Specialists

Researching on the web

Cited research that reads pages properly, escalating to a real browser when a page needs one.

Code
/research how does Stripe handle idempotency keys on retries
/research --deep migrating from Redis pub/sub to streams
/research --browser pricing on the competitor's site

What you get back

A written answer with its sources, saved to your project. Not a list of links — the pages are read, and the claims point at where they came from.

Because it's saved, "what did we find out about idempotency keys" works months later, and a teammate in the same workspace can ask it too.

The flags

FlagEffect
--quickA few sources, fast. For settling a specific question.
--deepWide sweep, more sources, longer. For "I need to understand this area."
--browserDrives a real browser, for pages that need JavaScript or a login wall.
--proHigher-capability models throughout.

Without a flag it picks a depth from the question.

When it beats reading the docs yourself

  • The answer is spread over several pages and you want it in one.
  • The official documentation is wrong or out of date and you need the issue thread that says so.
  • You want the answer recorded against the project rather than in a browser tab you'll close.

When it doesn't

Anything about your code. Use Searching your project — the answer is in the repository or in your own history, and the internet doesn't have it.

Cost

Deep research reads a lot of pages, and each is model input. --quick is usually the right default. /usage shows what a session spent.

If it doesn't work

A page came back empty. It needed JavaScript. Rerun with --browser.

The sources are thin. The question was too broad. Narrow it — a specific version, a specific error, a specific product.

It's taking a long time. --deep is minutes. Run it in the background and carry on with something else.

Next