How this helps you
Repeat website jobs use far less of your usage and keep full, checked data.
Help my Bots use less model work when extracting information from websites, without sacrificing accuracy or completeness. Use this order, stopping at the first suitable method that works: 1) a connected app, if one covers the task; 2) a plain script, for public data or an authorised official data feed, returning only what the task needs; 3) the browser, read by a script, if the page needs a browser; 4) computer use, only for steps that need visual interaction. Rules: check the result against the source and don't shorten, omit or invent anything; fail clearly and report problems before trying the next method; save tested scripts with instructions and checks for reuse; keep context lean with a routine or short fresh chat for repeat jobs; respect access limits and never bypass sign-ins, CAPTCHAs or bot protection; report the method you used and label estimates. Propose this rule for the relevant Bots and routines, then wait for my approval before changing their instructions.
Fill in: No placeholders. This is a condensed version of Michael's prompt; the full text is in his post.
What they did
@Michael_Fenech_ asked Grok Bot to extract the same data from a practice website three ways and recorded approximate tokens: computer use (screenshots, scrolling and clicking) took about 213K; opening the page in the browser and reading it with code took about 51K; a plain script that fetched the pages directly used no model tokens while it ran, with model usage still needed around the run. The scripts also kept full titles that computer use shortened. His conclusion: a Bot doesn't always need to look at a site to read it. He wrote a rule for the Chief of Staff with a fixed order (connected app, plain script, browser read by a script, then computer use) and six checks, ending with an instruction to propose the rule and wait for approval. You don't write the script; the Bot does.
Who it's for: Anyone with routines that read the same websites again and again.
Before you start
- A chief of staff or main Bot that can update other Bots' instructions with your approval.
- A repeat website task to try it on, such as a weekly price or listings check.
How to try it
- Paste the prompt below into your chief of staff.
- Read the rule it proposes for each Bot and routine. Approve only where it fits.
- Run one of your repeat website jobs and ask which method it used and why.
- When a script works, ask it to save the script with instructions and checks so the routine reuses it.
- Compare usage before and after in Settings, if your job runs often.
What to expect
A proposal listing which Bots and routines would get the rule, then, on the next run, a note like: used a plain script, 214 rows, matched the count on the page.
Common mistakes
- Expecting his exact numbers on your sites. They come from one practice-site test.
- Using scripts to get around sign-ins or bot protection. His rule says stop and ask instead.
- Letting the Bot rewrite every Bot's instructions without reviewing the proposal.
Related tips
- Connectors before the browser: One task got about 45x cheaper by switching to a connector.
- Make routines cheaper: events, slow schedules, change checks: A 5-minute routine wakes 8,640 times a month. Use an event trigger, the slowest schedule that works, and a cheap change check.
- Keep long jobs sharp: helpers, files and the cheapest tool first: How the Grok Bot harness avoids running out of context, and how to ask for work in the same shape.
Worth knowingThe token numbers come from one practice-site test, so treat them as a direction, not a promise. The prompt here is condensed; the full version is in his post.
Learn more: How to Use Grok Bot: A Beginner's Guide (2026) · All Grok Bot connectors tips