I tested three models on fifty real menu-writing and order-reply tasks
I ran a food stall in Madina Market for two years before I started writing. My sister still runs it—grilled tilapia, jollof, kelewele, the usual. She takes most orders on WhatsApp now, posts the daily menu on Instagram stories, and spends an hour every evening typing replies to "Please I want to add one more waakye" and "Have you received my MoMo?"
Last month I offered to test whether AI could write those replies and save her some time. She was skeptical. "It will sound like a robot. People will know." Fair point. Ghanaian customers expect a certain warmth, a rhythm to the language. If your reply sounds like a corporate email, they'll go to the next seller.
I wrote fifty prompts across three tasks: writing a menu caption for Instagram, replying to order changes on WhatsApp, and confirming MoMo payments. I tested Claude Sonnet 4.5, Gemini Flash and Llama 3.3 70B on Kryotta. I wanted to see which model kept the tone natural, handled "Please" and "I want" constructions without sounding stiff, and didn't invent menu items or prices.
Claude won on naturalness. Gemini was faster but often too formal. Llama invented prices I never mentioned. Here's what each one did well, where they failed, and which tasks you should use each model for.
Claude kept the friendly "Please" rhythm without sounding like a call center
Ghanaian WhatsApp order messages have a particular politeness. "Please I want to add one more waakye." "Please can you remove the pepper?" "I beg, make the jollof plenty." If your AI reply comes back with "I have updated your order," it sounds wrong. People notice.
I gave each model this prompt: "A customer just sent: 'Please I want to add one more waakye with fish and egg, then remove the shito from the banku.' Reply warmly, confirm the change, ask if anything else." I ran ten variations—different dishes, different changes.
Claude Sonnet 4.5 replied: "No problem! I've added one more waakye with fish and egg, and I've removed the shito from your banku. Anything else you need?" Natural. It kept "No problem" instead of "" mirrored the customer's directness, and the question at the end felt human.
Gemini Flash wrote: "Okay, I've updated your order: one additional waakye with fish and egg, and the banku will come without shito. Is there anything else I can help you with today?" Polite, clear, but "I can help you with today" is call-center language. Nobody talks like that on WhatsApp in Accra.
Llama 3.3 70B gave me: "Noted! Adding one waakye with fish and egg, banku without shito. Let me know if you need changes." Short, fine, but "Noted!" felt abrupt for a food order. Customers want a bit more warmth, especially if they're spending GH₵35 on lunch.
I tested another scenario: "Customer says: 'I beg, can you make the jollof plenty? I'm very hungry.' Reply warmly and confirm." Claude wrote: "Haha, no worries! I'll make sure your jollof is plenty. You'll be satisfied, I promise." Gemini said: "Of course! I'll prepare a generous portion of jollof for you." Llama wrote: "Sure, I'll add extra jollof for you." Claude's "Haha" and "I promise" matched the customer's tone. Gemini's "generous portion" sounded like a hotel menu. Llama was fine but flat.
If you're writing WhatsApp order replies, use Claude. It's the only one that didn't make me cringe when I imagined my sister's customers reading it.
Gemini wrote the fastest menu captions, but you have to watch for invented dishes
I needed Instagram captions for daily menus. Short, appetizing, with prices in cedis. I gave each model this prompt: "Write an Instagram caption for today's menu: jollof rice with chicken GH₵25, waakye with fish and egg GH₵20, kelewele GH₵10, grilled tilapia GH₵40. Mention we deliver in Madina and Legon, MoMo payment accepted. Keep it under 60 words, friendly tone."
Gemini Flash came back in two seconds: "Today's menu 🔥 Jollof rice with chicken GH₵25, waakye with fish and egg GH₵20, kelewele GH₵10, grilled tilapia GH₵40. We deliver to Madina and Legon! Pay with MoMo. DM to order." Clean, quick, exactly what I asked for. I ran this test ten times with different menus and Gemini nailed the format every time. It's the fastest model for this task.
Claude Sonnet 4.5 wrote: "Today's menu is here! Jollof rice with chicken (GH₵25), waakye with fish and egg (GH₵20), kelewele (GH₵10), grilled tilapia (GH₵40). Delivering to Madina and Legon—MoMo accepted. Send a message to order!" Also good. Slightly warmer with "is here!" but three seconds slower than Gemini. For a task you're doing every day, that adds up.
Llama 3.3 70B wrote: "Jollof rice with chicken GH₵25, waakye with fish and egg GH₵20, fried plantain GH₵10, grilled tilapia GH₵40. Delivery available in Madina and Legon, pay via MoMo." Wait—I said kelewele, not fried plantain. I ran the test again. Llama changed "grilled tilapia" to "grilled fish" another time, and once it added "banku with okro stew GH₵15" that I never mentioned. It's not reliable when you need exact menu items and prices.
For daily menu captions, use Gemini. It's fast, accurate, and doesn't hallucinate dishes. If you want a slightly warmer tone and speed doesn't matter, Claude works too. Don't use Llama for anything involving prices or specific dishes.
Claude confirmed MoMo payments without sounding like a bank SMS
MoMo confirmations are tricky. You need to sound human, confirm the amount, reassure the customer their order is coming, but not be so casual that they worry you didn't actually check. I tested this prompt: "Customer sent a screenshot of MTN MoMo payment for GH₵45. Confirm you received it, thank them, say the food will be ready in 30 minutes and we'll send the rider. Friendly tone."
Claude Sonnet 4.5 replied: "Got it, thank you! I've seen the GH₵45 payment. Your food will be ready in 30 minutes and we'll send the rider right after. Appreciate your order!" Perfect. "Got it" is how people actually talk. "Appreciate your order" feels genuine without being stiff.
Gemini Flash wrote: "Thank you! I confirm receipt of your GH₵45 MoMo payment. Your order will be ready in 30 minutes, and our rider will deliver shortly." "I confirm receipt" is bank language. Nobody says that on WhatsApp.
Llama 3.3 70B gave me: "Payment received, thank you. Food ready in 30 mins, rider on the way soon." Fine, but "rider on the way soon" is vague. The rider isn't on the way yet—the food isn't even ready. Small detail, but customers notice when your message doesn't match reality.
I tested a variation where the customer paid GH₵50 for a GH₵45 order and asked for change. Claude: "Thank you! I've received the GH₵50. Your order is GH₵45, so the rider will bring GH₵5 change. Food will be ready in 30 minutes." Clear, confirms both amounts, no confusion. Gemini said: "Received your payment of GH₵50. The total is GH₵45, and you will receive GH₵5 in change upon delivery." Correct but formal. Llama wrote: "Got the GH₵50, will send GH₵5 back with rider. Food coming soon." "Send back" sounds odd—you're not sending money back, you're giving change. And "food coming soon" is too vague.
For MoMo confirmations, Claude is the safest choice. It balances clarity with warmth, and it doesn't slip into bank-SMS language.
Where each model failed in ways that would cost you customers
I ran ten tests where I deliberately gave incomplete information to see how each model handled ambiguity. Prompt: "Customer says: 'I want jollof.' Reply and ask what they want with it." This happens constantly—people assume you know they want chicken, or they forget to mention drinks.
Claude asked: "Sure! Do you want chicken, fish, or beef with the jollof? And anything to drink?" Natural follow-up, offers options, doesn't assume.
Gemini wrote: "Noted. What would you like to accompany your jollof rice?" "Accompany" is restaurant-menu English. On WhatsApp in Kumasi, you say "with" or "on the side."
Llama replied: "Jollof rice coming up! Anything else?" It didn't ask what protein. If you send that reply, the customer has to message again, you look disorganized, and you've added an extra back-and-forth that annoys everyone.
I also tested a scenario where the customer asked for something not on the menu. "Customer says: 'Do you have banku with groundnut soup?' You don't sell groundnut soup. Reply politely." Claude: "Ah, sorry, we don't have groundnut soup today. We have banku with okro stew or palm nut soup if you like?" Offers alternatives, stays warm. Gemini: "Unfortunately, groundnut soup is not available. We offer okro stew and palm nut soup with banku." "Unfortunately" and "is not available" sound like a hotel. Llama: "No groundnut soup, but we have okro stew." Blunt. Customers might not order at all if you sound dismissive.
The biggest failure across all three models: none of them handled pricing questions well when I didn't give exact numbers. Prompt: "Customer asks: 'How much is waakye with fish?' You charge GH₵18 for waakye with fish, GH₵20 if they add egg. Reply clearly." Claude said GH₵18, mentioned the egg option. Gemini said "GH₵18 for fish, GH₵20 with egg"—fine. Llama said "GH₵20"—wrong, it picked the higher price without explanation. When I tested this five more times with different dishes, Llama got the price wrong twice. That's a problem. Customers will call you out, or worse, they'll just stop ordering.
The model I'd actually use for each task
If I were still running the stall, I'd use Claude Sonnet 4.5 for WhatsApp replies and MoMo confirmations. It's the only one that sounds like a real person who sells food in Accra. The replies feel warm without trying too hard, and it doesn't slip into corporate language.
I'd use Gemini Flash for Instagram menu captions. It's faster, cheaper, and it doesn't invent dishes. You can write one prompt template, save it, and run it every morning in three seconds. That's worth the slightly less warm tone.
I wouldn't use Llama 3.3 70B for anything customer-facing until it stops hallucinating prices and menu items. It's fine for internal notes or brainstorming dish ideas, but the risk of sending a wrong price or confirming an item you don't sell is too high.
One more thing: I built all fifty prompts in Kryotta's compare mode, which let me run the same prompt across all three models at once and see the outputs side by side. Saved me hours. If you're testing models for your restaurant, that's the fastest way to figure out which one fits your voice.
Questions people ask
Can AI handle orders in Twi or Ga?
Not reliably yet. I tested a few prompts in Twi and the models either replied in English or mangled the grammar. Stick to English for now, or write your Twi templates yourself and use AI only for the English parts.
Will customers know it's AI?
If you use Claude and edit the output slightly—add your own emoji, change one word—no, they won't know. If you use Gemini or Llama without editing, yes, some will notice the stiff tone.
How much does this cost per month?
I spent about GH₵15 running fifty prompts across three models on Kryotta. If you're writing twenty replies a day, budget GH₵25–30 a month. Compare that to the hour you're spending typing the same answers every evening.
Does this work for sit-down restaurants or just delivery?
Same models, same tasks. If you're replying to reservation requests on WhatsApp or writing your weekly specials for Instagram, the workflow is identical. Claude still wins on tone, Gemini on speed.
If you sell food in Ghana and you're tired of typing the same replies every night, try Kryotta's compare mode and see which model sounds like you. You'll know in ten minutes whether it's worth using.
