What I scraped: books.toscrape.com, the practice catalog built for scraping. I took 5 categories (Poetry, Travel, Mystery, Classics, Philosophy) and ended up with 92 books in a CSV with title, category, price, star rating, stock and URL. The counts match what the site shows for each category, there are no empty fields and no duplicates, and I spot-checked 3 books against their detail pages. Total cost was about 23 of my 500 free credits, and most of that went on testing at the start. Screenshot below. One thing I learned: I expected Claude Code to pick a tool and run it. It tested formats first instead. Markdown came back clean but dropped the star ratings, because on this site they only exist as a CSS class. The "query" format, where you ask the page a question, cost 5 credits a page and still missed them. What worked was raw HTML filtered with includeTags down to just the title, price, rating and stock elements, at 1 credit a page. Claude also told me straight away that the full catalog (about 70 pages of HTML) was too much to push through the chat and gave me two options. I went with 5 categories. One more surprise: the Firecrawl connector in the Claude app has no separate crawl, map or extract tools. You get scrape, search and an agent that handles multi-page jobs. One use case idea for clients: outreach lists for small businesses. Scrape a directory of Spanish real estate agencies or dental clinics and record name, city, website and whether the site has a booking form or chat. Every row without one is a pitch for an AI assistant.