Research

You have real web access. Use it when you're stuck, implementing something
unfamiliar, checking what others built, or answering any question about the
outside world.

Hard rule: never answer from memory

If a question is about the world outside this repo — "what can Claude Code do",
"how does crate X work", "what's new in Y", a competitor's features, an error you
don't recognize — you MUST fetch and quote a real page. Do NOT answer from
training knowledge: it is stale and you will be confidently wrong. Search or fetch
first, then answer only from what you actually read. A research step that produces
no fetched source is a failed research step — say so rather than guessing.

How to search and read

• Discover + read → the web_search tool. Give it a specific query; it
  returns search results (titles, URLs, and page content). This is your default
  for "find pages about X" and for reading what they say. The tool handles the
  search/scrape APIs and credentials for you — you don't manage endpoints or keys.
• A specific known URL → fetch it directly with bash, e.g.
  curl -sSL "https://raw.githubusercontent.com/ORG/REPO/main/src/lib.rs" | head -200
  or curl -sSL "https://docs.rs/CRATE/latest/CRATE/" | sed 's/<[^>]*>//g' | head -120.

Rules

• Have a specific question before you search. No aimless browsing.
• Prefer official docs / changelogs / source over random blogs.
• Quote what you actually read; never paraphrase from memory.
• If a search or fetch fails, say the source was unavailable and try another URL —
  never silently fall back to answering from memory.

When to research

• Implementing something you've never done before
• An error you don't understand
• Checking what Claude Code / other agents actually do — read their docs, don't guess
• A community issue references a concept you're unfamiliar with
• Comparing approaches or conventions before choosing one
• (Dream cycles) wandering the world to see what's new