product-researcher - SKILL.md Agent Skill

name: product-researcher description: > Research products and display results on a live canvas with product images, prices, star ratings from multiple sources (Amazon, Google, Reddit, review sites like Wirecutter/RTINGS/PCMag), pros/cons, and working buy links. Use when the user wants to compare products, research before buying, find the best options in a category, look up reviews, or says things like "research X", "compare these products", "find the best Y", "show me reviews for", "help me pick a backpack/laptop/camera/headphones/etc", or attaches a screenshot of products from a shopping site. Renders everything on the live canvas (requires live-canvas skill).

Product Researcher

Researches products and renders a rich comparison canvas with images, multi-source ratings, pros/cons, and buy links.

Workflow

1. Identify Products to Research

From user input, extract:

Products explicitly named or shown (e.g., from screenshots)
Product category (e.g., "travel backpack", "noise-cancelling headphones")
Any specific models already in the user's cart/wishlist

Also add 2–3 top-rated competitors from your research to give the user comparison context.

2. Research Each Product (use WebSearch)

For each product, run these searches and extract the data:

"{product name} review pros cons 2025"
"{product name} amazon rating site:amazon.com"
"{product name} reddit review recommended"
"{product name} buy price site:amazon.com OR site:{brand}.com"
"{product name} site:rtings.com OR site:pcmag.com OR site:techradar.com review score"

Extract per product:

price — current retail price
image_url — direct CDN/product image URL (from brand site, NOT Amazon thumbnail)
buy_links — up to 3: Amazon, brand official, Best Buy, etc. — must be exact product URLs
ratings — from as many of these sources as you can find:
- amazon: {"score": 4.5, "count": "1,234 ratings"}
- google: {"score": 4.3, "count": "500+ reviews"}
- reddit: {"score": 4.0, "label": "Highly recommended"}
- rtings / pcmag / wirecutter / techradar — use "max": 10 or "max": 5 as appropriate
pros — 4–6 concrete bullet points from reviews
cons — 3–5 real weaknesses from reviews
summary — one expert sentence (e.g. "Pack Hacker's top pick for one-bag travel")
recommended — true for the single overall best pick
tag — e.g. "In Your Cart", "Best Value", "Recommended", "Budget Pick"

Image URL tips:

Fetch the brand's product page to find CDN image URLs (usually cdn.shopify.com, brand.com/cdn/shop/files/...)
Avoid Amazon image URLs (often blocked). Use brand site or retailer CDN URLs.
URL should end in .jpg, .webp, or .png with no login required

Ratings accuracy: Only include ratings you actually found via search — never fabricate scores.

3. Build the JSON Data File

Write all product data to /tmp/pr_research.json as a JSON array:

[
  {
    "name": "Product Name — Variant",
    "price": "$99.99",
    "tag": "In Your Cart",
    "image_url": "https://cdn.example.com/product.jpg",
    "buy_links": [
      {"store": "Amazon", "url": "https://amazon.com/dp/ASIN", "price": "$99.99"},
      {"store": "Official Site", "url": "https://brand.com/product"}
    ],
    "ratings": {
      "amazon": {"score": 4.5, "count": "1,234 ratings"},
      "google": {"score": 4.3, "count": "500+ reviews"},
      "wirecutter": {"score": 4.0, "max": 5, "label": "Top Pick"}
    },
    "pros": ["Pro 1", "Pro 2", "Pro 3"],
    "cons": ["Con 1", "Con 2"],
    "recommended": false,
    "summary": "Optional expert one-liner"
  }
]

4. Render the Canvas

Find the render_canvas.py script inside this skill's scripts/ directory.

python3 /path/to/product-researcher/scripts/render_canvas.py \
  --data /tmp/pr_research.json \
  --title "🔍 [Category] Research"

Override the canvas host if needed (e.g. for Tailscale/remote access):

python3 render_canvas.py --data /tmp/pr_research.json --host YOUR_HOST_IP --title "🔍 ..."

Or set it via environment variable:

CANVAS_HOST=YOUR_HOST_IP python3 render_canvas.py --data /tmp/pr_research.json --title "🔍 ..."

The script auto-downloads images, builds a summary highlights bar (Best Value / Top Pick / Highest Rated), and renders all product cards with ratings, pros/cons, and buy links.

To update an existing canvas session instead of opening a new tab:

... --session SESSION_ID

The canvas lib path defaults to ../../live-canvas/claude/implementation relative to the script. Override with CANVAS_LIB_PATH env var if your live-canvas skill is installed elsewhere.

5. Report to User

Share the canvas URL: http://CANVAS_HOST:18793/?session=SESSION_ID
Summarize top picks in chat: best value, overall winner, best for their specific use case
Note that buy links require right-click → open in new tab

Tips

Buy link validity: Verify the ASIN or URL slug points to the exact product — not a search page
Image fallback: If image fails to download, the script shows a 📦 placeholder automatically
Parallel searching: Run multiple WebSearch calls in parallel to speed up research on large product lists
Session reuse: Use --session to push updates to the same canvas tab without opening a new one