How to Craft AI Prompts for a Marketing Masterpiece: Unleashing Creativity

Look, I’ve sat in enough content strategy sessions to know what’s happening. Someone types “write me a marketing email for our summer sale” into ChatGPT. Gets three paragraphs of beige corporate copy. Thinks AI doesn’t work for marketing. Moves on.

It’s not the AI. It’s the brief. A bad brief to a human copywriter produces bad copy. Same with AI — except the AI won’t push back, ask clarifying questions, or tell you the brief is garbage. It just writes confidently in whatever direction you pointed it.

The fix is mechanical, not magical. You already know what a good creative brief looks like for a human. AI needs the same thing: role, context, audience, constraints, format, success criterion. Spelled out. Every time.

Farfetch — luxury fashion marketplace, millions of subscribers — wanted to lift email engagement without losing their brand voice. Specifically their problem: they needed subject lines and body copy that felt like Farfetch wrote them, not like an AI wrote them and someone approved it in 30 seconds.

They used Phrasee, an enterprise AI platform built for exactly this — it learns the brand language over time by A/B testing different styles, tones, and phrases against real audience behavior. Not a single prompt, then done. An iterative loop between AI generation and performance data.

“The +31% open rate on trigger emails isn’t a subject line win — it’s a brief quality win. The AI knew which campaign type, which audience state, which brand constraints to operate within.”

Editorial synthesis — sources: Chain Store Age Farfetch/Phrasee case (2022); Retail TouchPoints generative AI retail analysis; Farfetch VP Matthias statement

The reason trigger emails outperformed broadcast by such a wide margin is worth understanding. Trigger emails have a defined audience state — this person just abandoned a cart, they showed intent, they’re warm. That context feeds directly into the prompt. A prompt with known context almost always outperforms a prompt aimed at a broad segment. The AI had more to work with.

These aren’t theoretical. They’re built on the same structure the Farfetch case demonstrates: role, audience state, brand constraint, format, test criterion. Swap in your context.

The label-the-hook instruction is the part most people skip. It forces the AI to be intentional rather than random, and it gives you a framework for evaluating the output without just going on gut feel. You’re not picking the one that “sounds good.” You’re picking the one with the right emotional hook for this audience state.

The three-variation structure matters. If you get one ad, you have one hypothesis. If you get three with different lead angles, you have a real test. The contrarian variation is the one most teams never write — and it’s often the one that performs.

The sentence length instruction sounds pedantic. It isn’t. Uniform sentence length is the fastest way to make AI copy detectable — and unpleasant to read. Mixing short punches with longer explanatory sentences is basic craft. You have to tell the AI to do it, because by default it doesn’t.

05 — The “why isn’t this working” diagnostic prompt

This one’s different. Not for generating copy — for figuring out why existing copy is failing.

That last instruction — “do not compliment the email before critiquing it” — sounds petty. It’s not. By default, AI will open with “This is a well-crafted email that…” and spend two sentences saying so. You don’t need the compliment. You need the diagnosis. Tell it to skip the preamble.

What the data actually says about AI in email marketing

Quick note on the numbers floating around in this space: some of them are wild. “250-320% increases in email engagement.” Sure. Directional, unverified, vendor-sourced. I’m not using those.

What I’ll use: Sources: Humanic AI analysis of 2024-2025 statistics aggregated from Tabular Email Marketing Stats, Wix Email Marketing Statistics; Artsmart AI compilation AI-driven email personalization is associated with a 13% uplift in click-through rates in consistent cross-study data, and a 41% revenue increase in some verticals where personalization was implemented at the campaign level, not just subject line level. Self-reported metric compilations, not independently audited RCTs — treat as directional

The Farfetch numbers (+31% opens, +37.9% CTR on triggers) are the most credibly sourced of any single case study in this space — they appeared in Chain Store Age quoting a Farfetch VP directly, which is closer to primary than most “case studies” you’ll find.

The failure no one writes about

You want the other side. Fine.

The failure mode I see repeated: marketing teams treat AI as a production shortcut instead of a brief-following copywriter. They write three-word prompts, get three paragraphs of glossy nothing, publish it anyway because the calendar says it ships today. Open rates tank. They conclude AI marketing doesn’t work.

The measurement problem compounds this. Source: DigitalApplied email subject line analysis, 2026; Mailchimp 2026 Email Marketing Benchmark Report referenced therein Apple Mail Privacy Protection — active on 50%+ of email clients as of 2026 — pre-loads tracking pixels and registers opens regardless of whether a human actually read the email. This inflates open rates by 15-30% depending on your audience’s Apple device share. If your list is 60% Apple Mail users and your AI email “improved open rates by 18%,” a chunk of that is measurement noise, not copy quality.

And one more: AI can write confident-sounding copy that is factually wrong about your product. I’ve seen product descriptions generated with hallucinated specs — materials that don’t exist in the product, sizing that contradicts the actual range. Every AI-generated output needs a product-knowledge check before it ships. Not an optional step. A required one.