AI Product Descriptions for Uzum and Other Marketplaces

The short answer

AI writes good product descriptions when you supply the facts and it only does the wording. A working pipeline looks like this:

1. A spreadsheet with specs and photos for each product.
2. A per-category prompt template that forbids adding anything not in the data.
3. Generation in Russian and Uzbek in one request or two passes.
4. Automated checks: length, banned words, numbers matching the spreadsheet.
5. Human review of a sample and of flagged cards.
6. Export in the marketplace format — import template or API.

Preparing the data

Description quality is set by spreadsheet quality. Minimum fields:

If specs are scattered across supplier PDFs, extract them into a table first (an LLM can help here too), check them manually, and only then generate copy.

Photos help describe appearance: color, shape, texture. Dimensions, materials and compatibility cannot be determined from a photo — forbid the model from drawing such conclusions.

Prompt template

Make a separate template for each major category: clothing, electronics and cosmetics need different emphasis.

You write a marketplace product description.
Use ONLY facts from the DATA block. If a fact is missing, do not mention it.
Do not use: "best", "No. 1", "guaranteed", medical promises.
Structure:
1. Title (up to N characters): product type + brand + model + key feature.
2. Description: 3-5 short paragraphs on buyer benefits.
3. Spec list as "parameter: value".
Languages: Russian and Uzbek (Latin script). Write Uzbek naturally, not as a literal translation.
Return JSON: {"ru": {...}, "uz": {...}}.

DATA:
{characteristics}

Tips:

• Structured output (JSON) makes checking and uploading easier.
• Give 1–2 examples of good cards from your category — they set the style better than long instructions.
• Check Uzbek separately: models may mix Latin and Cyrillic or carry over Russian phrasing. State the script explicitly.

Keeping facts accurate

Hallucinations in a card mean returns, complaints and a risk of the listing being blocked. Automated checks worth writing in code:

• every number in the text must appear in the source spreadsheet;
• brand and model match the data;
• no marketplace stop words or prohibited claims;
• title and description length within limits;
• both languages contain the same set of specs.

Do not auto-fix cards that fail the checks — send them for manual editing.

Marketplace rules

Each platform has its own requirements for titles, descriptions, photos and prohibited wording, and they change. Before launch:

• read the current content requirements in the Uzum seller account and on other marketplaces you sell on;
• turn them into a checklist and build it into both the prompt and the automated checks;
• remember that the same product often needs different titles and text lengths on different platforms;
• never copy competitor or manufacturer descriptions word for word.

Review before upload

Even with good checks you need a human. A sensible order:

1. Pilot: 20–30 cards from one category, full review, prompt tuning.
2. Scaling: spot checks from every batch plus all cards flagged by automation.
3. Feedback: buyer questions and "not as described" returns are signals to update the template.

Store the prompt version and source data for every card so you can trace an error quickly.

Common mistakes

• Generating "from the product name" without a spec sheet.
• One prompt for all categories.
• Uzbek copy as a literal translation of Russian.
• Uploading hundreds of cards without a pilot.
• Manual edits in the seller account that get overwritten by the next generation run.