How To Prompt An AI Email Builder Properly

An AI email builder can turn a rough campaign idea into a workable layout and first draft remarkably quickly. It cannot decide whether the campaign is welcome, relevant or appropriate for the people who receive it. Nor can it reliably distinguish a substantiated product claim from an optimistic one, or recognise a detail that has changed since the source material was written.

The quality of the prompt therefore matters, but “write a friendly promotional email” is not really a prompt. It is a request for the system to make a series of important decisions without enough information: who the recipient is, why they are receiving the message, what they already know, what may be promised, and what the email needs to achieve.

A better approach is to treat the builder as a fast drafting assistant. Give it a clear brief, trustworthy inputs and firm boundaries. Ask for a defined email artefact rather than generic “copy”, then review the output as carefully as any other proposed customer communication. This is consistent with the broader prompting guidance from OpenAI’s prompt engineering best practices: clear, specific instructions and iterative refinement produce more useful results than vague instructions.

Start with the job the email must do

Before writing a prompt, define the email’s operational purpose in one sentence. This stops the builder from blending several messages into one overlong campaign.

For example:

• Help a new subscriber choose the right plan after joining a permission-based list.
• Remind an existing customer that a trial ends in three days and show the relevant upgrade route.
• Announce a product update to active users, with one clear action to try the feature.
• Invite opted-in customers to a local event, making the location and attendance requirement unambiguous.

“Increase conversions” is a business outcome, not a sufficient job description. An email can support that outcome, but the prompt still needs to say what the recipient should understand or do after reading it.

Be equally clear about the message category. A service notice, a receipt, an onboarding email and a promotional offer may share a template, but they deserve different tone, urgency and calls to action. Combining a genuinely necessary account notification with a sales message can also create an awkward recipient experience. Separate messages are usually easier to explain, measure and suppress correctly.

Give the builder the information it cannot safely guess

A good prompt supplies a compact campaign brief. The exact fields vary, but the following information is usually more valuable than adding dramatic adjectives such as “exciting” or “punchy”.

Use approved source material, rather than asking the builder to research or improvise. A public product page, a current offer document, a release note and an internally approved list of claims make a much better foundation than “promote our new analytics feature”. Some email platforms explicitly support generating content from supplied text or a URL; for example, Mailchimp’s guidance on AI-generated email content recommends having source materials such as website URLs, product descriptions and relevant images ready before generation. Even then, the output should be checked against the original source.

Use constraints to protect accuracy and brand judgement

Constraints are instructions about what the builder must not do. They are not an admission that the technology is unreliable; they are how a marketing team documents decisions it has already made.

Useful constraints include:

• Use only the factual claims in the supplied source material. If a claim is unsupported or missing, insert [VERIFY] rather than inventing it.
• Do not create a discount, deadline, availability claim, customer quote, statistic or guarantee.
• Do not imply that the recipient has viewed, bought or abandoned something unless that event is confirmed in the campaign data.
• Do not use false urgency, misleading reply-style subject lines or language designed to obscure that the message is promotional.
• Keep the CTA consistent with the destination page. Do not label a link “Start free” if payment details are required before access.
• Return copy only; do not alter unsubscribe language, preference-centre links, sender details or other controlled footer content.

That last point is particularly important in an email builder. Footer and subscription components should normally be controlled by the sending platform or template, not regenerated for every campaign. For marketing email sent at volume to personal Gmail accounts, Google requires bulk senders to support one-click unsubscribe and include a clear unsubscribe link in the message body, alongside authentication and other requirements. Google’s email sender guidelines also recommend keeping user-reported spam rates below 0.1% and avoiding 0.3% or higher. Better prompting helps with clarity and relevance, but it cannot compensate for weak permission practices, poor segmentation or excessive frequency.

Ask for the parts of an email, not a wall of copy

Email is a constrained format. The subject line competes in the inbox; the preheader may add context; the opening needs to orient the reader; and the message must work when images are blocked or when it is scanned quickly on a small screen. A prompt that asks for “an email” often returns a long, article-like block that then has to be dismantled.

Specify the elements you need and their limits. Limits should serve the design and reading experience, rather than being treated as universal rules. For a simple promotion, you might request:

• five subject lines, each under 45 characters, with no emojis;
• five preheaders that add new information rather than repeat the subject line;
• one headline of up to eight words;
• a two-paragraph body with one product benefit per paragraph;
• one CTA label of two to four words;
• a short image direction for the designer, without text embedded in the image; and
• plain-text copy that preserves the offer, destination and required disclosures.

For a lifecycle email, the right output might instead be a reassuring subject line, a concise explanation of what happened, a list of what the recipient can do next and a support route. The correct structure follows the customer’s situation, not the capabilities of the tool.

Show the builder what “on brand” means

Brand voice guidance works best when it is concrete. “Use our warm, human tone” means different things to different writers and systems. Give practical instructions that a reviewer could apply:

• Write in British English.
• Use short sentences and everyday vocabulary; define technical terms when they are necessary.
• Prefer specific product outcomes to superlatives.
• Do not use exclamation marks, “revolutionary”, “game-changing”, “unlock”, “seamless” or “don’t miss out”.
• Address the reader as “you”; avoid claiming to know how they feel.
• Use sentence case for headings and CTA labels.

One or two approved examples are often more effective than a long list of adjectives. Label them clearly: “Here are examples of our preferred rhythm and level of formality. Do not reuse their product facts, offers or calls to action.” This gives the builder a style reference without allowing old campaign details to leak into a new message.

Keep the examples current. A successful campaign from three years ago may contain pricing, positioning or promises that no longer apply. Teams using a repeatable prompt should version it, record who owns it and update the brand examples and prohibited phrases as campaigns evolve. In production settings, prompts benefit from the same disciplined review as other operational assets; OpenAI’s prompting guide similarly recommends clear instructions, examples and systematic testing.

A prompt template for a campaign email

The following template is deliberately plain. It can be pasted into most email builders and adapted to their available controls.

Create a marketing email draft for the campaign below.

Audience and permission context:
[Who will receive this email, how they subscribed, and relevant lifecycle stage]

Campaign purpose:
[The one job the email should do]

Primary action:
[The action and destination]

Verified source material:
[Paste approved product facts, offer terms, dates, eligibility, URL and required wording]

Brand voice:
[Practical voice rules and one or two labelled style examples]

Content requirements:
- Provide 5 subject lines, each [limit].
- Provide 5 preheaders, each [limit], which do not repeat the subject line.
- Provide a headline, body copy, one CTA label and plain-text version.
- Use British English.
- Make the primary action clear in the first half of the email.

Constraints:
- Use only verified source material.
- Do not invent claims, prices, deadlines, statistics, testimonials or availability.
- Do not imply behaviour or purchase history unless stated above.
- Do not change controlled footer, subscription or legal content.
- If important information is unavailable, mark it [VERIFY].

Return the output with clear labels for each email component.

For a triggered email, add the event that caused it, the allowed delay, the relevant suppression rules and the information the recipient has already received. “Write a cart-abandonment email” is not enough. State whether the basket is reserved, whether the customer may have completed their purchase elsewhere, whether stock is live and whether the person has received a similar reminder recently. The point is not to make the prompt enormous. It is to include the facts that change what would be responsible to say.

Do not put sensitive customer data into a casual prompt

Personalisation can be useful, but a content brief rarely needs a spreadsheet of names, email addresses, purchase histories or support tickets. Use audience descriptions and approved merge fields where possible: {{first_name}}, {{plan_name}} or {{renewal_date}}. Send the minimum information needed to create the copy, and ensure the chosen tool, account configuration and internal process have been assessed for the data involved.

That is especially relevant where prompts contain personal data, confidential commercial material or information drawn from customer service interactions. The UK Information Commissioner’s Office has published guidance on AI and data protection; organisations should apply their own data-protection, security and supplier-governance requirements before sharing personal data with any service.

It is also sensible to prevent user-entered or imported content from silently changing the task. If the builder summarises reviews, survey comments or support notes, treat that text as source material to be analysed, not as instructions to follow. Ask the system to ignore instructions embedded in supplied content and to report uncertain or conflicting material for human review.

Review the draft in the context in which it will be sent

Generating copy is the beginning of the production process, not the end. Review subject line, preheader, sender name, body, CTA and landing page together. An accurate body can still be undermined by a misleading subject line, an out-of-date linked page or a button that does not match the promised action.

At minimum, a human reviewer should check:

• Factual accuracy: every price, date, claim, product capability, quote and eligibility condition matches an approved source.
• Audience fit: the email makes sense for the segment and excludes people for whom it would be irrelevant, sensitive or repetitive.
• Permission and pressure: the send respects consent, preferences, suppression lists and sensible contact pressure. An elegant message sent too often is still unwelcome.
• Accessibility and rendering: the hierarchy works on mobile, links are descriptive, images have appropriate alternative text and the essential message does not depend on an image.
• Deliverability hygiene: the campaign uses the correct authenticated sending identity, visible unsubscribe route and established sending practices.

Test the actual rendered email, not just the copy in the builder. A line that looks concise in a generation panel may wrap badly in a narrow inbox view; a preheader may be consumed by boilerplate; and personalisation may produce an empty greeting. Send test messages using realistic merge data and check the landing page before the campaign goes live.

If a sender is already experiencing spam-folder placement, rising complaints or provider throttling, producing more variations of copy may be the wrong immediate priority. Investigate the audience, recent send volume, authentication, campaign mix and engagement patterns first. This is the point at which an independent deliverability assessment, such as Email Foundry’s work for permission-based marketing and lifecycle senders, can be more useful than another creative iteration.

Improve prompts through controlled iteration

When a draft misses the mark, avoid responding with “make it better”. Say what failed and what must change: “The opening assumes the recipient knows this feature. Rewrite it for new users without adding features not listed in the source”; or “The CTA is too forceful for a renewal reminder. Offer a preference-management route as a secondary option.”

Change one or two instructions at a time, save the versions that work and build a small library of approved prompt patterns for common programmes: welcome, onboarding, product update, replenishment, renewal and re-engagement. Measure the campaign itself, not merely whether the generated prose sounds polished. Clicks, conversions, unsubscribes, complaints, replies, provider feedback and longer-term engagement all provide evidence about whether the message was useful to the people who received it.

The best prompt is not the most elaborate one. It is the one that gives the builder enough truthful context to produce a reviewable draft, while keeping important customer, compliance and deliverability decisions with the people accountable for sending the email.