
Second attempt: a prompt template people could paste into ChatGPT. About 10 lines of "You are a marketer at Summit Rx. We speak like..." with do/don't rules. That worked better — when people remembered to use it. Nobody wanted to copy-paste an 800-word prompt every time they started a draft. Friction killed adoption inside two weeks.
What the team actually needed: a dedicated chat tool that already knew the voice so they could just open it and start writing. That's a Custom GPT — OpenAI's feature that lets you create a specialised version of ChatGPT with your own instructions and reference files baked in. You build it once, share a link with your team, and they get a chat tool that already "knows" your style.
A Custom GPT — a specialised version of ChatGPT, accessible via a shared team link — built with three layers of voice instructions plus a small knowledge base of Summit Rx's best-performing past copy.
1. I started with data, not principles. Pulled the top 40 emails Summit Rx had sent in the last two years — ranked by a combined score of open rate + click rate + reply rate. These were the emails that worked with subscribers. Whatever voice they shared, that was the voice worth keeping.
2. Ran the corpus through Claude to extract patterns. Instead of writing the brand voice from scratch, I asked Claude (Anthropic's AI assistant) to read all 40 emails and tell me what they had in common: sentence-length patterns, recurring phrases the team uses, phrases that would feel out of place, whether the team opens with questions vs statements, how transitions are handled, how each email closes. Claude returned about 30 distinct voice traits. I kept the 15 sharpest ones.
3. Built a structured voice spec — rules, not principles. Not paragraphs of prose — a clean reference the AI could actually apply:
4. Uploaded the 40 best emails as GPT knowledge files. Custom GPTs let you upload reference documents the AI can read while it's working — like giving an employee a binder of past company work to reference. The instructions tell the GPT what to do; the knowledge files show it how, in actual context. When a marketer asks for a draft, the GPT has 40 real Summit Rx emails on tap — not abstract rules, real sentences from our best work.
5. Deployed as a team-wide Custom GPT. Published to Summit Rx's ChatGPT workspace and shared with everyone in marketing. Each person bookmarked it in their browser next to Gmail and HubSpot. When someone's writing a re-engagement email, they open the GPT first, paste the brief ("we're following up with subscribers who haven't opened in 60 days, the offer is X"), and get three draft variants that sound like Summit Rx. They pick one, tweak it, send it.
The qualitative win: new hires onto the marketing team get up to speed on Summit Rx's voice in days, not months. Instead of reading a brand guide, they use the GPT for their first week — and the way it writes back teaches them what the voice actually sounds like. The Custom GPT became an onboarding tool as much as a drafting tool.
I'd version the voice spec from day one. The first version of the spec was what I thought the Summit Rx voice was. After three months of actual use, some of those rules turned out to be wrong (the "never use exclamation points" rule was too strict — the GPT ended up sounding monotone). I'd treat the voice spec like a product: v1, v2, v3, with changelog. Right now it's just "the current prompt," which makes iteration harder.
I'd A/B test voice variations. The GPT currently produces one voice. A better version would produce slightly different variants for different use cases — a warmer voice for re-engagement emails, a crisper voice for transactional updates. I know the variants exist in our best past copy. The Custom GPT doesn't know to apply them contextually yet.
I'd wire it into Gmail, not just make it a bookmark. Opening a separate tab to get a draft and then pasting it back into Gmail still adds friction — enough that for really quick replies, people skip the GPT and type from scratch. A Gmail extension (or an MCP server surfacing the same voice) that offers the draft inside the compose window would close the last gap. The AI-drafted sales replies workflow is already doing this for sales — the marketing side hasn't caught up yet.
Biggest lesson: brand voice isn't a PDF or a set of principles. It's the pattern your best work already has. The job isn't to invent it — it's to notice it, capture it, and make it easy for the team to draw from. That's a knowledge-management problem, not a prompt-engineering problem.
I built this on OpenAI's Custom GPT feature — but the pattern is "encode brand voice via examples + structured rules into an LLM-powered drafting assistant the team uses every day." Custom GPT is one path. Other equally good paths:
None of those tools matter until you've done the upstream work: the analysis pass on past writing to extract voice traits, then turning those traits into rules and examples the model can actually use. Tooling is downstream of that thinking.