I thought abandoned-cart emails would be easiest—they weren't
I run a small Shopify store selling hand-poured candles, and last November I decided to let AI write my entire Cyber Monday email campaign. Not just one or two messages—I wanted three complete sequences: abandoned-cart reminders, VIP early-access announcements, and last-chance sale pushes. Eighty emails total across three customer segments. I'd been using Claude Sonnet 4.5 for product descriptions and it felt conversational enough, so I figured email would be simple. Write a subject line, write some body copy, export to Klaviyo, done.
It was not simple. The first sequence I generated—abandoned-cart emails for people who'd left a $40 candle in their cart during Black Friday—had open rates 11% lower than my usual template. The subject lines were fine. The body copy was fine. But "fine" doesn't move someone from browsing on their phone at 9 PM to pulling out a credit card. I scrapped that batch, tested Gemini Pro and GPT-OSS 120B on the same prompts, and discovered that the model you pick matters less than the segment you're writing for. Claude wrote better VIP emails. Gemini wrote better urgency-driven last-chance messages. And GPT wrote the worst abandoned-cart emails I've ever seen, which taught me more about prompt structure than the two that worked.
Here's what I tested, which model won for each goal, and the three prompt templates that actually drove clicks.
I started with Claude because I liked its tone—bad reason
My first prompt was too vague. I opened Claude Sonnet 4.5 in Kryotta and typed: "Write five abandoned-cart emails for a candle store. Cyber Monday sale. Friendly tone. Include subject lines."
Claude gave me five emails. They were polite, warm, and utterly generic. Subject lines like "Still thinking about that candle?" and "Your cart misses you 🕯️" that could've come from any store selling anything. The body copy mentioned the sale, mentioned the product, and ended with a discount code. I exported them to Klaviyo, scheduled them for the week leading up to Cyber Monday, and sent them to 340 people who'd abandoned carts between November 20th and 24th.
Open rate: 18%. My normal abandoned-cart template—one I'd been using for six months—averaged 29%. Click-through rate was 4.1%, down from 7.8%. I made $340 in recovered revenue, which sounds fine until you realize I'd normally make closer to $900 from that same list.
The problem wasn't Claude. The problem was that I'd asked for "friendly tone" without explaining what friendly meant for someone who'd already shown intent. Abandoned-cart emails aren't cold outreach. The person already wanted the thing. They got distracted, or they wanted to compare prices, or they were waiting to see if I'd offer a discount. My prompt didn't tell Claude any of that, so it defaulted to the same cheerful brand voice it would use for a welcome email.
I rewrote the prompt with three details that changed everything
I tried again. This time I gave Claude the context it needed:
"You're writing an abandoned-cart email for someone who added a $40 soy candle to their cart during Black Friday but didn't check out. They've bought from us before—average order value $52. They're price-sensitive but not bargain-hunting; they care about quality and they've opened 60% of our emails in the past three months. Write three emails: one sent 2 hours after abandonment (gentle nudge, no discount), one sent 24 hours later (mention the sale ends soon, still no discount), one sent 48 hours later (offer 15% off, urgent tone). Subject lines should reference the specific product if possible. Body copy should be under 80 words. Include one concrete sensory detail about the candle in each email."
Claude Sonnet 4.5 gave me three emails that felt like I'd written them. Subject lines: "That amber & sage candle is still waiting," "Cyber Monday ends tomorrow—your cart's ready," and "Last chance: 15% off the candle you picked." The body copy mentioned the scent, reminded them they'd already chosen it, and made the sale feel like a deadline they'd set for themselves, not one I was imposing. I sent these to a different segment—280 people who'd abandoned carts between November 25th and 27th.
Open rate: 32%. Click-through: 9.1%. I recovered $1,180 in revenue. The sensory detail—"the amber & sage candle you picked burns for 50 hours and smells like the first cold morning of the year"—showed up in two reply emails from customers who said that line was what made them come back.
Gemini crushed the last-chance urgency emails—and I didn't expect it
I moved on to the last-chance sequence: five emails sent in the final 36 hours of Cyber Monday, targeting people who'd opened at least one sale email but hadn't bought. This segment is harder. They're aware of the sale, they're not ignoring you, but something's stopping them from converting. Maybe they're waiting for a better discount. Maybe they're comparing you to three other stores. You need urgency without desperation, and you need to make the deadline feel real.
I tested the same prompt on Claude, Gemini Pro, and GPT-OSS 120B. The prompt:
"Write five last-chance emails for Cyber Monday. Sent to people who opened our sale announcement but haven't bought. Sale ends at 11:59 PM PT on Monday, November 27th. These people are on the fence—price matters, but so does missing out. Tone: urgent but not panicky. No fake scarcity (we're not lying about stock levels), but make the deadline feel real. Subject lines under 50 characters. Body copy under 70 words. Include the discount code in every email. No emojis in subject lines."
Claude gave me solid emails. Professional, clear, deadline-focused. Gemini gave me emails that felt like a friend texting you at 10 PM to say the thing you wanted is about to sell out. GPT gave me five emails that all said essentially the same thing in slightly different words, which is the worst kind of sequence because by email three, people are annoyed.
I sent Claude's batch to 400 people. Gemini's batch to another 400. GPT's batch to 200 (smaller test group because I had a hunch it wouldn't work).
Claude's results: 24% open rate, 6.2% click-through, $890 revenue.
Gemini's results: 31% open rate, 8.7% click-through, $1,340 revenue.
GPT's results: 19% open rate, 3.1% click-through, $180 revenue.
Gemini's subject lines were shorter and more direct: "11 hours left," "Tonight at midnight," "This is it." Claude's were slightly more formal: "Cyber Monday ends tonight," "Final hours: 30% off," "Last call for 30% off sitewide." The difference in open rates suggests that when someone's already engaged and on the fence, they don't need a full sentence—they need a countdown.
VIP early-access emails worked best with Claude's formality
The third sequence was VIP early-access: seven emails sent to my top 120 customers (people who'd spent $200+ in the past year) giving them 36 hours of exclusive access to the Cyber Monday sale before I announced it publicly. This segment is tricky because you're trying to make them feel special without sounding like you're begging. They've already proven they'll buy from you. You're not selling the product—you're selling the exclusivity.
I only tested Claude and Gemini here, because GPT had already failed badly enough that I wasn't giving it another $15 in token spend. The prompt:
"Write seven VIP early-access emails for Cyber Monday. Sent to customers who've spent $200+ in the past year. They get 36 hours of exclusive access to a 30% discount before we announce it publicly. Tone: You're letting them in on something valuable, not begging them to buy. Make them feel like insiders. Subject lines should emphasize early access, not the discount. Body copy under 90 words. Mention that this offer isn't going to the full list until Thursday. No countdown timers in the copy—these people don't need fake urgency."
Claude's emails leaned into the exclusivity. Subject lines: "You're in early (30% off starts now)," "Before anyone else: Cyber Monday preview," "VIP access live—shop now." The body copy thanked them for being loyal customers and framed the early access as a reward, not a sales tactic.
Gemini's emails were warmer but less formal. Subject lines: "You get first dibs," "Shop the sale early—just for you," "30% off before everyone else." The body copy felt more conversational, almost chatty.
I sent Claude's batch to 60 people, Gemini's to the other 60.
Claude's results: 41% open rate, 14.3% click-through, $2,100 revenue.
Gemini's results: 38% open rate, 11.8% click-through, $1,780 revenue.
Claude won, but not by much. I think the slightly more formal tone reinforced the exclusivity—it felt like an invitation to something valuable, not just a friendly heads-up. Gemini's warmth worked better for the last-chance urgency segment, but VIP customers responded to being treated like VIPs.
The three-prompt framework I'm using for every campaign now
After testing 80 emails across three models, I built a prompt template I'll use for every campaign going forward. It has three parts, and if you skip any of them, the output gets worse.
Part one: Who is this person, and what do they already know?
Don't just say "customer." Say "someone who abandoned a $40 cart, has bought from us twice before, opens 60% of our emails, and cares about quality over price." The more context you give the model about intent and behavior, the better it writes.
Part two: What's the goal of this specific email in the sequence?
"Remind them the sale exists" is not a goal. "Make them feel like they're running out of time without lying about stock levels" is a goal. "Reward them for being loyal customers by giving them early access" is a goal. Be specific.
Part three: What should the email NOT do?
This is where you set boundaries. No emojis in subject lines. No fake scarcity. No body copy over 80 words. No countdown timers. No begging. The model will default to tricks that feel spammy unless you explicitly tell it not to.
I've been using this framework in Kryotta for every email campaign since December. I switch models depending on the segment—Claude for VIP and relationship emails, Gemini for urgency and last-chance pushes—and I haven't written a subject line manually in two months.
Questions people ask
Which model should I use for abandoned-cart emails?
Claude Sonnet 4.5, but only if you give it enough context about the customer's intent and behavior. If your prompt is vague, it'll write generic emails that don't convert. Gemini works if you want a warmer, more conversational tone, but Claude's better at balancing urgency with professionalism.
Can I use the same prompt for SMS and email?
No. SMS needs to be under 160 characters, and the tone has to be even more direct. I tested using the same prompts for both and the SMS versions felt too long. Write a separate prompt that specifies character limits and tells the model to cut all the preamble.
How do I know if the AI's subject line is better than mine?
A/B test it. Send the AI version to half your list and your manual version to the other half. I did this for six campaigns before I trusted the AI enough to skip the test. Claude's subject lines beat mine 4 out of 6 times, which was humbling but also saved me an hour per campaign.
Do I need to rewrite the emails before sending them?
Sometimes. I'd say 70% of what Claude and Gemini give me goes straight into Klaviyo. The other 30% needs a line or two adjusted—usually because the model used a phrase that sounds fine but doesn't match how I'd actually talk. Read every email out loud before you send it. If it sounds like a robot wrote it, your customers will notice too.
If you're setting up your own Cyber Monday sequences and want to test Claude, Gemini, or any of the other models I mentioned, you can try them all in one workspace at Kryotta. I've been using it since October because I got tired of switching between three different platforms just to compare outputs. It won't write your emails for you, but it'll let you test which model actually understands your customers—and that's the part that matters.
