The setup: Jumia, three models, fifty listings and one spreadsheet
I run a small electronics and home goods business in Nairobi. We sell through Jumia and a WhatsApp catalogue, mostly kitchen appliances, phone accessories and bedding. In November I had fifty new products to list—everything from Bluetooth speakers to duvet sets—and I was staring at fifty blank description boxes on the Jumia seller dashboard.
Writing them by hand would take days. I'd written maybe twelve descriptions before I gave up and opened Kryotta. I had three models sitting there: Claude Sonnet 4.5, Gemini Flash and DeepSeek V3. I decided to split the job: fifteen listings per model, test them live for a month, track which ones got clicks, which got sales and which got returns. Then write the last five with whichever model won.
I'm not a copywriter. I wanted descriptions that answered the questions Kenyan buyers actually ask—"Does this speaker work with my Samsung?", "Is the duvet warm enough for July?"—and that didn't sound like they were translated from Mandarin by a tired intern. I also wanted to spend less than KSh 500 total on the AI side, because margins are tight and I wasn't sure this would work.
Spoiler: it worked. One model got me a 22% conversion rate on electronics. Another wrote descriptions that sounded fine but got three returns in two weeks. The third was cheap, fast and completely missed what Nairobi buyers care about.
Prompt one: electronics (Bluetooth speakers, earbuds, chargers)
I started with the hardest category. Electronics buyers on Jumia read every word. They want to know if the charger fits their specific phone, if the speaker battery lasts through a weekend upcountry, if the earbuds will survive a boda ride in the rain.
I gave each model the same input: product name, specs sheet, price and this prompt:
"Write a Jumia product description for a Kenyan buyer. Product: [name]. Specs: [paste]. Price: KSh [amount]. Answer these questions in 120 words: What does it do? What phones/devices does it work with? How long does the battery last (if applicable)? Is it durable? Why is this price fair? Write like a Nairobi seller who knows their stock, not a marketing agency. No hype."
Claude Sonnet 4.5 wrote descriptions that sounded like a patient shopkeeper. For a KSh 1,200 Bluetooth speaker, it opened with "This speaker connects to any phone with Bluetooth—Samsung, Tecno, iPhone, doesn't matter." Then it listed battery life (six hours), mentioned the rubberised shell that survives a few knocks, and closed with "At KSh 1,200 you're paying for sound that's clear enough for a small room and a battery that won't die halfway through a playlist." Practical, specific, no fluff.
Gemini Flash wrote faster and cheaper, but the tone was off. Same speaker, same specs: "Experience premium sound quality with this portable Bluetooth speaker! Features 6-hour battery life and universal compatibility. Durable design perfect for on-the-go listening. Great value at KSh 1,200!" It wasn't wrong, but it read like every other listing. A buyer scrolling through fifty speakers wouldn't stop.
DeepSeek V3 gave me technically accurate descriptions that missed the point. It listed every spec—Bluetooth 5.0, 5W output, 1200mAh battery—but never answered "Will this work for my phone?" or "Is it loud enough?" One description spent thirty words explaining frequency response. Nobody shopping on Jumia at midnight cares about frequency response.
I posted five electronics listings per model. After two weeks, Claude's descriptions had a 22% click-to-sale conversion. Gemini's had 11%. DeepSeek's had 9%. The difference wasn't the specs—it was that Claude anticipated the buyer's next question and answered it in plain Swahili-English.
Prompt two: fashion (bedding, curtains, kitchen linens)
Fashion and home goods need a different approach. Buyers want to picture the item in their space. They want to know if the duvet is warm enough for Nairobi's cold season, if the curtains block out morning sun, if the colours match the photo.
New prompt:
"Write a Jumia listing for [product]. Material: [fabric]. Colours available: [list]. Price: KSh [amount]. In 100 words, describe what it looks like, how it feels, what room/season it suits, and what sizes/colours are in stock. Write for someone furnishing a Nairobi flat on a budget. No adjectives like 'luxurious' or 'elegant'—just tell me what I'm buying."
Claude wrote descriptions that felt like a friend showing you around their shop. For a KSh 2,800 duvet: "This duvet is thick enough for June and July when Nairobi gets properly cold. The outer cover is cotton, the filling is hollow fibre—warm but not heavy. It's a double size, fits a standard Kenyan bed. We have it in plain white and a grey-blue stripe. At KSh 2,800 it's cheaper than most bedding stores in town and it'll last a few years if you wash it gently."
Gemini gave me descriptions that were pleasant but vague. Same duvet: "Stay cozy with this comfortable double duvet. Soft cotton exterior, quality filling, available in white and blue. Perfect for Nairobi's cooler months. Affordable at KSh 2,800." It wasn't bad, but it didn't tell me if the duvet was thick or thin, heavy or light.
DeepSeek wrote descriptions that sounded translated. "High-quality duvet with cotton shell and fibre filling. Double size, suitable for cold weather. Colour options: white, grey-blue. Economical price KSh 2,800." Accurate, but stiff. Nobody talks like that.
I posted five bedding and linen items per model. Claude's descriptions got more favourites and fewer "Is this thick or thin?" messages in the Jumia chat. Gemini's got decent traffic but a lot of questions. DeepSeek's got the fewest clicks.
Prompt three: handling returns and complaints (the test I didn't plan)
Three weeks in, I had my first returns. Two duvets and one set of curtains. The curtains were a DeepSeek listing: the description said "blackout curtains" because that's what the supplier called them, but they weren't fully blackout—they blocked most light, not all of it. The buyer messaged: "These don't block the sun, I want a refund."
I went back to the listing. DeepSeek had copied the supplier's term without questioning it. Claude, when I tested the same product later, had written: "These curtains block most light—your room will be dim in the morning, not pitch black. If you need total darkness, look for our lined blackout range." That one extra sentence would've saved me a return and a refund argument.
The two duvet returns were both Gemini listings. The descriptions said "warm" and "comfortable" but didn't specify thickness. One buyer expected a heavy winter duvet, got a medium-weight one, sent it back. Again: not wrong, just not specific enough.
Claude's listings got zero returns in the first month. I don't think that's a coincidence.
The cost breakdown: KSh per listing and what you actually pay
I tracked the cost for every description. Kryotta shows token usage per prompt, and I converted that to shillings.
Claude Sonnet 4.5: roughly KSh 8–12 per listing, depending on length. For fifty listings, that's around KSh 500. The most expensive option, but the best conversion rate and zero returns meant I made that back in the first week.
Gemini Flash: KSh 2–4 per listing. Fast, cheap, decent for high-volume work. If I were listing two hundred items and didn't care about perfect descriptions, I'd use Flash. For fifty items where each sale matters, the savings weren't worth it.
DeepSeek V3: KSh 1–2 per listing. The cheapest by far, and the output was fine if you edited it. But I spent ten minutes per listing rewriting stiff phrases and adding context. That time cost more than the KSh I saved.
I ended up writing the last five listings with Claude. Total AI cost for fifty descriptions: around KSh 400. I've spent more on matatu fare in a week.
What I'd do differently: one model for drafts, another for final polish
If I were doing this again, I'd use Gemini Flash to generate a rough draft—fast, cheap, hits the basic structure—then feed that draft to Claude with a prompt like "Rewrite this to sound more like a Nairobi seller. Add specific details a buyer would ask about. Cut any hype."
That workflow would cost about KSh 5 per listing and give me 80% of Claude's quality without paying full price for every word. I tested it on three products after the experiment and it worked.
I wouldn't use DeepSeek for customer-facing copy. It's great for summarising supplier emails or pulling specs from a PDF, but it doesn't write for people—it writes for datasets.
The result: which model got sales and which got questions
After one month on Jumia:
Claude Sonnet 4.5: 22% conversion on electronics, 18% on home goods. Fifteen listings, eleven sales, zero returns. Buyers clicked through, read the description, bought without messaging me first. That's the dream.
Gemini Flash: 11% conversion on electronics, 14% on home goods. Decent traffic, but I got a lot of "Is this compatible with my phone?" and "How thick is this duvet?" messages. The descriptions weren't specific enough, so buyers hesitated.
DeepSeek V3: 9% conversion on electronics, 10% on home goods. The lowest traffic and the most returns. The descriptions were technically accurate but didn't answer the human questions.
I've since rewritten the Gemini and DeepSeek listings with Claude. Conversion has evened out across the store, and I'm spending less time answering the same five questions in the Jumia chat.
Questions people ask
Can I use these prompts for an Instagram shop or WhatsApp catalogue?
Yes. The structure works anywhere—just swap "Jumia buyer" for "Instagram buyer" or "WhatsApp customer." Instagram descriptions can be shorter (80 words instead of 120), and you'll want to add a line about payment (M-Pesa number, bank details) since Instagram doesn't have a checkout.
Do I need to edit the AI descriptions or can I copy-paste straight to Jumia?
I edited about 20% of Claude's descriptions—usually just tweaking a phrase or adding a spec the model missed. Gemini and DeepSeek needed heavier edits. If you're listing fifty products, expect to spend ten minutes per listing on edits, even with a good model.
Which model is best if I'm listing hundreds of products and need speed?
Gemini Flash. It's fast, cheap and good enough for volume work. You'll sacrifice some conversion, but if you're running a high-turnover store where you list new stock every week, speed matters more than perfect copy.
Does this work for services or only physical products?
I've only tested it on physical products, but the same logic applies. If you're selling a service—graphic design, M-Pesa agent training, boda hire—you still need to answer the buyer's questions (What do I get? How long does it take? What does it cost?) in plain language. Claude's better at that than the other two.
If you're sitting on a pile of unsold stock because writing descriptions feels like homework, try this. Pick three products, write three prompts, test all three models in Kryotta's Compare arena and see which one sounds like you. You'll know in ten minutes, and you'll have your first three listings done before lunch. Start your free trial here—no card, no M-Pesa upfront, just the models and a blank prompt box.
