Keep long jobs sharp: helpers, files and the cheapest tool f

How this helps you

Big jobs finish with a clean summary and a file, and the chat stays usable for the next step.

✦ Try it

This is a big job: [job]. Hand the heavy parts ([heavy parts]) to helpers so this chat stays short. Save the full results to [file name] in your workspace and send me a summary of no more than [lines] lines plus the file link. Use the cheapest reliable method first: what you already know, a connected app, web search, the browser, then the full computer. Only ask me if all of those fail.

Fill in: [job] is the whole task. [heavy parts] are the steps that read or produce a lot. [file name] is where the full output goes. [lines] keeps the chat summary short, such as 10.

What they did

In a Stanford CS146S lecture, Lee Robinson explains how Grok Bot's always-on harness works. A Bot's main chat has a limited context window, so expensive work is handed to helpers (sub-agents) with fresh working memory; only their result comes back to the main conversation, which is what makes it feel like the context never runs out. Long outputs go into files rather than the chat, because the model is good at searching files later, and screenshots or other costly work go to a helper. The harness tries the cheapest, most reliable option first: what it already knows, then a plugin or API, then web search, then the browser, then the full desktop, and only then asks you. On web pages it can read the accessibility tree instead of screenshots to save tokens. Your messages take priority and interrupt the current turn, so you can steer mid-task. Most of this happens automatically; the steps below are our suggestions for asking for big jobs in the same shape.

Who it's for: People running long research, data or coding jobs who notice quality dropping late in a chat.

Before you start

A job big enough to matter: many pages, many files, or a long research run.
Optional: Michael Fenech's website-reading rule (see related) saved on your Bots.

How to try it

Split the job: say which parts are heavy (reading 50 pages, processing screenshots, long shell runs).
Ask for those parts to go to helpers, each returning a short result (we haven't verified that naming helpers changes behaviour).
Ask for the full output saved as a file in the workspace and only a summary in chat.
Tell it the order of tools to try, cheapest first, and to ask you only when all of them fail.
If priorities change mid-run, just send a message; it takes priority over queued work.
For repeat jobs, start a fresh chat or a routine rather than piling onto a long thread.

What to expect

A ten-line summary in chat, a link to a file with the full table, and a note on which method it used for each source.

Common mistakes

Pasting huge outputs back into the main chat, which fills its context and makes later turns worse.
Treating this as a guarantee. Lee describes how the harness is built; how a given job is split is up to the Bot.
Asking it to bypass sign-ins or bot protection when the cheap methods fail. Ask it to stop and tell you.

Related tips

A /clean-start skill for long chats: Save the goal, decisions, files and next action to a handoff file, then pick up in a fresh chat.
Connectors before the browser: One task got about 45x cheaper by switching to a connector.
Make your Bots read websites with a script, not screenshots: One test: 213K tokens by clicking around, 51K with a browser script, almost none with a plain script. Give your Bots the rule.
How to trust your Bot's work more: Put most of your effort at the top: code structure, then lint/tests/CI, then prose rules, then human review.
Turn your YouTube Watch Later pile into a learning site: A Bot reads the transcript of every video in your Watch Later list and builds a site with summaries, timestamps, a quiz and a watch or skim verdict for each.