.jpg)
The challenge
Vanden Borre stocks an enormous range of audio, TV, multimedia and telephony products, and wanted an appealing description plus pros and cons for every single one. Writing that manually was never realistic, but generic AI output wouldn't meet the bar for a retailer of this scale. The challenge was building prompts precise enough to generate consistent, on-brand copy across an entire product range, not just a one-off text.
If you want proof that consistency builds brand recognition, look no further. For years, hundreds of thousands of Belgians driving to work (or sitting in traffic) heard the same unmistakable line crackle through their car speakers: "U heeft goed gekozen/Vous avez bien choisi" (You've made a great choice). The tagline hasn't been in active use since around 2016, but it's stuck: still instantly recognisable to anyone who grew up with Belgian radio. Vanden Borre's current slogan, in place since 2022, is "Voor 't leven/C'est pour la vie" (For life), chosen to underline lasting relationships with customers and to highlight how durable their products are. This focus on sustainable innovation even earned the company third place at the 2023 Comeos Mercurius Awards, a prestigious innovation prize within Belgian retail.
One glance at their website tells you everything you need to know: they stock a lot of products. An enormous amount. And Vanden Borre likes to give every single one of them an appealing product description. Rather than sentence their marketing team to that Herculean task (thousands of texts, plus pros and cons for every product, with a collective burnout as the likely outcome), the company went looking for a solution.
Raccoons was able to provide exactly that.
Generating thousands of marketing texts and pros and cons for a huge range of audio, TV, multimedia and telephony products? Surely that's no big deal in 2026. Just type "write me a creative text for this range hood" into a chatbot, and ten seconds later you've got your copy.
But for a company like Vanden Borre, one of the most iconic names in Belgian retail, that simply won't cut it. These days, anyone with an internet connection can get AI to churn out a bit of text. What's no longer a secret, though, is that the quality of that output depends heavily on the quality of the prompt you feed the Large Language Model in the first place. And that's exactly where our expertise comes in: our contribution to this project was building high-quality "prompt templates" that allow the model to generate consistent, high-quality marketing texts and pros and cons, time after time. In both Dutch and French.
Thanks to those prompt templates, Vanden Borre could take the entire backend flow into their own hands: they develop and manage the product information, and ultimately the generated text itself.
If we wanted our prompt templates to be just as durable as the products the retailer sells, there were a few things we had to take into account:
To arrive at these prompt templates, we worked closely with Vanden Borre's marketing specialists. They provided a "golden dataset": a set of product information covering 24 products, from brand and commercial name to price and attributes. We used that dataset to test our prompts, refining them step by step based on their feedback.
Initially, that internal product data was the model's only source of information. It quickly turned out to be too little: the model didn't understand the product well enough, and had no idea how it compared to similar products on the market. In short, the "writing prompt", the prompt responsible for actually writing the text, needed a helping hand.
That helping hand came in the form of a "research prompt". Starting from the internal product data, this prompt heads out onto the internet to do its own digging. Whatever it turns up then serves as extra fuel for the writing prompt.
By this point the content was in good shape, though the odd small language or spelling slip still crept in here and there. Hence a third and final prompt: the "language correction prompt". It runs one last check and polishes away any remaining rough edges.
The result is a three-step flow: research, writing and language correction. Here's exactly what that flow looks like:
Incidentally, the marketing texts and the pros and cons were two entirely separate tracks. Same flow, but the prompts and their development ran completely independently of each other.
And for both, the same rule applied: the vast majority of the iterations happened on the Dutch version. Once that was sharpened to a fine point, the French counterpart followed close behind.
Because the research phase needed tool calling, specifically a web search tool, we configured a number of agents within Azure Foundry (Microsoft's central cloud platform for building, testing, managing and deploying AI agents and other AI applications). These agents let us connect the built-in Bing Search tool to an LLM, so the model could independently look up relevant information online during the research phase.
Azure AI Foundry also offers a wide selection of LLM models that can be deployed within the environment. In the early stages of the project, the entire flow still ran on GPT-4.1, but it quickly became clear that the sheer volume of instructions and guidelines in the prompts called for a stronger model. After some puzzling, we eventually landed on GPT-5.5, which handled the complex instructions far better and delivered the output we were after.
Good to know: a separate agent can be configured for each phase in the flow. That gives us the flexibility to deploy the most suitable model for each phase. The language correction phase, for instance, needs a lot less computing power than the research or writing phase, which means the available models are put to work far more efficiently.
In what follows we show an example of the flow applied to a dishwasher.
In the end, the project doesn't come down to a single prompt template, but to twelve: six in Dutch, six in French, built to work across Vanden Borre's entire product range. Every one of those templates was built with variables, so a single template can generate copy for any product in that range, no matter the category.
And the result? Vanden Borre can now give thousands of products copy that's accurate, persuasive and, above all, unmistakably on-brand, without needing an army of copywriters on the verge of a breakdown to make it happen. That leaves Vanden Borre's marketing team free to focus on checking, fine-tuning and putting the final polish on every text.
This was no simple "write me a creative text for this range hood". This project proves once again that writing a good prompt takes just as much craftsmanship as writing a good text. Interested in these prompt skills yourself? Feel free to get in touch.
Want to know more?
