AI-Powered Ad Description Generation for Classifieds Platform

leboncoin · E-commerce · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Classifieds ad creation

Leboncoin wanted to help sellers overcome blank-page syndrome and write better ad descriptions without losing the human tone of a peer-to-peer marketplace.

Constraints:

Architecture

Leboncoin built ad-description assistance directly into listing creation: seller inputs feed Claude Haiku through Bedrock, prompts are tuned for marketplace voice, legal review shapes disclosure, and sellers keep final editing control.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  seller-listing-flow["Seller listing flow"]
  claude-haiku-generator["Claude Haiku generator"]
  prompt-iteration-loop["Prompt iteration loop"]
  legal-and-data-use-review["Legal and data-use review"]
  generated-description["Generated description"]
  seller-edit-flow["Seller edit flow"]
  published-ad["Published ad"]
  seller-listing-flow --> claude-haiku-generator
  prompt-iteration-loop --> claude-haiku-generator
  prompt-iteration-loop --> legal-and-data-use-review
  legal-and-data-use-review --> seller-listing-flow
  claude-haiku-generator --> generated-description
  generated-description --> seller-edit-flow
  seller-edit-flow --> published-ad
  published-ad --> seller-listing-flow
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
Seller listing flow Collects photos, title, and item details from the seller. Claude Haiku generator
Claude Haiku generator Uses Bedrock-hosted Claude Haiku to draft marketplace-appropriate ad descriptions. Generated description
Prompt iteration loop Refines tone and wording through interviews and user tests. Claude Haiku generator, Legal and data-use review
Legal and data-use review Handles disclosure, user data, and AI partner questions from the start of design. Seller listing flow
Generated description Provides a structured natural-language starting point for the seller. Seller edit flow
Seller edit flow Lets sellers tweak, add personal detail, or restart. Published ad
Published ad Carries the human-approved description into the marketplace. Seller listing flow

Data flow

Photos, title, and item details are passed to an AI generator that creates a structured description, then the seller edits or accepts it before publishing.

“The algorithm considers your photos, title, and item details to generate relevant, structured descriptions that feel natural and engaging.”article, stated
“The AI-generated description is just a starting point. Sellers maintain full control: they can tweak the text, add personal details, or start fresh if they prefer.”article, stated

Why this architecture

Prompt and legal loops were needed because early outputs sounded like product catalogs and AI/data disclosure had to be designed up front.

“During user testing, our first attempts at AI-generated descriptions fell flat — they were too verbose, too formal, reading more like product catalogs than leboncoin’s ads.”article, stated
“Instead of the traditional “build first, legal later” approach, we invited our legal team to join the design process from day one.”article, stated

Models used: Claude Haiku

Scale

Alternatives considered (roads not taken)

Writing tips or template descriptions only

Rejected because: Leboncoin chose generative descriptions to move beyond basic templates while keeping the human touch.

“We could have stuck to the basics — offering writing tips or template descriptions. But in a world where AI is transforming how we work, why not dream bigger?”article, stated
“The AI-generated description is just a starting point. Sellers maintain full control: they can tweak the text, add personal details, or start fresh if they prefer.”article, stated

Late legal review after implementation

Rejected because: Leboncoin included legal in design from day one to resolve AI partner data and disclosure questions.

“Instead of the traditional “build first, legal later” approach, we invited our legal team to join the design process from day one.”article, stated

Formal catalog-style AI copy

Rejected because: User testing rejected verbose formal outputs that did not match the marketplace voice.

“During user testing, our first attempts at AI-generated descriptions fell flat — they were too verbose, too formal, reading more like product catalogs than leboncoin’s ads.”article, stated

Patterns

human-in-the-loop-gate (workflow)

The generated description is explicitly a starting point that sellers can edit, personalize, or discard.

“The AI-generated description is just a starting point. Sellers maintain full control: they can tweak the text, add personal details, or start fresh if they prefer.”article, stated

prompt-versioning (prompting)

The team iterated AI outputs through interviews and user testing to move from verbose product-catalog copy toward Leboncoin's concise marketplace tone.

“Working with Nafi, we conducted multiple rounds of interviews, testing different versions of our AI outputs.”article, stated

privacy-preserving-processing (safety)

Legal joined from the start to shape user-data handling, disclosure that AI writes descriptions, and compliance requirements.

“How do we handle user data with our AI partner? Where and how do we tell users that AI is helping write their descriptions?”article, stated

Tradeoffs

How much authorship to give the AI.

Chose AI co-pilot with seller editing control over fully automatic ad publishing.

Rationale: Seller control preserves personal tone and marketplace authenticity while still reducing blank-page friction.

“But here’s the key: we’re not replacing the human touch.”article, stated

Whether to use simple templates or generative descriptions.

Chose generative AI descriptions over writing tips or template descriptions.

Rationale: The team wanted to rethink ad creation rather than only patch the blank-page problem.

“We could have stuck to the basics — offering writing tips or template descriptions. But in a world where AI is transforming how we work, why not dream bigger?”article, stated

Gotchas (what bit them)

Generic AI copy did not match Leboncoin's marketplace voice.

Root cause: First attempts were too verbose and formal, sounding like product catalogs rather than concise peer-to-peer ads.

Resolution: Run multiple rounds of interviews and output testing to tune toward concise, complete, conversational copy.

“During user testing, our first attempts at AI-generated descriptions fell flat — they were too verbose, too formal, reading more like product catalogs than leboncoin’s ads.”article, stated

Late legal review would have forced redesign.

Root cause: Consumer AI features raise data-handling and disclosure questions that affect product design.

Resolution: Bring legal into the design process from day one as an active collaborator.

“Instead of the traditional “build first, legal later” approach, we invited our legal team to join the design process from day one.”article, stated

Evaluation

Leboncoin reports user testing feedback and production lift for ads using generated descriptions.

Metrics: 20% increase in inquiries · 20% increase in completed transactions

“ads using this feature are seeing a 20% increase in both inquiries and completed transactions.”article, stated

Outcomes

Company tech-blog post; impact claims are self-reported.

“The algorithm considers your photos, title, and item details to generate relevant, structured descriptions that feel natural and engaging.”article, stated