Search for ChatGPT email marketing prompts and you get a thousand variations of the same thing: “write me a cold email to a marketing director.” Fine. You’ll get an email. You’ll also get an email that has no idea what comes after it, what came before it, or why it exists.
Campaigns are not copy. Campaigns are architecture — a set of steps, each with a job, spaced across days, escalating or de-escalating pressure, with a defined exit. The copy is the last thing you write, not the first. And this is the part people miss: ChatGPT is genuinely good at architecture if you prompt it as a planner instead of a copywriter. Ask it for an email and it writes an email. Ask it for a sequence design and it will argue with you about step spacing, which is exactly what you want.
Here’s the prompting method I use to go from a blank doc to a full multi-step campaign, plus the actual prompts at each stage.
The campaign-prompting method: brief, then architecture, then copy
Three passes, in this order. Skipping the brief is why most ChatGPT email output sounds like it was written for nobody — because it was.
- Brief. One master prompt that loads the model with your ICP, offer, proof and objections. Everything downstream inherits it.
- Architecture. Steps, timing, and a one-line job per step. No copy yet. Argue here, because fixing structure is cheap and rewriting five emails is not.
- Copy. One prompt per step, each referencing the approved architecture so step 3 knows what step 1 already said.
Keep all three in a single chat. The context is the asset. Starting a new chat for each email is how you end up with four emails that all open with the same “I noticed you’re scaling your team” line.
Pass 1: the master brief prompt
Run this once. Everything else in the thread builds on it.
You are acting as a senior B2B email strategist. I’m going to brief you on a campaign, and you’ll hold this context for everything that follows. Do not write any copy yet — reply with a summary of what you understand plus the three things you’d most want clarified.
ICP: [job title, company size, industry, region]
What they’re currently doing about this problem: [status quo]
Offer: [what I sell, in one sentence, no adjectives]
Proof I can honestly use: [case studies, named clients, specific outcomes I can substantiate — do not invent any others]
Top three objections I hear: [1, 2, 3]
What a “win” looks like: [reply / booked call / trial signup]
Constraints: [tone, banned words, compliance rules, length limits]
That last instruction — do not invent any others — matters more than the rest of the prompt combined. Left unconstrained, ChatGPT will happily manufacture a statistic to make your value prop land harder, and you will not always catch it before it ships.
Pass 2: the sequence architecture prompt
Still no copy. You want a table you can push back on.
Using the brief above, design a [4/5/6]-step email sequence. For each step give me: step number, days after previous step, the single job of that email (one sentence), the psychological posture (curious / helpful / challenging / releasing), the CTA type (soft question / resource / direct ask / breakup), and what should trigger skipping this step entirely.
Output as a table only. Then, below the table, list the two structural weaknesses in this sequence and what you’d change if the reply rate on step 1 came back near zero.
The self-critique at the end is the highest-leverage sentence in this whole article. Models default to confident output; explicitly asking for the failure modes gets you a second, more honest draft for free.
Pass 3: per-step copy prompts
One prompt per step, each anchored to the approved architecture.
Step 1 — the opener
Write step 1 from the approved architecture. Constraints: under 90 words, no greeting fluff, no “hope this finds you well,” no compliment about their company. Open with the observation or problem, not with me. The CTA is a single low-friction question. Write the body only — I’ll do subject lines separately. Then explain in one line why this opener earns the second sentence.
Step 2 — the value-add
Write step 2. This email must not reference the fact that I emailed before and must not ask “did you see my last note.” It stands alone and gives one concrete, useful thing — a specific tactic, a benchmark, a teardown observation — that’s valuable even if they never reply. Under 100 words. End with a question that’s easier to answer than to ignore.
Step 3 — the objection flip
Write step 3. Take objection #2 from my brief, name it out loud in the first sentence before they have to, and then reframe it. Do not resolve it with a claim I can’t substantiate — resolve it with a question, a reframe, or a concession. Under 110 words. Tone: peer-to-peer, slightly blunt, zero salesmanship.
Step 4 — the breakup
Write the final step: a release email. No guilt, no “I’ll assume you’re not interested,” no fake deadline. Give a genuine exit, restate in one clause what I’d have helped with, and make it easy to say “not now, try me in Q3.” Under 70 words.
The supporting prompts
A/B variant prompt
Variants are only useful if they differ on one axis. Otherwise you learn nothing.
Take the approved step [N] and produce variant B. Change exactly one variable: the [opening angle / CTA type / length / proof element]. Hold everything else — tone, structure, offer, word count band — constant. Below both versions, state in one sentence precisely what this test measures and what result would make you keep variant B.
Subject-line batch prompt
Generate 15 subject lines for step [N]. Rules: 2–5 words, lowercase, no colons, no questions in more than three of them, no curiosity-gap clickbait, nothing that would look out of place from a colleague. Group them into three buckets — literal, oblique, and reference-to-their-world — then flag the two you’d actually send and why.
Tone-calibration prompt
This is how you stop every campaign sounding like the same LinkedIn ghostwriter.
Here are three emails I’ve written myself that got replies: [paste]. Extract my voice as a rule set — sentence length, punctuation habits, how I open, how I hedge, what I never say. Then rewrite the full sequence to match those rules. Show the rule set first so I can correct it before you rewrite anything.
Reply-triage prompt
Once the campaign runs, replies are the real work. Use a prompt as a sorting layer, not an auto-responder.
Classify each reply below into: interested, not now (with a date), wrong person (with a referral), objection (say which one), or hard no. For each, draft a two-sentence response in my voice. Do not send anything — output drafts only, and flag any reply where you’re less than confident in the classification.
From copy to execution
Here’s where most people’s ChatGPT workflow stalls. You’ve got a beautiful five-step sequence in a chat window and now you’re copy-pasting it into a campaign builder, mapping delays by hand, re-pasting subject lines, and re-uploading a prospect list.
That gap is closing. With MCP — the open standard that lets an assistant like ChatGPT or Claude call an external product’s tools directly — the model can build the campaign in the platform rather than in a text box. In plain language you can say: create the sequence I just designed, verify these addresses, enroll the ones that come back valid, and start it.
The important part is what the agent can’t do. The agent never sends a message itself and can never raise a sending limit. Creating and launching a campaign just writes it and hands it to the platform’s scheduler, which paces every send inside safe caps and a gradual ramp — identically whether a human clicked the button or an agent asked for it. That division is what makes agent-run outreach viable at all: the creative layer is fast and conversational, the sending layer stays boring and governed.
Where WarmySender fits
WarmySender is the execution layer for exactly this workflow. Connect it to ChatGPT, Claude, Cursor, Codex or any agent that speaks MCP, and your assistant can create cold email, LinkedIn and Instagram campaigns, verify addresses, enroll prospects, launch and pause, and read back the stats — while WarmySender’s scheduler paces every action within safe limits regardless of who’s driving. Connecting and disconnecting accounts stays in the app, where it belongs. It’s self-service: you build it, the agent helps, the scheduler keeps it sane.
FAQ
How many steps should a ChatGPT-designed sequence have?
Ask the model to justify the count against your ICP rather than picking a number first. As a starting frame: shorter sequences for senior buyers with obvious pain, longer for categories where you’re doing education. Then test — the architecture prompt above gives you the table to change one variable at a time.
Will ChatGPT-written emails sound generic?
Only if you skip the tone-calibration pass. Feeding it three of your own replies-that-worked and having it extract a rule set before rewriting is the single biggest quality jump available. Generic output is almost always an under-specified prompt, not a model limitation.
Can I let an agent launch campaigns unsupervised?
You can let it build and launch, but review the copy first — you’re accountable for what goes out under your name. The safety model isn’t “trust the agent’s judgment,” it’s that the agent physically cannot send a message or raise a limit; the platform’s scheduler owns pacing either way.
Start building
Copy the seven prompts above into a doc, run them in order in one chat, and you’ll have a full campaign — architecture, copy, variants, subject lines and a triage layer — in about an hour. Then wire your assistant to WarmySender so the last mile stops being copy-paste.