← Back to library

Make your agents read websites the cheapest reliable way

Michael Fenech (@Michael_Fenech_) ran one extraction job on a practice site three ways and reported roughly 213K tokens with computer use, about 51K with a browser read by code, and zero model tokens while a plain script ran. He turned that into a ranked-method rule for Grok Bot. We show how to use it on any agent platform.

  • OutcomeGive your lead agent a standing rule for pulling information from websites: use a connector first, then a plain script, then a browser page read by a script, and only fall back to screenshots and clicking when nothing else works, with result checks, clear failure reports and saved scripts for repeat jobs.
  • SetupUnder an hour
  • LevelIntermediate
  • CreditMichael Fenech (@Michael_Fenech_)
View original
Works with
Grok BotClaude CoworkChatGPT agentClaude Code

Published Oct 29, 2026Updated Oct 29, 2026AI dry run on sample input: passed Oct 29, 2026Source verified Oct 29, 2026

When should you use it?

  • Your agents collect prices, listings, reports or tables from the same sites every week
  • Usage allowance runs out quickly and runs show long stretches of screenshots and clicks
  • Results from browsing agents sometimes come back with shortened titles or missing rows

Not for: Getting around sign-ins, CAPTCHAs, paywalls or a site's terms; the rule tells agents to stop and ask instead; One-off lookups where writing and testing a script would take longer than just browsing; Sites whose data you have no right to collect or reuse

What do you need, and what do you get?

You need

  • A lead agent that can edit or propose instructions for your other agents and routines
  • A list of recurring website jobs and the fields each one must return
  • Any connectors or official data feeds you already have for those sites

You get

  • A proposed extraction rule for each relevant agent or routine, waiting for your approval
  • For each job, the method used and why cheaper options did not fit
  • Record counts and field checks against the source, with problems reported rather than hidden
  • Saved, tested scripts with instructions for repeat jobs, and measured usage where available

How does it work, step by step?

  1. List the website jobsWrite down every routine that reads a site, how often it runs and what it must bring back.
  2. Paste the ruleGive Fenech's prompt to your lead agent so it can propose the method order for each agent and routine.
  3. Approve the changesReview which jobs move to a connector or script and which still need a browser, then approve the instruction updates.
  4. Check the first runsCompare counts and key fields with the site by hand on one or two runs before trusting the new method.
  5. Keep the scriptsMake sure working scripts are saved with their checks so the next run reuses them instead of rebuilding.

The prompt

The original prompt by Michael Fenech (@Michael_Fenech_), word for word, with a link to where it was published. Below it is our adapted version with fill-in blanks, if you want it. Works in Grok Bot, Claude Cowork, ChatGPT agent and others. Values in lime are placeholders.

Original prompt by Michael Fenech (@Michael_Fenech_)

Word for wordby Michael Fenech (@Michael_Fenech_)
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.

Use an existing connector if it covers the task and returns the information we need.

2: A plain script.

For public data or an authorised official data feed, fetch and process it directly. Return only the information needed for the task. Avoid opening a browser unnecessarily.

3: The browser, read by a script.

If the page needs a browser, check whether supported browser scripting can read its contents accurately instead of repeatedly analysing screenshots.

4: Computer use.

Use screenshots, scrolling and clicking for the remaining steps that need visual interaction or cannot be completed reliably by the methods above.

Apply these rules:

Check the result. Verify record counts, required fields and key details against the source. Don’t silently shorten, omit or invent information.

Fail clearly. If a method fails or returns incomplete or suspicious data, report the problem before trying the next suitable method.

Reuse successful scripts. Save tested scripts with instructions and validation checks so recurring jobs don’t require rebuilding them.

Keep context lean. For repeat jobs, use a routine or a short, fresh chat with the necessary instructions and inputs. Avoid carrying unrelated conversation history into every run.

Respect access limits. Use supported access methods. Don’t extract protected credentials or bypass sign-ins, CAPTCHAs or bot protection. Stop if human action is required.

Report the method. Tell me what you used, why earlier options weren’t suitable, and any limitations. Report measured usage where available; label estimates.

Propose this rule for the relevant Bots and routines, then wait for my approval before changing their instructions.

Exactly as published by Michael Fenech (@Michael_Fenech_) in X post: a method order that makes Grok Bot read websites with less model work.

Our adapted version (optional)

Fill in the blanks (optional; the copied prompt uses your values)

Adapted prompt2 variables
Set a rule for how my agents on {platform} collect information from websites. For each job in {job_list}:

- Prefer, in this order: an existing connector or official API, a script that fetches the pages directly, a browser page read by a script, and only then screenshots and clicking for the steps that need it.
- After each run, compare row counts and required fields with the source and tell me about gaps instead of filling them.
- If a method fails, say so and why before you try the next one.
- Save working scripts with short instructions and their checks so we reuse them.
- Never handle my passwords, solve CAPTCHAs or bypass sign-ins; stop and ask me.
- Tell me which method you used and the usage it took, marking estimates.
Show me the proposed instruction changes for each agent and wait for my OK.

What can go wrong?

  • A script that silently breaks when a page layout changes is worse than a slow agent. Keep the count and field checks in every run.
  • Token savings depend on the site and the job. Fenech's figures come from one practice-site test, so measure your own before and after.
  • Some sites forbid automated collection even when it is technically easy. Check terms before switching a job to a script.
  • Do not let the rule become a reason to store passwords in scripts. Sign-in steps stay with you.

Our take

Watching an agent screenshot, scroll and click its way through a table it could have read in one request is one of the more expensive habits in agent setups. Fenech's test puts rough numbers on it, and his prompt turns the lesson into a rule the agent can apply on its own. We like the order of preference, but we like the guardrails more: verify counts and fields against the source, report failures before switching methods, keep sign-ins and CAPTCHAs with a person, and label usage estimates as estimates. Those lines stop a cheaper method from quietly becoming a less accurate one. Two cautions from us: check each site's terms before moving a job to a script, and expect some maintenance, since scripts break when pages change. In our dry run with a fictional supplier catalog of 40 public product pages and a supplier portal behind a login, the agent moved the catalog to a plain script that returned 40 of 40 items with full titles, stopped at the portal's CAPTCHA and asked for help, saved the catalog script with a count check, and proposed the rule for two bots, waiting for approval before changing anything.

Where the time goes back: Recurring website jobs stop paying for screenshots and clicks on every run, and saved scripts with checks make the next run faster and more complete.

FAQ

Do I need to know how to code?
You do not. The rule tells the agent when a script is the better choice; the agent writes, tests and saves it, and you approve the change.
Does it work outside Grok Bot?
Yes. Any agent that can run code and also drive a browser, such as Claude Cowork, Claude Code or ChatGPT agent, can follow the same order.
Why keep computer use at all?
Some steps really need visual interaction, such as a custom date picker. The rule keeps it for those steps only rather than whole jobs.

Source and credit

The original prompt on this page is by Michael Fenech (@Michael_Fenech_), shown word for word with a link to where it was published. The summary, notes and adapted version are ours.

Original
X post: a method order that makes Grok Bot read websites with less model work
Creator
Michael Fenech (@Michael_Fenech_)
Published on
X
Original date
Oct 28, 2026
License
Not openly licensed; the original prompt is shown word for word with credit and a link to the original
Originally for
Grok Bot
AI dry run
Oct 29, 2026: passed. Dry-run tested by our AI test agent on sample input. Followed the method order, verified counts and fields, failed clearly at the CAPTCHA, saved a reusable script, labeled estimates, waited for approval.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails.
Verified
Oct 29, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.

Are you the creator? Claim or correct this page.

More in Operations & Admin

All 203 Operations & Admin playbooks →

Related playbooks

Operations & AdminWorkflow✓ AI dry run

Log a sales order after checking its paperwork matches

Give one spoken instruction and get the order logged, cross-checked against its paperwork, the commission tracked and a client email drafted.

Credit: Elvis Rigglet (@rassailor)

An afternoon+
Operations & AdminPrompt✓ AI dry run

Audit how your agents pick tools before buying more usage

Have your lead agent review the recurring jobs it can see, note which apps, scripts, browser sessions or full computer control each one relies on, flag jobs that take a heavier route than needed or rely on remembered facts, check permissions and fallbacks, and then propose three side-by-side experiments before anything is changed.

Credit: Michael Fenech (@Michael_Fenech_)

Under an hour
Operations & AdminPrompt✓ AI dry run

Audit your agent routines for steps you can cut

Have your lead agent study one recurring routine, write down what a good result looks like, propose a shorter path to it, test old and new side by side on the same inputs, and switch only after you approve, with the old version kept for restore.

Credit: Michael Fenech (@Michael_Fenech_)

Under an hour
Operations & AdminSkill✓ AI dry run

Price a custom manufacturing job with every cost traced

Use the quote-build skill from Anthropic's Claude for Manufacturing plugin to price a make-to-order job from the intake form, similar past jobs and cost history, citing the purchase order or work order behind each cost, spreading setup across every quantity break, keeping margin on a separate internal line and checking the promised date against shop load, with a person releasing the quote.

Credit: anthropics

An afternoon+
Operations & AdminPrompt✓ AI dry run

Turn one recurring task into a repeatable Claude workflow

Choose one chore you repeat every week, such as a status update or a meeting brief, and set it up in Claude as a reusable project: a clear deliverable and reader, saved background, an approved example, tidy inputs, questions before drafting and a review step, so later runs need only light edits before you think about scheduling it.

Credit: Paul Storm

Under an hour