Grok Bot handbook, part 2: 14 practical uses for its new features, from Vault and Teach a Task to shared bots

This article is the second installment of a two-part Forward Future guide to Grok Bot's recent features. Each section explains what is new, why it matters and how to make use of it, and includes a sample prompt. The text received is cut off at item 38, the last case shown, so this summary covers only items 25 through 38. The features are grouped into three blocks: Vault and 1Password, Computer Use with Teach a Task, and Shared Bots with multi-agent workflows.

Block 1: Vault and 1Password. According to the article, Grok Bot added Vault, a store for credentials, payment methods and contact details, and expanded its integration with 1Password. The core idea is that the bot can work inside authenticated websites without the user pasting passwords or card details into a chat. The text presents this as more significant than it seems, because a lot of repetitive work happens on websites that lack useful integrations.

Item 25: preparing purchases and handling checkout. With Vault, the bot can fill in saved details in supported payment flows. It can search for the product, check availability, prepare the cart and show the total, within price, brand and delivery limits. The suggested prompt asks it to stop before payment and not to accept substitutes or additional terms. The article includes a security warning. Vault cards can be used without fresh approval for each transaction. Anyone who wants to authorize every payment personally should use a single-use payment flow instead of a saved card.

Item 26: accessing websites through 1Password. This is the case the author finds most compelling, because it does not involve spending money. It can be used, for example, to log into a supplier portal, download invoices, pull reports from a business account or collect data from a service without a convenient API. The user approves the login and the credentials are filled into the bot's browser without exposing the password to it. The integration is rolling out gradually and requires a compatible version of 1Password. Some sites will still require the user to take over for additional verification.

Block 2: Computer Use and Teach a Task. Grok Bot can operate websites and desktop software. Teach a Task lets users turn a demonstrated browser workflow into a reusable skill. It is especially useful when there is no API or ready-made integration.

Item 27: finding buyers for something you are selling. One user described using the bot to sell a domain name. It compared marketplaces, researched companies that might want it and drafted personalized messages. The prompt asks it to compare five marketplaces and their fees, identify up to 20 plausible buyers, look up public contacts and draft an email for each one. It also requires links to sources, flagging uncertain matches and not contacting anyone or creating paid listings without approval. The author recommends a short list of credible buyers over mass outreach.

Item 28: negotiating a lower internet bill. The article cites the example of a user on X (@mikepat711) who asked the bot to negotiate with his provider through its official support chat. He supplied a redacted bill, explained what he wanted and spelled out what he would not accept. The bot reportedly took about 30 minutes and found an offer that would save about $58 a month. The text itself notes that this is one person's result. Even if a better rate is not secured, the bot can explain the options and prepare a script for a follow-up call.

Team pick: a bot for health insurance claims. A member of the Forward Future team proposes a bot that untangles insurance coverage, becomes an expert on the plan and helps with claim disputes. It would organize policies, explanations of benefits, bills and correspondence, flag discrepancies, prepare questions or draft appeals and track pending deadlines. The author stresses treating it as an assistant for document analysis and claims administration, not as an authoritative interpreter of insurance law or medical necessity. He also warns that great care is needed with the personal data it accesses.

Item 29: adding a phone number to the “Do Not Call” registry. An entrepreneur (@MinimalDuck) asked the bot for help cutting down on spam calls. According to the account, the bot completed registration on the US National Do Not Call Registry, including the confirmation step. The article notes that the registry does not stop illegal calls and that the process varies by country. It recommends using only the official registry, donotcall.gov.

Item 30: having the bot operate another AI tool. One user reportedly directed from his phone as Grok Bot installed, configured and ran Hermes Desktop on the bot's computer, then captured the results. The author treats this as an advanced experiment that, according to reports, is possible but involves a lot of trial and error. The suggested prompt restricts installation to official sources and requires the bot to ask before installing additional software, using credentials or starting paid services.

Item 31: teaching the bot a task by showing it. This is one of the features the author finds most interesting. Teach a Task records browser interactions for up to ten minutes and turns them into a draft skill. The article mentions that Kevin Stratvert's YouTube channel covered it using accounting tasks in QuickBooks. The example is a weekly task of opening a supplier portal, downloading a report, renaming it and saving it to Drive. The steps to get started are to open a one-on-one conversation with the bot and the computer view, choose Teach a task if available, explain the goal and demonstrate it. Next, review the generated skill, especially its exception handling, and test it before setting up a recurring routine. The recording captures visible browser interactions, not microphone audio. The feature is still rolling out. If it is not available, the skill can be created from written instructions. The author advises starting with a simple, predictable workflow and highlights the feature as a way to lower the barrier to building automations without touching APIs.

Block 3: Shared Bots and multi-agent workflows. “Team Bots” appeared in late September and are now called Shared Bots. They allow teams to share specialized assistants with common instructions, tools and knowledge. The text cautions against elaborate collections of bots that spend more time coordinating than working.

Item 32: updating the CRM after every customer call. Working from a sales demo, the bot would extract new contacts, objections, commitments and next steps from the transcript. It would compare them against the CRM, propose changes with references to the supporting passages and prepare a follow-up email. Nothing is written or sent without approval. The author urges caution with inferred commitments and deadlines: if the customer did not explicitly agree to something, it should not be logged as a confirmed decision.

Item 33: pairing a research bot with a sales outreach bot. A testimonial from the company Above the Line describes one bot that researches events and decision-makers and another that turns that research into personalized messages. The pattern works for prospects, partners or candidates. The advantage is that the research can be reviewed and reused before drafting. The advice is to start with two bots, not five.

Item 34: turning a request for proposal (RFP) into an itemized estimate. Above the Line says it completed in about 30 seconds a job that previously took a producer much longer. The article itself notes that actual time savings will depend on each business. The prompt asks the bot to use a rate card and comparable historical estimates, list assumptions, exclusions and risks, and treat everything as a draft until reviewed by an authorized person.

Item 35: routing unusual sales terms to the right approver. A “Deal Bot” would review non-standard requests, such as discounts, payment terms or contract changes. It would classify them as standard, requiring finance review or requiring legal review, and prepare a summary for the approver. The bot never approves discounts or makes legal decisions. It only streamlines the process.

Item 36: a research-to-publication pipeline. One creator described four bots. The first monitors X accounts and gathers news. The second evaluates and drafts. The third checks that editorial approval exists. The fourth hands approved work to a coding agent that prepares the website update. Nothing is published without a final human decision. The author urges maintaining substantial editorial review, because running an inaccurate story through four bots does not make it more accurate. One team member highlights the workflow for connecting tasks that are usually handled separately and for being reusable across newsletters, articles and social content.

Item 37: grouping related support tickets and preparing batch replies. The Grok Bot team reportedly described the case of an engineer who approved replies to 204 tickets in about ten minutes. The bot would group tickets with a likely common cause, investigate and draft consistent replies. The author recommends being conservative when grouping, because a mistake sent to 200 people is far worse than one in a single draft. He also cautions that the figures come from the company itself and should be read as an example, not as a universal productivity benchmark.

Item 38: a shared AI teammate for engineering. The Grok Bot team reportedly used a Shared Bot connected to Slack, Notion, Linear, Datadog and Cursor to follow project discussions, triage bugs, create tickets and launch coding agents. The company says a five-person team shipped more than 100 pull requests a day during the project. The text describes this as a vendor-supplied figure. The content received cuts off just as the prompt for this item begins, so there is no further detail or concluding section from the article.

Cross-cutting takeaways. Nearly all the prompts in the article share the same design: narrowly scoped tasks, sourced results, human approval before paying, sending, publishing or modifying records, and explicit prohibitions on what the bot must not do. Several figures come from user testimonials or the companies themselves, and the author presents them with caution. For readers, the practical lesson is to start with simple, predictable workflows, test before scheduling a routine, and keep access to sensitive data in mind when connecting credentials, payments or personal documents.