web-data

Web data jobs with Spidra

The Spidra server already tells Claude which tool does what. This skill is about doing the whole job well: the steps, the defaults, and what a good answer looks like. Every Spidra call spends the user's credits, so the habit throughout is a small first step, then more only if the user wants it.

Lead lists and data from popular sites

Businesses on a map, products, listings, jobs, reviews, profiles:

1. spidra_find_templates with a word or two from the request. Spidra adds templates often, so look before deciding there isn't one.
2. spidra_get_template to see its inputs. Fill them from the request; don't invent inputs it doesn't list.
3. For anything over a few dozen rows, run it with "estimate": true first and tell the user what it will cost.
4. Run it. When the user gave no number, start small (10 to 20 rows) and say more are available.
5. Answer with a Markdown table of the columns the user asked for, then one line with how many rows came back and the credits used. Offer the rest of the fields or more rows; don't dump them.

If no template fits, a link from the site with "output": "json" and no prompt often comes back with the page's data already in named fields. Try that before writing an extraction prompt.

Research with sources

When you already know the page that answers the question (official docs, release notes, a company's own page), read it directly: scrape it with no prompt to get Markdown for 1 credit, and summarize it yourself. A summarizing prompt on a long page can cost 20 credits or more.

Otherwise:

1. spidra_search first. Use the news source for anything recent, and a time range when the user says "this week" or "latest".
2. Answer from the snippets when they're enough. When they aren't, read the 3 to 5 most useful results in one spidra_batch_scrape as Markdown, not one scrape each and not every result.
3. Write a short answer where each claim links to the page it came from, and say how recent the sources are. If sources disagree, say so.