How to write prompts for email generation

Most advice about prompting AI is advice about wording. Pick the clever verb, add a role-play preamble, tell the model it's a world-class copywriter, and the good email supposedly falls out. That advice is mostly wrong, and marketing email is where it fails fastest. A model already knows how to write a competent promotional email in the abstract. What it can't guess from your phrasing is your audience, your offer, your voice, and the shape your team actually ships. The quality of what comes back is governed by the context you supply, not the sentence you wrap around the request.

This guide is about supplying that context on purpose, and only the context that earns its place. You'll get a repeatable structure for an email prompt, before-and-after examples you can adapt, and a short method for iterating when the first draft misses. None of it depends on a magic phrase, because there isn't one. The stakes are simple enough: 63% of marketers already use AI tools in email, and 87% of businesses that use AI apply it to email workflows, so the difference between a good draft and a generic one is increasingly the difference between shipping and rewriting.

What to put in an email prompt, and what to cut

An email isn't a paragraph of prose. It is a sender name, a subject line, preview text, a headline, body blocks, one call to action, and a footer, all governed by brand and often by legal and accessibility rules. A prompt that produces that has to carry six kinds of context, ordered context first, task second, constraints third, because the model reads top to bottom and anchors on what it sees early. Carry those six and cut the rest. The temptation is to paste your entire brand guide and hope, which buries the parts that matter under the parts that don't. Pick the version that fits this email.

The audience comes first: who is this for, and what do they already know about you. "Existing customers on the free plan" and "cold prospects from a webinar list" produce different emails, and the model can't infer which you mean. The goal and single action come next: one email, one job. Name the outcome, drive demo bookings, recover an abandoned cart, register for an event, and the one call to action that serves it. Competing links dilute the result, so say one CTA out loud. The offer is the specific thing you're putting in front of the reader: the feature you're announcing, the discount, the event date and time. Vague offers produce vague emails, and if a date or price is involved, put the exact value in so the model doesn't invent one. The tone and brand are where most drafts go generic. Give three or four voice adjectives and one thing to avoid, "confident, plain, a little dry, never exclamation points," or better, point at an approved email and say match the voice and structure of this one. One real example teaches more than a paragraph of description. The structure is the part wording will never produce on its own, so spell it out: subject line under 50 characters, preview text, one hero headline, two body sections, one CTA label, standard footer. Name the shape and you get the shape. The constraints load last: terminology you require or ban, character limits, compliance lines, and accessibility rules like alt text on every image. In enterprise email these aren't optional, and the model has no way to know them unless you say so.

A prompt structure you can reuse

Put those six pieces in a consistent order and the prompt becomes a form you fill in rather than a fresh act of invention each time:

• Audience: who it is for and their relationship to you.
• Goal and action: the one outcome and the single CTA.
• Offer: the specific thing, with exact dates, prices, or names.
• Tone and brand: three or four adjectives plus one to avoid, or a linked example.
• Structure: the blocks you want back, named explicitly.
• Constraints: terminology, limits, compliance, and accessibility.

You won't need every line every time, and that is the point of selecting rather than dumping. A quick internal reminder doesn't need a compliance clause. A customer-facing launch email needs all six. The structure is a starting point, not a formula, and the judgment about which lines matter is yours.

Before and after

The contrast between a wording-led prompt and a context-led one is the clearest way to see it. A wording-led prompt reaches for adjectives and hopes: "You are an expert email copywriter. Write an amazing, high-converting launch email that really pops." Every word there is about the model's persona or the email's vibe, and none of it tells the model anything about your product, your reader, or your brand. What comes back reads like every other AI email, because that is exactly the information you gave it.

A context-led prompt is duller to read and far more useful: "Write a launch email for [Product], announcing [feature] to existing customers. Match the voice and structure of the attached example. Goal: drive demo bookings, one CTA (Book a demo). Subject under 50 characters, no exclamation points. Include preview text and alt text for the hero image." Nothing in it is clever. It names the audience, the goal, the single action, the tone reference, and the structure, and that is why the draft comes back close enough to edit rather than close enough to throw away.

Here is a second pairing for a different job. Weak: "Write a great re-engagement email to win back inactive users." Strong: "Write a re-engagement email for trial users who signed up 60 days ago and never activated. Voice: helpful, plain, no hype. Goal: get them to book a 15-minute onboarding call, one CTA (Book onboarding). Structure: subject under 50 characters, preview text, short empathetic opener, two-sentence value reminder, CTA button, footer. Add alt text for any image." The strong version reads like a spec because it is one, and a spec is what produces a usable email.

How to iterate when the first draft misses

Even a well-built prompt rarely nails it on the first pass, and that is normal. The move is to correct with more context, not stronger adjectives. If the tone is off, don't tell the model to make it punchier; show it, paste a subject line you love and say match this energy. If the structure drifts, name the block that broke: the CTA got buried, move it above the fold and make it a single button. If the copy invents a claim, feed it the real fact and tell it to use only what you provide.

Change one thing per pass so you can see what moved the output. Correcting tone and structure and the offer all at once leaves you guessing which edit did the work. And keep the reader's single action fixed across iterations, because the most common way a draft degrades on revision is a second call to action creeping back in. One caution belongs here plainly: even perfect context doesn't guarantee a correct email. The model will confidently produce a wrong date, a stale price, or a compliance line that doesn't match your current legal language. Verify every fact, offer, and legal line before anything goes out. This human-in-the-loop discipline is exactly why 92% of marketing-ops professionals expect AI to reshape their roles rather than replace them. A thoughtful AI approach to email gets you most of the way, and a human finishes the job.

Where the context comes from

The honest problem with everything above is labor. Pasting brand voice, hunting for the right example, and re-specifying structure on every prompt is friction few marketers sustain past the second week. It tracks with a broader gap: fewer than 30% of marketers feel they have the tools and systems to manage content effectively, and hand-assembling context on every prompt only widens it. The context is real work to assemble, and assembling it by hand each time is where good intentions quietly die.

This is the gap the Knak app for ChatGPT closes. When your assistant is connected to your Knak instance, it can already see your brands, themes, and approved templates, so the context you would otherwise paste is present before you type a word, already selected and already approved. You supply the intent, audience, goal, offer, one action, and the app supplies the brand and structure the model lacks, then returns a real, on-brand email asset inside Knak rather than a block of text you rebuild by hand. The prompt stays simple because the context is automatic, and you refine and ship the result in Knak Studio.

The principle holds no matter which tool you use: name the audience, the goal, the offer, the tone, the structure, and the constraints, then iterate with more context and fewer adjectives. Carry what the email needs and leave the rest out. If you want the context handled for you so the prompt can stay short, see how the Knak app for ChatGPT builds on-brand email into your workspace at knak.com/mcp.