compare-products
Compare N products from URLs and render a comparison table (HTML or Markdown)
Instructions
You will receive product_urls (array) and dimensions (array of attribute names). For each URL:
- Visit the page using
browser.navigate. - Take a
browser.screenshotof the hero region (used as thumbnail). - Run
web.extractwith a schema that includes the requesteddimensionsplus title, primary image URL, and current price. - Normalise units (e.g. convert "R$ 1.234,56" → 1234.56 as number,
detect currency).
Once all products are extracted, render a comparison artefact in output_format. Mark the cell as *old*when the value is the best on that dimension (lowest price, highest spec score). Add a short "Notes" paragraph at the bottom summarising which product wins for which use case.
If any product fails to load (timeout, captcha), keep going with the remaining ones and add a footnote listing the skipped URLs.
Steps
- Initialise an empty result table with columns = dimensions.
- For each URL, attempt extraction with up to 2 retries on transient
failure. After 2 retries, mark as skipped and continue.
- Detect "winner" per dimension (lowest price, longest warranty,
highest weight capacity, etc.) and apply bold formatting.
- Render and save to
comparison.
Examples
*ser:*"comparar essas 3 cadeiras [3 URLs]" *xpected:*HTML page with 3 columns, pricespecsphoto, winner highlighted, "Notes" paragraph with a buy recommendation.
*ser:*"compare these two laptops in markdown" *xpected:*Markdown table with 2 columns, plus a 2-3 sentence notes section.
Caveats
- Do *ot*include personal opinions in the Notes paragraph; stick
to objective trade-offs (e.g. "X is heavier but supports more weight").
- Do *ot*fabricate specs — if a dimension wasn't extracted, write
"Not specified" in that cell, never invent.
- Currency: always show the original currency string in parentheses
for trust ("R$ 4.993,00 (≈ USD 1,050)").
References
- Manus screenshot 2026
0514 (the original cadeira-comparativa flowinspiring this skill)
- TOOLS-013 (skill format)
- SANDBOX-010 (browser tool)
- AICORE-114 (web.extract)