A retailer we work with had a real problem, and it is a common one. Nearly two hundred products, and for most of them the "description" was a bare model number. Anyone selling online, sending a catalogue on WhatsApp, or listing on a marketplace hits this wall: writing two hundred tidy descriptions by hand is a week of dull work nobody wants to do.
We did it in an afternoon with a chat assistant. No integration, no developer, no code - this is stage 1 of the ladder: AI as a fast assistant you brief and check. Here is exactly how, including the parts that went wrong.
The messy starting point#
The real input was not clean. It was a spreadsheet export where each row had a model number, a rough category, sometimes a price, and often a typo. A few rows had no category at all. This is what real catalogues look like, and it is fine - you do not need to tidy it first.
The only preparation that mattered was deciding what a good description should contain. For this shop we settled on: one plain sentence of what it is, three short benefit-led lines, and a consistent friendly tone. Nothing fancy.
Write the first five by hand#
Before touching the AI, we wrote five descriptions ourselves - properly, the way we wanted all two hundred to read. This is the most important step and the one people skip.
Those five become the example. An AI assistant is superb at copying a pattern you show it and mediocre at guessing a pattern you only describe. Show, do not just tell.
The prompt that worked#
With the five examples in hand, the instruction was simple and specific:
You are helping write product descriptions for a Kenyan electronics shop.
Here are five examples written in the style and length I want:
[paste your five hand-written descriptions]
Now write descriptions in exactly that style for the products below.
For each one give: one sentence of what it is, then three short benefit
lines. Keep the friendly tone. If a detail is missing, do NOT invent it -
leave a [CHECK] tag so I can fill it in.
Products:
[paste the next 20 rows]Two details in that prompt did the heavy lifting. The five examples set the style precisely. And the instruction "do not invent - leave a [CHECK] tag" is what kept it honest, which brings us to the part everyone glosses over.
The ones we had to fix#
It was not perfect, and pretending otherwise would be dishonest. Out of the two hundred, roughly one in six needed a human touch:
- Invented specs. A few times it confidently stated a screen size or battery
capacity that was not in the input. The
[CHECK]instruction caught most of these - but not all, which is why you still read every line. - Wrong category guesses. For the rows with no category, it occasionally filed a product in the wrong place. Quick to fix once spotted.
- Repetition. Feed it twenty similar items and the benefit lines start to sound the same. We nudged it: "vary the wording so they do not all read alike."
Why batches of twenty, not two hundred#
We fed it twenty rows at a time, not the whole list. Two reasons. Quality holds up better on a smaller batch - accuracy drifts on very long lists. And it is far easier to check twenty fresh descriptions than to face a wall of two hundred at the end. Do a batch, check it, paste the next.
What this is, and what it is not#
This is stage 1: a capable assistant doing a language-shaped, low-risk, repetitive job while you supervise. Notice what it is not. Nothing is connected to the shop's systems. The AI did not update the catalogue itself - a person pasted the finished descriptions back in. That connected, hands-off version is stage 3, further up the ladder, and it is a different kind of work.
For now, the win is simple and real: a week of dull writing became an afternoon, and the shop owner spent that afternoon checking and improving instead of typing from scratch.
Try it this week#
Pick twenty products, write five descriptions by hand, and run the prompt above. If it saves you an afternoon, you have your answer - and you have taken your first real step up the ladder. For where this sits in the bigger picture, see the five stages of AI adoption.
