AI prompts for email marketing

Subject lines: prompt for the character window, not just the hook

Most email clients truncate subject line previews well before the full line, with roughly 40-60 characters visible on desktop clients like Gmail and Outlook, and considerably less — often 25-35 characters — on mobile inbox previews. A subject line that reads brilliantly at 80 characters may show as an incomplete, confusing fragment on the device most opens actually happen on.

The instruction against misrepresentation is not a style note — the CAN-SPAM Act specifically prohibits deceptive subject lines, and a subject line promising something the email body does not deliver is both a compliance risk and a fast way to train subscribers to stop opening future emails.

Preview text is a second headline most prompts forget

The preview text — the snippet shown after the subject line in most inbox views — is set separately from the subject line but frequently left as the email's opening sentence by default, wasting a second opportunity to earn the open. A deliberate preview text prompt treats this field as its own creative decision.

Pairing a curiosity-driven subject line with benefit-driven preview text, or vice versa, tends to outperform two fields saying the same thing, because the combination gives the recipient two separate reasons to open rather than one reason stated twice.

CAN-SPAM: what a prompt can never override

The CAN-SPAM Act, enforced by the FTC since 2003, sets requirements no amount of clever prompting should work around: accurate "From," "To," and routing information; a subject line that does not mislead about the email's content; clear identification that the message is an advertisement where applicable; the sender's valid physical postal address; and a working, honored opt-out mechanism that must be processed within 10 business days of a recipient's request. As of 2024 FTC guidance, penalties for violations run up to $53,088 per separate email in violation.

None of these requirements live inside the email copy itself in a way a copywriting prompt would naturally produce — they are structural and platform-level (footer content, list management, sender authentication), which means the discipline here is less about prompting differently and more about never asking a model to write copy that omits or obscures the required unsubscribe link and sender information your email platform is responsible for including.

CASL and international lists: opt-in is not opt-out

Canada's Anti-Spam Legislation, in effect since 2014, requires express or implied consent before sending commercial email to a Canadian recipient — a meaningfully stricter standard than CAN-SPAM's opt-out model, which permits sending until someone unsubscribes. A prompt template built around US compliance assumptions will not flag this difference, because a model has no inherent awareness of which subscribers on your list are in Canada or the EU, where GDPR imposes its own, separately stringent consent requirements.

This prompt is deliberately a checklist reminder rather than a task the model can complete on its own, because consent verification requires access to real list and consent-tracking data no prompt can supply.

Welcome sequences: use the actual signup trigger, not a generic new-subscriber frame

A welcome sequence written from "someone just joined our list" produces generic onboarding copy. A welcome sequence written from the actual trigger — downloaded a specific guide, abandoned a specific cart, signed up via a specific referral partner — produces copy that continues a conversation the subscriber already started, which performs measurably better because it matches their actual, recent intent.

The instruction not to invent a discount matters because a model asked to write a "compelling" welcome email 3 will often manufacture a limited-time offer that does not exist, which is both a deceptive-content risk under CAN-SPAM and a trust problem the first time a subscriber notices the "limited time" offer is still running two months later.

Personalization beyond a first name requires real data, or it becomes invention

Merge-tag personalization — inserting a subscriber's first name — is table stakes and easy to prompt for. Genuine behavioral personalization, referencing what a subscriber actually browsed, purchased, or clicked, drives measurably stronger engagement, but only when the underlying data is real; a model asked to "personalize this email" with no data attached will produce plausible-sounding but fabricated behavioral claims.

Restricting the model to supplied data is the same discipline that prevents fabricated statistics elsewhere — a personalization email referencing behavior the subscriber did not actually engage in is not just inaccurate, it is the specific kind of error that makes a subscriber feel surveilled and misrepresented at once.

A useful middle ground exists between full behavioral personalization and generic copy: segment-level personalization based on real aggregate data, such as "subscribers who purchased in the last 90 days" or "subscribers who signed up via the pricing page," which requires far less granular data per individual than full behavioral tracking while still outperforming completely generic copy sent identically to an entire list.

Re-engagement emails: name the actual gap, not a generic "we miss you"

A re-engagement or win-back email addressed to "we haven't seen you in a while" performs worse than one referencing the specific inactivity window and, where available, what the subscriber previously engaged with before going quiet — because the generic version could apply to any subscriber on any list, while the specific version demonstrates the sender actually tracked the relationship.

Offering a frequency-adjustment option alongside the win-back message is worth including explicitly, since it gives disengaged subscribers a middle path between full re-engagement and unsubscribing, often recovering more of the list than an all-or-nothing message does.

Testing subject lines: check the sample size before trusting a "winner"

A/B testing subject lines is a genuinely valuable practice, but the result is only meaningful with sufficient list volume — testing two subject lines on a 300-person segment and declaring one the winner based on a handful of extra opens is drawing a conclusion the sample size cannot support. Most email platforms' own statistical significance calculators, or a simple significance calculator applied to open-rate data, will confirm whether a result crossed a meaningful threshold before it gets treated as a durable insight.

Asking the model to flag insufficient sample size rather than assume significance is the same discipline used elsewhere in prompting: a confident-sounding conclusion is not evidence the underlying data supports it.

Deliverability: copy choices that trigger spam filters

Certain copy patterns — excessive exclamation points, all-caps subject lines, phrases historically associated with spam ("FREE," "ACT NOW," "100% guaranteed"), and a poor text-to-image ratio — can trigger spam filter scoring even when the email is entirely legitimate and compliant. A model with no constraint will sometimes reach for exactly these patterns because they read as attention-grabbing in isolation.

This review does not guarantee inbox placement, since deliverability also depends on sender reputation, authentication (SPF, DKIM, DMARC), and list hygiene factors entirely outside the copy itself — but it removes the content-level risk factors that are within a copywriter's control.

A complete sequence prompt, combining the pieces

Each constraint here targets one of the specific failures this page walks through: the real trigger keeps the copy from reading as generic, the no-fake-urgency rule closes off a deceptive-content risk, and naming CAN-SPAM keeps compliance visible even though the actual opt-out mechanism belongs to the platform template rather than to anything the model writes.

What to verify before sending any AI-drafted campaign

Four checks apply to every send. First, confirm the subject line accurately represents the email's actual content — no promise the body does not deliver. Second, confirm any discount, deadline, or scarcity claim is real and currently accurate, not a plausible-sounding invention. Third, confirm the unsubscribe link and sender postal address render correctly in the actual template, since these are CAN-SPAM requirements the copy prompt does not control. Fourth, for lists including Canadian, EU, or other opt-in-jurisdiction subscribers, confirm consent records are current before the send goes out.

A fifth check worth adding for any sequence longer than a single email: read the full sequence in order, not just each email in isolation, since a claim made in email 1 ("this offer is not available anywhere else") can be quietly contradicted by email 3 if the two were drafted in separate prompts without cross-referencing each other, and a reviewer checking emails one at a time will miss the inconsistency a subscriber reading the full sequence would catch immediately.

Newsletter and recurring content: keep the format stable, vary the substance

A recurring newsletter benefits from a predictable structure subscribers learn to expect — the same section order, roughly the same length, a recognizable sign-off — while the actual content inside that structure should vary meaningfully issue to issue. AI-assisted newsletter drafting works best when the prompt separates these two things explicitly, supplying the fixed format once and the week's actual content fresh each time.

The explicit permission to run with 2 sections rather than fabricate a third matters in practice — a model asked for exactly 3 sections with only 2 real inputs will often manufacture a plausible-sounding third topic rather than flag the gap, which is the newsletter-specific version of the same invention problem that shows up whenever a prompt has a quota it must fill regardless of whether the underlying material exists.

When a template beats a fresh prompt every time

Transactional and lifecycle emails that must be legally precise — order confirmations, shipping notices, password resets, billing receipts — are usually better served by a fixed, legally-reviewed template than by a fresh AI-generated draft each time, because the specific wording of a receipt or a data-related notice can carry compliance weight (accurate pricing disclosure, correct tax treatment, required regulatory language) that benefits from being locked once and reused, rather than regenerated with the risk of drift.

AI is best applied to the marketing layer sitting on top of these transactional emails — a friendly note accompanying a shipping confirmation, for example — while the transactional core itself stays fixed and reviewed.

The distinction has a useful rule of thumb: if getting a single word wrong in this email could create a legal, financial, or safety problem — the wrong tax amount on a receipt, an inaccurate delivery promise, a misstated cancellation deadline — treat it as transactional-core content and lock it down. If getting a word wrong would only make the email slightly less engaging, treat it as marketing-layer content and let AI iterate on it freely.