The unpaid job inside most companies is glue work. Two calendars. A shared inbox. A three-page email from a teacher, a coach, a sponsor, or a utility. Someone has to read it, decide whether it matters, put it on the right calendar, and not drop the conflict. That someone is usually you, and you will make mistakes.
Grokbot AI agents are for that layer. Matt Berman’s rule is blunt: do not hand the creative brief to a model. Automate the back office, the follow-ups, and the research you keep postponing because it is not worth a Saturday. Connect email and calendar once, drop a specialist bot on a recurring mess, and keep a human on voice, price, and anything that has to sound like you.
If you do this well, you stop being the search engine between Slack, Notion, and your inbox. You spend the recovered attention on the story, the deal, or the decision only you can make.
Key takeaways
- Use Grokbot AI agents on tedium first. Berman’s team automates sponsor management, meeting notes, follow-ups, and proposals; they still invent the work themselves.
- Connect email and calendar once. After that, the bot should already know where to look instead of making you re-explain the stack.
- Summaries are not enough. A useful bot recommends the action (“tell the coach if you can show up”) and writes the event to a shared calendar.
- Personal glue problems map onto business ones: Marketplace haggling is vendor negotiation; school-email triage is sponsor-thread triage; a utility-plan audit is a supplier audit.
- Prefer a cloud agent with no local setup over Mac Minis, VPS boxes, and homemade service accounts. Logins still happen; passwords should sit in a vault the agent cannot see.
- When too many bots come back at once, talk to one orchestrator (a chief of staff bot) and let a bot-factory bot spawn specialists that report in.
- When a bot writes slop, do not just fix the output. Ask why, then add a one-line skill so the failure does not repeat.
- Computer use beats custom connectors for many sources. Point the bot at the directory, the utility portal, or the ATS and let it drive the browser.
Who this is for and who should skip it
This is for a founder, operator, marketer, or ops lead who is already drowning in email, calendars, and half-finished follow-ups. It also fits a lean content business (Berman’s example is Forward Future and sponsors) and anyone running outbound that looks like a job search: find the person, tailor the pitch, keep a critic on voice. You need a Grokbot account, willingness to connect mail and calendar, and comfort granting site access when the bot asks you to log in. You do not need to stand up a home server.
Skip it if you want a model to decide what to make. Berman tells people not to use AI for the actual creative process of coming up with ideas. Skip it if you will not grant inbox or calendar access, or if you want a fully hands-off customer-facing agent with no human on negotiation, brand voice, or facts. The Cursor and Remotion path later in this playbook assumes you have a repo and can stand a coding agent next to Grokbot; that part is optional.
Automate back-office work before you touch the creative brief
What you are trying to accomplish: strip managerial tedium so human time goes to craft. Berman’s observation is that the valuable skill is still how well you can make a compelling story. His team (he names Brian and Alex) only recently got real leverage on video editing with AI: dictation, multiple edits in parallel, work that he says did not even seem possible a few weeks earlier. Everything behind the scenes was already fair game.
Prepare a short list of work that is repetitive, text-heavy, and currently living in your head: sponsor threads, meeting notes, follow-ups, proposals, inbox triage. Do not start by asking a bot what content to publish.
- Draw a line: idea selection and story stay human.
- Put Grokbot on notes, follow-ups, inventory, and proposals.
- Treat new creative-tooling (editing, motion, ads) as a second wave, after the back office is boring.
Good looks like this: you still choose the work; the bot arrives with a draft, a summary, or a calendar event. Failure looks like using the model as a substitute for taste. Berman’s line is worth keeping on the wall:
Don’t use AI for the actual creative process when you’re coming up with the ideas and deciding what to make.
The host’s matching rule for everything else:
The back office stuff, that is easy to automate, you should automate it.
Connect email and calendar once, then drop a Grokbot on the glue
What you are trying to accomplish: stop being the human middleware between inboxes and calendars. Berman’s highest-use bot is a family bot, but the mechanics are the same if you swap “school and sports” for “customers and sponsors.”
Prepare access to email, a shared calendar (he uses one with his wife), and a channel other people can talk to without learning bot lore. They use a shared Telegram so she can treat it as an assistant, not a configuration project.
- Stand up one bot whose only job is a defined slice of mail (family, sponsors, a single product line).
- Have it read every email and decide whether it belongs in that slice.
- Force a one-to-two-sentence summary plus an action recommendation, not a digest of facts.
- Write events to the shared calendar automatically.
- Notify on conflicts and help resolve them.
- Let it keep preferences (Berman’s example: not taking meetings in the morning) so you are not re-briefing it.
- Share access with the other human who lives in that calendar.
Good output is short, action-shaped, and already on the calendar. “Here’s a three-page email” is a fail. “You need to tell the coach if you can show up” is the bar. Berman says he cannot tell you how much time this single bot has saved him; treat that as a qualitative checkpoint, not a benchmark.
The setup contrast matters. Before Grokbot, the host’s version of this was write a script, parse mail, create service accounts, and maintain the mess. Earlier personal assistants also pushed people toward Mac Minis, VPS boxes, browser control, and local agents that are a poor fit if you want to send a task while walking. Grokbot’s pitch in this conversation is cloud, no setup, complexity abstracted. You still authenticate to the sites it must use. You should not be pasting passwords into the model. The host’s security contrast: older demos dumped passwords in plain text; here, entry goes into a vault, the agent cannot see the password, and the blast radius stays smaller if something is compromised.
Microsoft Recall is the anti-pattern they name: screenshot everything, then ask the computer where an email went, and eat the privacy backlash. The Grokbot version is an opt-in connection to the system of record, not a screen recorder.
Personal Grokbot workflows that map straight onto the business
What you are trying to accomplish: pick problems you already understand in civilian life, then run the same agent pattern at work. The host’s one-liner:
Personal problems map almost directly to business problems.
Sell idle inventory the way you would sell a PlayStation
Berman walked his house, photographed things he had not used (a PlayStation, a MacBook, a mountain bike), and sent the photos to Grokbot with a request to list them. After that, the human work was logging into Facebook Marketplace and eBay. The bot researched the items, pulled comps (including mountain bikes sold in the last six months), wrote complete listings, posted them, watched the threads, tried to answer buyer questions, and followed up indefinitely. If it could not answer, or if a buyer came in low, it asked him how to handle it. He sold the PlayStation and the MacBook; the bike was still in negotiation because he would not ship a 70-pound frame and wanted an in-person handoff.
The business clone is vendor haggling: have the bot learn raw materials, internal deal history, and teammate pricing, then sit in the live thread so you are not pinging three people for context.
Turn meetings into sponsor proposals
For Forward Future, Grokbot knows ad units, pricing, inventory, and whether the company has sponsored before. It joins meetings via Fathom, takes the transcript, and produces a proposal. Berman still corrects when extra context is missing. The checkpoint is not “untouched forever.” It is “right enough that you are editing, not assembling.”
Stop searching Slack to do your job
Onboarding and day-to-day retrieval are the same pattern as family email. Instead of opening the wiki, wandering Slack, and asking a manager, you ask the bot who owns a thing, how to get it done, whether others have been blocked on the same permissions, and who unblocked them. Look in Slack. Look in Notion. The speaker’s result: messages to a manager or teammates for random questions dropped to nearly zero. Multiply that across a company and you get fewer interrupts, not just a faster individual.
The same retrieval habit kills context switching in a busy inbox. You should not hunt a sponsor thread across two email addresses, Drive, and Notion. Ask the bot what they said last time.
Audit bills and subscriptions you would otherwise ignore
Berman forwarded the pattern to money. PG&E kept sending long emails about other plans because he has an electric car. He screenshotted the email, told Grokbot to decide whether to switch and to which plan, and pointed it at twelve months of usage and bills. The bot asked him to log into PG&E, ran the review, and recommended a plan aimed at off-peak charging. He had it double-check the math twice, then switched. He puts the savings at $1,000 a year and the attention cost at about 60 seconds. Treat both as his claim, and copy the method: screenshot, usage history, login, recommendation, human confirmation.
He generalizes from there. Have the assistant scan mail for subscriptions you have not used in six months, even at $9.99, and cancel them so you are not answering seventeen “why are you leaving” questions. Same move on a medical bill you cannot parse: do not spend Saturday morning on it, and do not pay a few hundred dollars just to make it go away.
The ROI point he wants operators to hear: you used to spend scarce attention only on the $10,000 problem and dismiss the $1,000 one. If both take little attention, you do both.
| Use case | Bot / tool | Input needed | Output | Owner |
|---|---|---|---|---|
| Household or team glue | Family-style Grokbot | Email, shared calendar, preferences | 1–2 sentence summary, action, events, conflict alerts | Operator + partner |
| Sell unused gear or excess inventory | Listing bot on Marketplace and eBay | Photos, site logins, ship vs pickup rules | Researched listings, buyer replies, escalations | Owner |
| Sponsor / sales proposals | Grokbot + Fathom | Ad units, pricing, inventory, relationship history, transcript | Draft proposal | Founder or sales |
| Onboarding and retrieval | Workplace Grokbot | Slack, Notion, email | Owners, how-tos, permission history | Each employee |
| Bills and plans | Audit bot | Screenshot, portal login, 12 months of usage | Switch / stay recommendation | Ops or household |
| Outbound and hiring | Finder, writer, critic, applier | Resume, stories, directories, JDs | Email CSVs, letters, applications | Candidate or recruiter-style seller |
| On-brand motion ads | Chief of staff → Dr. Eggbot → Remotion ads bot | Code, design assets, aspect ratios | Square / portrait / landscape spots | Founder or marketing |
Cut context switching with a chief of staff and a bot factory
What you are trying to accomplish: get more done without living in ten threads. The host’s complaint is the real operating constraint: more agents means more work finished and more cognitive load as results stream back.
Lauren’s pattern, as the host tells it, copies a coding-agent setup. You do not talk to every sub-agent. You talk to an orchestrator. She calls hers Steve rather than “chief of staff.” A second bot, Dr. Eggbot, is a bot factory: it knows how she likes to create bots. Steve asks Dr. Eggbot to stand one up, the specialist does the job, and the report lands in the one thread you actually watch.
Prepare a single morning thread and a written taste for how you want new bots built. Then:
- Put daily intake in the orchestrator: what to audit, what to ship, what to research.
- Have it call the factory bot to create a specialist (“energy-bill auditor”) instead of you configuring every agent.
- Require specialists to report back to the orchestrator.
- Resist opening every child thread unless you are debugging.
Berman still peeks. Chief of staff is his morning kickoff; during the day he drifts into individual bots because he likes topics separated. His stated direction is to push more concise summaries into that one thread. He also refuses a slick answer: if you are an engineer with ten cloud agents running, you have become an overloaded manager. People he talks to are a little stressed and trying to do more. That is normal. Step away when you overheat, and use the busyness to ask which threads deserve a human at all.
Good looks like one conversation, several jobs in flight, and short reports. Failure looks like a dashboard of half-read agent chats and no idea what needs you.
Outbound pipelines: find the person, write like yourself, then criticize the draft
What you are trying to accomplish: run sales-shaped work without becoming a full-time researcher. A software intern on the stream treats the job market as four paths: cold apply (big companies drown in hundreds of applications, so speed matters), startups that need people who ship, direct recruiter or employee coffee chats, and being visible in the room. The host’s analogy is correct: you are selling yourself the way you sell a product.
His Grokbot team is a pipeline, not a single prompt.
Recruiter finder and alumni finder
The recruiter finder controls a browser on a VM, searches recruiters, looks them up on LinkedIn, uses the Apollo browser extension to see whether an email is in the database, and copies rows into a CSV. It walks companies hiring software engineers. The alumni finder does the same against a university alumni directory: search people at Google, Microsoft, and similar, take the email the directory returns, append the CSV. Next step he planned: Outlook MCP to send coffee-chat mail from the student account.
There is no custom connector on the alumni side. He logs out of anything sensitive, gives Grokbot the directory, and lets computer use drive search. That is the whole integration story: access to the page, then later access to mail.
Job applier, cover letter writer, and cover letter critic
A SWE job applier hunts roles on GitHub, Greenhouse, and Ashby, reads the job description, and hands it to a cover letter writer. The writer already holds the resume plus a bank of real stories (internships, projects, games). It may do light research on company values, then sends work through a cover-letter repo that coding agents use to generate the PDF and pass it back. The same repo can build a resume. His hard rule:
Do not fake your resume.
Cover letters are the harder thing to automate because they have to connect experience to the role in your voice. He added a cover letter critic as a guardrail: it rejects motivational filler and lines that “don’t sound like you.” He trained it the slow way. Generate a letter, mark the bad patterns (short, empty, inspirational AI slop), feed it papers he wrote before that slop was common, and lock in habits he actually has: compound-complex sentences, some gerunds, enthusiasm that still names the work. The critic then scores four paragraphs: role and why; experience one tied to the JD and to an experiences file; experience two, same test; close with impact you want to have at that company, specific to the description.
The transferable operating rule, from the host: when an agent does something you dislike, ask why at the root, then write a one-line sentence into the skill it loads. Repeat. Bots compound. If you only fix the draft and never tell the agent, you re-teach it forever.
A related knowledge-work habit: a daily bot that reads Slack, Notion, GitHub PRs, and public posts, then writes bullets that roll up by week. You get a log of what you actually did, which is the same raw material the cover-letter writer needs and the same raw material a standup needs.
When the asset is creative, make code the source of truth
What you are trying to accomplish: on-brand motion without living in After Effects. The livestream team’s go-to-market experiment was ads for a game, using Remotion so agents write animation as code. The host’s reason is practical: agents are strong at code and weak at mysterious GUI tools. If Lauren changes the product assets, you pull the new source into the ad instead of exporting a stale screenshot.
Prepare a design system or landing hero, a scratch directory of assets, and an orchestrator that already knows Dr. Eggbot. Then:
- Ask the chief of staff to have Dr. Eggbot onboard a Remotion ads bot.
- Install the Remotion plugin and skills so the agent is not guessing the tool.
- Start with a 1:1 square for X, then generate portrait and landscape from the same composition.
- Prioritize the landing hero; keep motion light until the layout is right.
- Use a tight visual loop (they use Cursor design mode) to strip extra shadows, highlights, and clutter.
- Treat the first square as a template, not a finished media buy.
They describe this as a minutes-long loop once the repo exists, not a production-ready ad account. Play to the agent: text pipelines for letters, code pipelines for motion. If you can put a programming interface on a tool, the bot goes further and burns fewer tokens doing it.
Prompts, bot instructions, and workflows
None of these were pasted from a system-prompt screen. They are reconstructed from how the speakers described the bots. Use them as starting instructions, then add the one-line skills you learn in production.
Family or team glue bot. Reconstructed from the speaker’s description.
Read every incoming email. If it is family-related (kids, school, sports, after-school activities), summarize it in one to two sentences.
Do not only tell me what I should know. Recommend when I must take action. Example: I need to tell the coach whether I can show up to this event.
You are connected to the shared family calendar. Put events on that calendar automatically. Notify us if there are conflicts and help resolve them.
Remember preferences as you see them, including that I prefer not to take meetings in the morning. Do not make me re-explain where to look after email and calendar are connected.Listing and negotiation bot. Reconstructed from the speaker’s description.
Here are photos of items I want to sell. List them.
Research each item. Gather comparable sales. Write a complete listing.
Post to Facebook Marketplace and eBay after I log in. For items that should not be shipped (for example a heavy mountain bike), list for local in-person pickup only. Ship the rest.
Watch incoming questions. Answer them when you can. If you cannot answer, or if a buyer offers a lower price, ask me how I want to handle it. Follow up indefinitely until the item sells or I stop you.Utility-plan audit. Reconstructed from the speaker’s description.
I am attaching a screenshot of a utility email. Figure out whether I should switch plans and which plan I should use.
Look at my previous 12 months of electricity usage and bills. I have an electric car. If a plan is tailored for off-peak charging, include it in the comparison.
Ask me to log into the utility account if you need to. Show the math. Double-check the math before I switch.Orchestrator to bot factory. Reconstructed from the speaker’s description.
Steve: I want to understand my energy bill. Talk to Dr. Eggbot and create a new bot focused on auditing that. Have it report back in this thread.
I will interact with you, not with every sub-agent. Summarize results here.Cover letter critic. Reconstructed from the speaker’s description.
You are a cover letter critic. Reject writing that does not sound like me, including short motivational filler and vague lines such as “I want to work on agents” with no specifics.
Use my resume, experiences.md, and my pre-2022 writing. Prefer compound-complex sentences, some gerunds, and enthusiastic but specific language.
Check four paragraphs:
1. Opening: names the role, why I want it, and what I want to work on.
2. Experience one: uses a real story from experiences.md and connects it to the job.
3. Experience two: same test, impact included.
4. Close: briefly restates those two experiences, then states the impact I want to make at this company, specific to the job description.
If a sentence could apply to any company, send it back.Remotion ads kickoff. Reconstructed from the speaker’s description.
Use Remotion for X ads with one-to-one aspect ratios and the current scratch assets. Use light motion across platforms, prioritizing the hero on the landing page.
Install Remotion skills if a plugin exists. Build from code in the repo so asset changes in the product can flow into the ad.Measurement and business impact
The speakers did not give a dashboard. They used checkpoints you can copy.
Time and attention: the family bot is described as a large, unquantified save. The PG&E switch is described as about 60 seconds of mental attention and $1,000 a year, after two math checks. Unused-subscription cancels are framed as recovering money you would not have chased at $9.99 a month. Sponsor proposals arrive assembled; the human tax is correction, not construction. Workplace retrieval is measured as nearly zero “who owns this?” pings.
Revenue and throughput: Marketplace listings produced sold items (PlayStation, MacBook) with the bike still open. Video editing moved from serial work to parallel edits in a matter of weeks. A separate customer clip on the same stream (Nokia, using Cursor rather than Grokbot) claimed initial analysis of a 50-million-line-plus monolith in a couple of weeks by two people, versus months and a dozen experts with bespoke tooling. That is a speaker claim about a different product; do not treat it as a Grokbot result.
Quality: “right, with notes” on proposals; “sounds like me” on letters; comps on listings; double-checked arithmetic on money. If the bot cannot answer a buyer or a critic flags slop, that is a pass, not a failure, as long as it escalates.
Pitfalls and guardrails
- Do not outsource idea selection. Creative judgment is the part Berman wants humans to keep, and he thinks it gets more valuable as the back office disappears.
- Do not accept summaries without actions. You will still be the glue.
- Do not run customer-facing or money-facing output unsupervised. Listings, proposals, plan switches, and cover letters all had a human on facts, price, or voice. He double-checked utility math on purpose.
- Do not fake credentials or inventory. The intern’s interview warning applies to sales decks too: the first live conversation will expose it.
- Do not ship the mountain bike. Decide channel rules up front (meet in person vs ship) so the bot does not optimize for a listing you will not fulfill.
- Do not build a connector when computer use will do. Also do not dump passwords into chat. Vault, small blast radius.
- Do not collect ten specialists and live in all of them. Orchestrate, or you will recreate middle management in your pocket.
- Do not only fix bad output. Write the skill. Brand voice dies in unlogged corrections.
- Expect cognitive overheating. More completed work is not the same as a calmer day. Step away; cut threads that should never have been agents.
7-day implementation plan
Day 1. Pick one glue problem that already arrives as a three-page email or a lost calendar event. Connect email and one calendar to Grokbot. Write three preferences (working hours, what counts as actionable, who shares the calendar). Do not create a bot army.
Days 2–3. Ship a single specialist: family/team triage, Marketplace-style listings, or meeting-to-proposal with Fathom. Require one-to-two-sentence summaries and an explicit next action. Put events or tasks in the system of record automatically. Share Telegram-style access if a partner has to live in the same calendar.
Days 4–5. Add the human checkpoints the speakers actually used: site logins, price escalations, math double-checks, a critic on voice. Persist one failure as a one-line skill. If money is on the table, run one audit (plan, subscription, or vendor) end to end and confirm the numbers yourself.
Days 6–7. Stand up a chief of staff thread and, if you are spawning more than two specialists, a Dr. Eggbot-style factory. Route reports back to one place. Optionally add one pipeline (outbound finder plus writer plus critic, or Remotion ads from repo assets). Spend a day using only the orchestrator. If you cannot, your bots are too noisy.
Close the loop on one workflow
The business outcome is not “more AI.” It is getting the $1,000 problems and the $10,000 problems done without donating your attention to three-page emails, and then spending the leftover hours on work that still requires taste. Connect one inbox. Make one bot recommend actions. Keep a human on voice and price.
Watch the video for the walkthrough, then start with a single Grokbot on the mess you already hate.
FAQ
Do I need a Mac Mini or a VPS to run Grokbot AI agents?
Not in the setup the speakers describe. Earlier personal assistants pushed people toward local boxes, browser control, and a lot of maintenance. Grokbot is framed as a cloud agent that already has a computer, with no setup the broad audience wants to think about. You still log into the sites it must use. You should not be running a homelab just to summarize mail.
Grokbot vs older personal AI assistants: what actually changed?
The promise is similar: connect sources, read messages, draft mail, do the chores. The complaint about the first wave was security, maintenance, and glue (scripts, service accounts, plain-text secrets). Here the pitch is that those difficulties are gone, connections persist so you are not re-explaining where to look, and passwords belong in a vault the agent cannot see. Treat that as the speakers’ comparison, not a lab test.
Should I use AI to decide what content to make?
Berman says no. Invest time in the ideas and the story. Use Grokbot on editing logistics, sponsor ops, notes, follow-ups, and proposals. If your bottleneck is taste, a bot will not fix it. If your bottleneck is three-page emails, it will.
Do I need custom connectors for email, alumni directories, or job boards?
Often no. The alumni finder and recruiter finder were computer use against a website, LinkedIn, and the Apollo extension, with results copied to a CSV. Utility access was “log into PG&E.” Alumni mail was planned through Outlook MCP, not a homegrown API. Build a connector when the bot cannot see the system of record. Start by giving it the page.
Do I need a chief of staff bot on day one?
No. Build one specialist that already recommends actions. Add an orchestrator when kicking off agents creates its own context-switching problem. Lauren’s pattern is Steve plus Dr. Eggbot: you talk to one bot, a factory stands up the rest, they report back. Berman still peeks into topic threads; his aim is to push more summaries into the morning thread.
Is it safe to connect Grokbot to email, calendar, and logins?
The speakers contrast this with Microsoft Recall (screenshots of everything) and with demos that printed passwords. They want a secure connection to the system of record, vaulted credentials, and a smaller blast radius. That is still access to real mail and money. Share bots only with people who already share the calendar, and keep a human on anything that spends, publishes, or speaks as you.
Do I need Cursor if I am already using Grokbot?
Only if the work is a codebase, a PDF pipeline, or motion defined as code. Grokbot is the glue agent for mail, calendar, browser tasks, and orchestration. Cursor shows up here for cover-letter repos, Remotion ads, HTML-to-React, and a Nokia clip about inspecting a huge monolith. Many operator workflows in this playbook never touch an IDE.
What should I do in the first 24 hours?
Connect email and one calendar. Pick the recurring message you dread. Instruct the bot to summarize in one to two sentences and to tell you the action. Do not generate ads, spawn a factory, or automate cover letters until that loop is boring and correct.