1. Overview

This agent reviews product listings submitted to an online marketplace and returns an instant moderation decision: approve, reject, or escalate for manual review. Each decision includes the specific policy violations found and a seller-facing explanation written in clear, actionable language.

2. Business value

• Speed: moderates listings in seconds instead of the minutes-to-hours a manual review queue requires, so sellers go live faster.

• Consistency: applies the same policy rules to every listing, which eliminates reviewer-to-reviewer variance and reduces appeals.

• Scalability: handles traffic spikes (holiday seasons, flash sale onboarding) without adding headcount.

• Seller trust: provides specific, cited reasons for any rejection so sellers can fix and resubmit rather than guess what went wrong.

3. Operational context

• When it runs: triggered every time a seller submits a new listing or edits an existing one.

• Who uses it: marketplace operations teams, trust & safety teams, seller support.

• How often: continuous, high-volume. Built for hundreds to thousands of listings per day.

4. Inputs

5. Outputs

6. Execution steps

1. Validate inputs. Confirm Title and Description are non-empty. If either is blank, return a rejection with violation type missing_required_info.

2. Check for prohibited items. Scan title and description for references to weapons, drugs, counterfeit goods, recalled products, or hazardous materials using the prohibited terms list in Appendix A.

3. Check for offensive language. Flag profanity, hate speech, sexually explicit terms, or discriminatory language.

4. Check for misleading claims. Identify unsubstantiated phrases like "100% genuine," "official," "certified," "FDA approved," or "doctor recommended" that appear without verifiable context.

5. Check formatting compliance. Flag ALL CAPS titles, excessive punctuation (3+ consecutive ! or ?), emoji-only descriptions, or descriptions under 20 characters.

6. Check category-specific rules. Apply category-specific policies (e.g., Food & Beverage requires allergen disclosure; Health & Beauty cannot claim to cure diseases).

7. Aggregate violations. Compile everything found across steps 2-6.

8. Determine status:

   • Zero violations and high confidence → approved

   • One or more clear violations → rejected

   • Borderline or ambiguous cases (e.g., brand name that doubles as a common word, vintage items that could be counterfeit) → escalated

9. Generate seller explanation. Write a clear, specific explanation for each violation. Include what rule was broken and what the seller should change.

10. Set recommended action. Map status to action: approved = publish, rejected = remove, escalated = manual_review.

7. Validation checks

• Every violation found must appear in both the violations array and the seller explanation. No silent flags.

• Seller explanations must cite the exact text or phrase that triggered the violation, not generic messages.

• Common words that happen to be brand names (e.g., "Apple" as a color, "Coach" as a verb) should be evaluated in context before flagging.

• If a listing has multiple violations, document all of them. Don't stop after the first one.

8. Edge cases

• Multilingual listings: if the listing contains non-English text, escalate for manual review. Don't auto-reject.

• Empty description with images only: reject with insufficient_description. Listings must have text descriptions.

• Vintage/secondhand claims: "Vintage" or "pre-owned" items making authenticity claims about luxury brands should be escalated, not auto-rejected.

• Seller-provided certifications: if a seller mentions a certification (e.g., "USDA Organic"), escalate for verification rather than rejecting.

• Duplicate listings: out of scope. Duplicate detection is handled separately.

9. Example

Input

• Product Title: "GENUINE ROLEX WATCH - BEST PRICE!!!"

• Product Description: "100% authentic Rolex Submariner. This is the real deal. Ships from warehouse. Certificate of authenticity available on request."

• Product Category: Other

Execution

• Step 5: title is ALL CAPS with excessive punctuation (3x !) → violation: spam_formatting

• Step 3: no offensive language found

• Step 2: no prohibited items found

• Step 4: "100% authentic," "GENUINE," and "the real deal" are unverifiable authenticity claims for a luxury brand → violation: misleading_claim

• Step 8: multiple clear violations → status: rejected

Output

• Status: rejected

• Violations: [spam_formatting, misleading_claim]

• Seller Explanation: "Your listing was not approved for two reasons: (1) The title uses ALL CAPS and excessive punctuation, which violates our formatting guidelines. Please rewrite the title in standard capitalization. (2) The phrases '100% authentic,' 'GENUINE,' and 'the real deal' are unverifiable authenticity claims. If you can provide proof of authenticity (authorized dealer documentation or verifiable certificate), please resubmit with that documentation attached."

• Recommended Action: remove

• Confidence: high

---

Appendix A: Violation types

Appendix B: Decision matrix