web

Imported from roiguri/jarvis-agent (tools/web/SKILL.md). Install upstream with npx skills add roiguri/jarvis-agent --skill web. Copyright stays with the author.

• Having a URL and searching for it are different problems. A URL or a post id is an address — fetch_url reads it. web_search finds an address you don't have yet. Searching for an address cannot work, however many times you rephrase it.
• A link appearing in conversation is not by itself a reason to fetch it. But before you act on one — summarizing it, filing it, judging it, describing what it is — read it. Producing a description of a page you have not opened is the failure to avoid, not leaving a link unread.
• Never describe a page you could not read. If fetch_url fails, say so plainly, record the bare URL, and ask the owner what it was. A guessed description that lands in memory outlives the conversation and is worse than an admitted gap — this has already happened twice, and both times the guess was wrong.
• Never infer a page's content from the conversation around it. The link the owner sent before this one tells you nothing about this one; two links in a row are not by the same author.
• When a search does not find what you need, stop and say so. Rephrasing the same query repeatedly costs real money and does not turn an unanswerable search into an answerable one.