A standalone composite, session-by-session walkthrough of the full agenda — Dubai edition — two friends, six weeks into their home-based kunafa business, solving a very small, very real problem: too few boxes some evenings, too many leftover others.
Maram & Aisha — Founders, Noor's Kunafa Corner (home-based bakery, Dubai — kunafa, luqaimat, karak; Dubai Economy & Tourism home-business licence, no physical shopfront, orders via Instagram and WhatsApp)
"Some Thursdays we sell out our kunafa boxes by six and people who messaged late get nothing. Other nights we've got four boxes left and Maram was up since Fajr for nothing. We don't want an app. We want to know how many boxes to make."
Composite persona built from realistic first-time Dubai home-business patterns, for illustration — not a real business.
How the tools are used
Everyone meets all four platforms, but builds with two. DeepSeek and GitHub Copilot are shown in short live demos. Manus turns the first prompt into a working draft and a project structure; Claude Code takes that project from GitHub and does everything after — the fixes, the shared data, the protection pass and the deploy.
"Aisha says it out loud once and they both nod — they've been having this conversation over WhatsApp for a fortnight."
Their brief:
A simple page where regulars can reserve today's kunafa box for pickup, so we know how many to bake and nobody messages us for nothing.
"Four short live demos, one question each time: is this the tool for a first product, built by two people who have never written code?"
| Tool | Role today | What they saw |
|---|---|---|
| DeepSeek | Demo only | Strong, low-cost writing and reasoning — shown side-by-side on an Instagram caption. Not fed any customer data until its data terms are checked against UAE PDPL. |
| GitHub Copilot | Demo only (free tier) | Suggests code line by line inside an editor. Most useful once you have a project and are writing code yourself; the full agent features need a paid plan. |
| Manus | Builds | One plain-English prompt to a working first draft — and, just as importantly, a project folder they can keep. |
| Claude Code | Builds | Works inside that project on GitHub: makes precise changes, fixes what's broken, and deploys it — all directed in plain English. |
PLAN: Meet all four, build with two — Manus for the first draft, Claude Code for everything after it.
"Three drafts before the Instagram bio sounds like them and not like every other home bakery on the timeline."
Exercise A — Instagram bio and caption, three passes (in Claude, with the first prompt run side-by-side in DeepSeek)
Write short, warm Instagram bio copy for a small Dubai home bakery. We make kunafa, luqaimat, and karak. Orders by DM, pickup only, evenings. Under 40 words, no clichés about 'artisan' or 'passion.'
Cut anything that sounds like every other home bakery on the grid. We're new, not precious about it.
Now write one caption line about the kunafa specifically, for someone who hasn't tried it before.
"For exercise B they looked at how other Dubai home bakeries handle a daily item that sells out, before assuming they needed a whole system for it."
Exercise B — deconstructing the incumbents, business-level
How do small Dubai home bakeries typically handle a daily limited-batch item selling out on Instagram? Break down the mechanics.
REALITY CHECK: Most of what came back was delivery-app integration, payment gateways, and multi-location inventory sync — none of which two people and a home kitchen need. A live count, a name, and a pickup slot was the entire job.
What they actually typed into Manus:
Build a page for a home bakery called Noor's Kunafa Corner. Show today's menu: kunafa, luqaimat, karak. Add a reservation form just for the kunafa: today we're baking 24 boxes. Let someone pick a pickup time (5:30, 5:45, 6:00, 6:15, 6:30, 6:45) and give their name. Show how many boxes are left after each reservation. Limit to 1 box per person.
Twenty minutes later: a working page, and the project folder behind it pushed to a new GitHub repository, ready for Claude Code.
REALITY CHECK: Two things looked right and weren't. The same name could reserve twice — a warning showed, but nothing stopped it. And the count only lived in the browser: Aisha's laptop said 22 left, Maram's phone still said 24. Both caught by testing on two devices, not by reading the description.
If Manus stalls
The free tier can run out or time out mid-build. Even then it usually leaves a project structure, which goes to GitHub and on to Claude Code as normal. If nothing usable comes back, the same brief goes straight to Claude Code to build the project from scratch — the rest of the day doesn't change.
"Aisha drove this one herself — Claude Code on the web, connected to their GitHub repo. No terminal, nothing installed."
What she asked for, in plain English:
The box count only updates on the phone that made the booking. Make it one shared count that every customer sees, stored online.
Fix the one-per-person limit — right now the same name can reserve more than once, it just shows a warning that doesn't stop it. It needs to actually block the second reservation.
Add a simple way for us to set tomorrow's batch number ourselves each evening, protected by a PIN so only we can change it.
Deploy it to Cloudflare Pages and give me the link.
Tested by the whole room:
A QR code on the screen, everyone booking a box from their own phone — Kunafa — today's batch: 9 of 24 boxes left, the same number on every phone at once.
REALITY CHECK: Not every catch is dramatic. These were just: does it actually do what the label says it does, on more than one phone? Most bugs are that, not a security hole — worth checking anyway.
"Half their regulars read Arabic first. The page had to work for them too."
Add an Arabic version of this page, right-to-left, with a language switch. Keep it warm and Gulf-appropriate, not formal.
REALITY CHECK: The first pass was grammatically fine and sounded like a government form. Maram rewrote the menu lines herself, and they tested the same copy in all four platforms before deciding whose Arabic to trust.
HOMEWORK: Send the link to three regulars tonight — not the public grid — and write down exactly what they do with it.
"Two of the three booked without asking a single question. The third sent a WhatsApp anyway — to ask whether she could pick up at 7. That's a real answer: the last slot was too early."
Back in Claude Code:
Add a 7:00 pickup slot. Then add a simple way for us to set tomorrow's batch number ourselves each evening, without asking anyone for help.
Clear customer names from the system every Sunday night.
REALITY CHECK: The batch setting worked first time — for anyone. There was no check on who was changing it, so any customer with the link could have set tomorrow's batch to zero. One more prompt put it behind a PIN only Maram and Aisha know.
Their five-question checklist, filled in:
| Market signal — is this a real complaint? | Yes — 9 people have DM'd "do you take orders for pickup?" in the past 10 days, unprompted. |
| Feasibility — can they run this themselves? | Yes — Aisha can check her phone between her other job and the school run. |
| Cost signal — what does this replace? | A couple of days of their time, against a delivery-app commission (25–30%) or a proper booking/EPOS platform they don't need for a three-item menu. |
| User reaction — how will they actually test it? | Soft-launch to regulars via Instagram Stories and a WhatsApp broadcast list — not posted to the public grid yet. |
| Technical risk — what could go wrong? | Low — no online payment (cash or transfer on pickup); no personal data beyond a first name and a time slot, cleared every week. |
VERDICT: Turn it on for regulars this week via WhatsApp. Only post it to the public Instagram grid after a week of it working without a hitch.
"Before: right-click, View Source, and the whole page is there to copy. After: nothing a competitor could lift in an afternoon."
Before each deploy, obfuscate the page's code and add a hidden watermark unique to this build, and keep a log of which build went live when.
REALITY CHECK: This doesn't make copying impossible — it makes it slow, and it makes a copy provably theirs. For a three-item bakery page that's the right amount of protection, and it's the same pass the programme's founder ran on their own products before launch.
Their self-check:
| What they checked | What they found / did |
|---|---|
| The AI-drafted kunafa description | Said "contains gluten and dairy" — true, but never mentioned the crushed pistachio topping, because nobody had told the AI that detail. Aisha corrected it by hand against the real recipe and added a clear nut warning. |
| The luqaimat and karak descriptions | Matched the real recipe and ingredients — no issue, kept as-is. |
| The business name, "Noor's Kunafa Corner" | A quick search confirmed no existing UAE trademark conflict. Flagged as a reminder: register the name with the Ministry of Economy if the business grows beyond the home-business licence. |
| Photos used on the page | A kunafa photo pulled from a search that looked close enough. Swapped for their own phone photo, taken that evening. |
| Customer data | First names and pickup slots only, cleared weekly. None of it pasted into any AI tool whose data terms they hadn't checked against UAE PDPL. |
REALITY CHECK: AI can only work with what it's told, not what's actually in the kitchen. It doesn't know the pistachio topping is there unless someone says so — checking the copy against the real ingredients isn't optional here. It's an allergy risk, not a style preference. And a business name worth building on is worth a five-minute trademark check before it's on every box.
DECISION: Stay on the home-business licence for now. Revisit a free zone or mainland licence only if they start selling through retailers or hiring — and know which door (Dubai SME, in5) to knock on when they do.
"The link goes in the Instagram bio and the WhatsApp broadcast, not on a flyer in the building lobby — same production-versus-test lesson as everyone else this cycle, just with rosewater on it."
The facilitator then ran the obfuscation-and-watermark pass on their build as a live demo: View Source before, View Source after.
ACTION: Live on Cloudflare Pages, linked from Instagram and WhatsApp only. Lobby flyer held back a week, until they trust it under real Thursday-evening pressure.
Their AI build log:
| When / Tool | What happened |
|---|---|
| Day 1, 11:40 — Manus | Built the menu page, reservation form and count from one prompt; project pushed to GitHub. Count lived in one browser only; the per-person limit didn't actually work. |
| Day 1, 13:40 — Claude Code (Aisha driving) | Made the count shared and online, made the one-per-person limit real, deployed to Cloudflare Pages. |
| Day 1, 15:20 — Claude + Maram | Arabic version added; menu lines rewritten by hand. |
| Day 2, 10:05 — Claude Code | Added a 7:00 slot, a PIN-protected batch setting and a weekly clear of customer names. |
| Day 2, 12:10 — Claude Code | Obfuscation and per-build watermark added to every deploy. |
| Day 2, 13:40 — Manual | Corrected the allergen description against the real recipe; checked the business name; swapped in their own photo. |
| Day 2, 15:20 — Cloudflare Pages | Linked from Instagram and WhatsApp only, lobby flyer held back a week. |
| Time / Tool | What happened |
|---|---|
| 11:10 — Manus | Built the menu page, reservation form and count from one prompt; project pushed to GitHub. Count lived in one browser only; the per-person limit didn't actually work. |
| 13:05 — Claude Code (Aisha driving) | Made the count shared and online, made the one-per-person limit real, added a PIN-protected batch setting, deployed to Cloudflare Pages. |
| 14:50 — Manual | Corrected the allergen description against the real recipe; checked the business name; swapped in their own photo. |
| 15:40 — Facilitator demo | Obfuscation and watermark pass run on their build; linked from Instagram and WhatsApp only. |
"Manus got them from a sentence to a page in twenty minutes. Claude Code turned it into something they could trust on a Thursday evening."
"They didn't build a booking platform. They built a way for the people who specifically message for gran's kunafa to actually get a box. Twenty-four boxes, real names, and nobody left waiting on a reply at six on a Thursday."
Composite illustrative example for the Where to Start and Grow programme. Noor's Kunafa Corner, Maram, Aisha, and the numbers above are constructed for teaching purposes — not a real business.