{
  "source_agent_id": "wsa_ba4e006c41b5468ca49c2efdae5ffa71",
  "exported_at": "2026-10-15",
  "agent_name": "retail-review-history",
  "display_name": "Retail review history",
  "description": "For a retailer, product and date window, returns every customer review published in the window on that retailer's product page, one row per review, with date, star rating, title, text, verified flag and source page, plus liked and criticised features tagged from a fixed, versioned skincare taxonomy (v1). Stars, dates and text are copied exactly; tags are in separate fields.",
  "use_case": "dataset_building",
  "effort": "x-high",
  "sources": {"allow": [], "block": [], "prioritize": "The named retailer's own product page and its review pages", "avoid": "Brand websites, review aggregators, price comparison sites and any retailer other than the one named in the input"},
  "suggested_questions": ["Retailer: cultbeauty.com. Product: Medik8 Crystal Retinal 6 (ID 11899386), https://www.cultbeauty.com/p/medik8-crystal-retinal-6/11899386/. Window: 2025-10-01 to 2026-09-30. Return every review in the window."],
  "output_schema": {"type": "array", "items": {"type": "object", "required": ["retailer", "product_id", "product_name", "review_date", "stars", "source_url"], "properties": {
    "stars": {"type": "integer", "description": "Star rating 1-5 exactly as shown"},
    "author": {"type": ["string", "null"]},
    "retailer": {"type": "string", "description": "Retailer domain"},
    "product_id": {"type": "string"},
    "source_url": {"type": "string", "description": "URL of the page the review was read from"},
    "page_number": {"type": ["integer", "null"], "description": "Review page number the review was read from, if paginated"},
    "review_date": {"type": "string", "description": "YYYY-MM-DD exactly as shown on the review"},
    "review_text": {"type": ["string", "null"], "description": "Review body, first 400 characters"},
    "product_name": {"type": "string"},
    "review_title": {"type": ["string", "null"]},
    "liked_features": {"type": "string", "description": "Semicolon-separated LIKED labels from the fixed taxonomy (4-5 star reviews only), or empty"},
    "pdp_rating_avg": {"type": ["number", "null"], "description": "Overall average rating shown on the PDP at read time"},
    "pdp_review_count": {"type": ["integer", "null"], "description": "Total review count shown on the PDP at read time"},
    "taxonomy_version": {"type": "string", "description": "Always the string v1"},
    "variant_mentioned": {"type": ["string", "null"], "description": "Product strength or size the reviewer says they used, if stated (e.g. 'Retinal 3')"},
    "verified_purchase": {"type": ["boolean", "null"]},
    "criticised_features": {"type": "string", "description": "Semicolon-separated CRITICISED labels from the fixed taxonomy, or empty"}
  }}},
  "skill": "Purpose: collect every customer review published within the date window on one retailer product page, exactly as displayed, one row per review, and tag each review with the product features it praises or criticises from the fixed taxonomy below.\n\nORDER OF WORK (mandatory): finish PART 1 for the whole window first, run the continuity check, and only then tag the collected rows in PART 2. Never skip review pages to save time; tagging must not reduce collection.\n\nPART 1: COLLECTION (copy exactly, no judgement)\n- Do NOT sample. Return every review whose date falls inside the window.\n- Copy stars as the integer shown and the date as shown (YYYY-MM-DD). review_text: the first 400 characters of the review body, verbatim (or the shorter limit given in the input).\n- Start at the product page URL given in the input. Reviews are listed newest first. Page through ALL review pages in order, one page at a time, until the oldest review on a page is older than the window start, then stop.\n  - cultbeauty.com: review pages are server-side: append ?pageNumber=N to the product URL (N = 1, 2, 3 ...). Each page shows 10 reviews with date, stars, title, text and \"Verified Purchase\". The page states \"Page X of Y\".\n  - lookfantastic.com: ?pageNumber is ignored. The page HTML embeds only the latest ~10 reviews as JSON-LD; older reviews load in the rendered review section via \"load more\" / next-page controls or the review widget's own pages. Use the rendered page and its review controls to reach older reviews, advancing one page at a time; do not jump ahead.\n  - Other retailers: use the site's own review pagination or \"load more\" control, sorted by most recent.\n- On every row record pdp_review_count and pdp_rating_avg as shown on the product page, page_number (the actual review page the row came from, counted from 1 at the newest reviews) and source_url of the page the review was read from.\n- If a review mentions which strength or size the reviewer used (e.g. \"Retinal 3\", \"number 6\"), put it in variant_mentioned.\n- Continuity check before tagging: page_number values must run without gaps from the first page holding in-window reviews to the last; review dates must not jump forward between consecutive pages; there must be no calendar gap of more than 3 days with zero reviews in a window where neighbouring days average more than 2 reviews a day. If any check fails, go back and read the missing pages before continuing.\n- Completeness: the oldest review returned must be older than or equal to the window start. If you could not page back that far, still return every review you reached; never invent or estimate reviews.\n- No duplicates: the same author, date and text is one review. Never take reviews from another retailer, the brand's own site or a review aggregator.\n\nPART 2: FEATURE TAGGING (taxonomy v1; set taxonomy_version to \"v1\" on every row)\nTag from the review title and text only. Use ONLY these labels, spelled exactly; separate multiple labels with \"; \"; leave the field empty if none apply.\n\nliked_features (only for 4 or 5 star reviews; empty for 1-3 stars):\n- Visible results: skin looks or got better: glow, radiance, brighter, smoother, fewer or softer fine lines/wrinkles, clearer skin, more even tone, faded pigmentation, smaller pores, improved skin texture, compliments on skin, \"amazing results\", \"skin has never looked better\".\n- Texture / absorption: praises how the product itself feels: lightweight, absorbs or sinks in quickly, not greasy or sticky, silky, creamy, nice consistency. (Improved SKIN texture is Visible results.)\n- Hydration / softness: skin feels hydrated, moisturised, plump, nourished, dewy, soft.\n- Gentle / no irritation: gentle, mild, soothing, calming, no irritation/redness/peeling/purging/breakouts, suits sensitive skin, well tolerated.\n- Strength ladder: praises the range of strengths or progressing through them (started on 3, moved up to 6, worked my way up, good for beginners).\n- Value / lasts: worth the price, good value, lasts a long time, a little goes a long way.\n- Packaging / pump: praises the pump, packaging or dispenser.\n- Scent: praises the smell, or that it is fragrance-free.\n- Under make-up / SPF combo: works well under make-up, or values the built-in SPF.\n\ncriticised_features (any star rating, only when the reviewer reports a problem they experienced):\n- No results: no difference, no visible effect, did nothing, skin got worse.\n- Irritation / dryness: irritation, redness, peeling, purging, breakouts, burning, stinging, dryness or flaking caused by the product, allergic reaction or rash.\n- Price: too expensive, overpriced.\n- Packaging / pump: pump, lid or packaging broke, leaked, stopped working, got cheaper in quality, or tube arrived/looked half empty.\n- Staining / pilling: stains pillowcases or skin, yellow tint, pills under other products.\n- Scent: dislikes the smell.\n\nTagging rules:\n- Tag only what the reviewer states. Do not infer from the product name or the star rating.\n- Negations matter: \"no irritation\" is Gentle / no irritation, not Irritation / dryness. \"Not sure of results yet\" or \"early days\" is neither.\n- Describing one's skin type (\"I have dry skin\") is not a criticism. Saying a problem went away (\"flakiness is gone\", \"doesn't stain\") is not a criticism.\n- Generic praise with no feature (\"love it\", \"my favourite\", \"highly recommend\") gets empty liked_features.\n- Reviews in other languages: tag the same way from their meaning.\n- If the star rating and text clearly contradict (e.g. 1 star with entirely positive text), still tag criticised_features from the text only and leave liked_features empty for 1-3 stars."
}