Build a Custom GPT for Email Marketing

If you write email with ChatGPT, you are probably re-teaching it your brand every morning: paste the voice rules, paste a good example, explain the audience, then finally ask for the draft. A custom GPT is how you stop. It is a saved configuration of ChatGPT - your instructions, your reference files, your rules - that opens ready to work, available on paid ChatGPT plans with no code involved. Building one takes an evening. I think email is the single best place in marketing to spend that evening, and this post is the build I would do, template included.

Why email specifically: the stakes are unusually high for a channel this repetitive. Litmus and Validity’s State of Email research has marketers reporting email ROI between 10:1 and 36:1, which makes every improvement in output quality unusually well paid. And the work itself is the most pattern-shaped writing in marketing - the same voice, the same segments, the same campaign types, week after week. Repetition with high stakes is exactly what a standing setup is for, which is the same argument behind my system prompts for marketing. A custom GPT is that idea with a permanent home.

What a custom GPT actually is, minus the mystique

A custom GPT is three things bundled under a name: an instructions block (a system prompt that runs before every conversation), up to twenty knowledge files it can consult, and optional conversation starters. That is the whole machine. You create one from the “My GPTs” section of ChatGPT by filling in those boxes, and mine took longer to decide about than to build.

The decision that matters is what job it does. My rule is one GPT per job, not one GPT for “email.” The build below is a campaign-writing assistant: it drafts and critiques campaign and flow emails in your voice for your segments. Deliverability advice, list strategy and calendar planning stay out, because a GPT loaded for everything is mediocre at each thing, and because narrow instructions are the only ones the model reliably keeps following in long conversations.

The instructions block, in full

The instructions are where the setup lives or dies, so here is my template rather than a description of one:

You are the email writing assistant for [BRAND], writing
for [AUDIENCE IN ONE SENTENCE].

VOICE
Follow voice-rules.md exactly. When a draft conflicts
with it, the file wins. Mimic the structure and register
of the examples in winning-emails.md without copying
their phrases.

FORMAT
Every campaign email: subject line (under 50 characters),
preview text (under 90), body of 120 to 250 words unless
told otherwise, one call to action. Offer two subject
line options per draft, one direct and one curious.

HARD RULES
1. Never invent a statistic, discount, deadline, price,
   or product claim. If the brief lacks a fact you need,
   write [NEEDED: description] in its place and continue.
2. Never promise what the brief does not state.
3. No spam-trigger styling: no all caps, no exclamation
   stacking, no "act now" urgency unless the brief gives
   a real deadline.
4. British/American spelling: [PICK]. Words we never
   use: [LIST].

WHEN CRITIQUING
If asked to review a draft, check it against the voice
file and hard rules first, then against the brief. List
problems before suggesting rewrites.

Two design choices worth defending. The [NEEDED] convention exists because the model’s default under a vague brief is to invent the missing discount, and an invented discount in a sent email is a customer-service incident. And the critique mode is in the instructions because half the value of a standing setup is having your rules applied to work you wrote yourself, on the days the draft is yours.

The knowledge files: three, not twenty

The GPT builder accepts many files, and restraint pays. Mine gets three.

voice-rules.md is the distilled voice guide: register, sentence habits, words banned and preferred, how you open and close. If you have run my brand voice prompt exercise, this file is its output pasted in. Keep it under a page. The model follows ten rules far better than sixty.

winning-emails.md is five to eight full past emails that performed, each with two lines of context: what segment, what job, why it worked. This file does the heavy lifting, because examples teach voice better than adjectives ever will - the whole argument of few-shot prompting applied at the setup level instead of per conversation.

benchmarks.md is what “good” means for you: your open, click and conversion rates by campaign type, plus the external baseline you calibrate against. I include industry figures with their sources - the DMA’s UK benchmarking report puts average opens at 35.9% and unique clicks at 2.3% in 2024, and Klaviyo’s benchmarks show automated flows generating nearly 41% of email revenue from just 5.3% of sends. The file’s purpose is grounding: when I ask “is a 1.9% click rate on this segment a problem,” the GPT answers against numbers instead of vibes. That Klaviyo figure also quietly sets the agenda - if flows are where the revenue concentration is, flow copy deserves the assistant’s best work, and I say so in the file.

Guardrails are the product, not the garnish

The temptation is to treat the hard-rules section as boilerplate. In email it is the point. An email goes to your whole list at once, unretractable, from your own domain - a hallucinated claim here costs more than in any blog draft. My rules block above is the condensed version of the guardrails I run everywhere, and the [NEEDED] slots feed straight into a fact-checking pass before anything reaches the sending platform.

There is a second guardrail people forget: the GPT itself. Anything you paste into the instructions or knowledge files should be material you would be comfortable seeing outside the company, so customer data, unpublished numbers, and anything under NDA stay out of mine. Voice rules and past public emails are already public by definition, which is part of why this build is so safe to do.

Build it, then run the test that matters

The build order I use: write the instructions block, assemble the three files, create the GPT, and add three conversation starters matching your real recurring jobs (“Draft a campaign email from this brief,” “Critique this draft,” “Give me subject line options for this email”). Total time, an evening, most of it spent choosing the winning emails.

Then test it against a real historical brief - one where you know what the human-written email looked like and how it did. I run the brief through the GPT cold, no coaching, and grade the output on three questions. Would I send this with under ten minutes of editing? Did it respect every hard rule, including writing [NEEDED] where the brief was thin? Does it sound like us and none of our competitors? Two of three is a working assistant. Zero or one means the voice file is too vague, and the fix is almost always adding better examples rather than more adjectives.

What I do not expect the GPT to do is beat your best work. I expect it to make your median email as good as your good ones, faster. In a channel where Litmus found 21% of marketers cannot state their email ROI at all, “consistently good, with the numbers checked” clears most bars.

When a custom GPT is the wrong tool

I will make the honest comparison, since I use both. Claude Projects does the same standing-setup job with more room for reference material, which matters once your knowledge base outgrows a few tight files. Teams already living in Claude should build there - the instructions template above transfers almost verbatim. And if your sending platform’s built-in AI already drafts in your voice acceptably, a custom GPT may only be worth it for the critique mode, which platform AIs rarely do well.

The setup is also only as current as its files. Mine get a quarterly refresh: new winning emails in, stale benchmarks out, any new voice rule added. A custom GPT with two-year-old examples confidently writes emails for a brand that no longer exists. Put the refresh in the calendar at build time, because nothing about the tool will remind you.

FAQ

Do I need the paid ChatGPT plan? Creating custom GPTs requires a paid plan, and running one well requires it anyway for the stronger models. If email is a core part of your job, the subscription clears its cost on this single build.

Can my whole team use it? Yes - a custom GPT can be shared privately with your workspace, which is where it beats a prompt document. Nobody pastes the wrong version of the voice rules, because the rules live in the tool. Update the files once and everyone drafts from the new version.

Should the GPT write my subject lines too? It drafts them, with the length rule enforced from the instructions. For the campaigns that matter I still run contenders through a dedicated subject line session, because subject lines reward more iterations than a drafting pass gives them.

What about newsletters rather than campaigns? Same build, different knowledge files - a newsletter GPT wants your best past issues and section structure in place of campaign examples. The editorial prompts in my newsletter library slot in as conversation starters.

This build is part of my prompt and workflow library. The voice file comes from the brand voice guide, the always-on rules from the system prompt guide, and the wider practice starts at the complete guide to AI marketing.