Generated descriptions all sound alike for four reasons, and each is fixed separately — at the input, in the prompt, at the acceptance gate and in the cost report. A more expensive model does not help.
A description sounds generic when the model is handed someone else's marketing copy and a generic prompt. The fix is at the input and in the prompt, not in a more expensive model: feed raw specifications instead of a finished description, keep a separate prompt per category, put the result through human review, and track what one product costs.
Most stores do the same thing. They take the description from the supplier's price list, drop it into a prompt along the lines of "write a product description for this item", and publish the output. Three stores working with the same supplier get three near-identical texts — they started from the same text and passed it through the same model with the same prompt.
The result is recognisable: three paragraphs that sound confident and say nothing. "This product combines high performance with elegant design." The buyer learns nothing, and the search engine sees a duplicate.
The supplier's description was written to sell to you, not to your customer. It is already a retelling. Feed it to a model and you get a retelling of a retelling — which is exactly where the specifics disappear.
The right input is raw data: part number, brand, technical specifications, the supplier's category, available images. From those a model can write something true and specific. From marketing copy it can only write more marketing copy.
A motherboard and an office chair have nothing in common. Write them with one prompt and that prompt is necessarily generic — so the text comes out generic. A prompt has to know what matters for its category: for a processor that is socket, cores and TDP; for a chair, the mechanism, the upholstery and the rated load.
Automation with no acceptance gate produces volume, not quality. You need a queue where a person sees what was prepared and confirms it. Not to rewrite, but to catch the cases where the model assembled something plausible and wrong.
If you do not know what one description costs, you cannot decide which ones deserve a stronger model. Spend has to be visible per prompt, per supplier and per day — otherwise the choice of model is superstition.
Generation is not one prompt but six types, each with its own job: the product name, the description, extraction of filter values, category assignment, recognising the same product across different suppliers, and merging their data.
Prompts live in the database, are versioned, and are edited from the admin panel. Each can be universal or category-specific, has a "test with a real product" button before activation, and its own model choice — cheaper for mechanical work, stronger for writing. Everything is produced in Bulgarian and English at once.
The input is the raw product plus context from the store itself: the existing filters with their real values, the category tree, examples from the same category. That is why the output fits the catalogue instead of proposing categories that do not exist.
The finished product enters a review queue. The operator sees what the AI prepared, corrects what needs correcting and confirms. Only then does the product move to the store.
Every call is logged — which prompt, how many tokens, what it cost — with budgets and warnings. You know what you are spending before the bill arrives.
The descriptions will not become literature, and they should not. The goal is text that is true, specific to the product and different from the store next door. That comes from data at the input and a prompt that knows what it is writing about — not from a more expensive model.
Why do AI descriptions all sound the same? Because the input and the prompt are the same. Most stores feed the supplier's description into a generic prompt, and the supplier is the same one everybody uses. Different text starts from a different input — raw specifications instead of borrowed marketing copy.
Does a more expensive model fix it? No. A stronger model writes the same empty content more fluently. Specificity comes from the data at the input and from a prompt that knows what matters for the category.
How many prompts does one catalogue need? At least one per meaningful group of products, not one for everything. In Kronixon PI prompts can be universal or category-specific and are tested against a real product before activation.
Does a person have to review every product? Review is a required step, not a recommendation. Automation prepares, a person confirms. That is also where plausible mistakes get caught.
Can it write in two languages at once? Yes. The name, description and specifications come out in Bulgarian and English from a single call.
How is the cost tracked? Every call is recorded with its prompt, token count and price. Spend is visible per supplier, per prompt type and per day, with budgets and warnings.
Tell us what eats your team’s day. We will show you — on our own live system — how much of it can simply disappear.