12 AI Assistant Use Cases That Feel Illegal

summarized

TLDR

Personal AI assistants can now negotiate bills, cancel subscriptions, and even control a Tesla when given API access and a structured prompt. Matthew Berman reports that his Grockbot setup found over $1,000 a month in unwanted subscriptions, recovered hundreds of dollars in unclaimed property, and switched him to a cheaper electric plan. The real takeaway is that these workflows are mostly reusable prompt templates that hand the assistant real account access while keeping humans in the approval loop.

Key points

The Tesla bot controls climate, navigation, and garage door through Tesla's official fleet API.

Meeting summary bot pulls action items from Fathom transcripts into Todoist.

The family bot puts school events on the shared calendar with spouse invites.

The computer cleanup bot recommends weekly deletions without deleting anything.

All prompts and Grockbot templates are shared in one Box document.

Tools mentioned

Techniques

  • Structured agent prompts divided into setup, job, and rules sections
  • External API integrations such as Tesla fleet API, Fathom API, and Todoist API
  • Risk-tiered email triage with autoarchive and approval-based drafting
  • Human approval gates before high-stakes actions
  • Meeting transcript action-item extraction into a to-do list
  • Utility usage analysis for rate plan comparison
Transcript (captions)

0:00 Personal AI assistants like Grockbot, Muse, and Dot have taken AI mainstream. They are helping millions of people get more done, be more productive, and also do less of the tedious, boring stuff

0:13 that life just requires. I have bots saving me money, planning my calendar, booking me appointments, and even controlling my Tesla. All right, so everything I'm going to show you is in

0:24 Grockbot. That is the personal AI assistant that I am using right now. I also use Muse. I also use Dot. But my primary is Grockbot. But all of these use cases can apply to whichever

0:37 personal assistant you're using. And so the first use case is the money saver. Here it is. Look through all of my emails for all of my subscriptions and then go negotiate with any and all of

0:48 them that you can to try to get the price down. Now, you can also tell it to look for any subscriptions that look like you're not using them and just go ahead and cancel them for you. It found

1:00 over $1,000 a month in fixed seats across my three Gmail accounts. Everything from my insurance to type form, Chad GPT, Anthropic. Obviously, a lot of these I need, but you'd be

1:13 surprised if you just tell one of these subscriptions that you're going to leave. They generally offer you a discount to stay. And so now the bot can do it for you. So here's one. It coded

1:24 up an email to Beehive, which is what we use for our newsletter. And I'm saying, "We're on Beehive Max at $459 a month for Forward Future and has become one of our biggest software costs. We're happy,

1:36 but we need to bring down the bill." And it drafts up the email. All I have to do is click send. All right. And as part of the moneysaver bot, I also had it go to the California unclaimed property

1:49 website and find anything that is under my name that is unclaimed property. And it actually found hundreds of dollars in money that was just sitting there. I mean, look at this. We have returned

2:00 more than $8 billion. And this is obviously just for California, but many states have websites just like this. And you just tell your bot to go do it. Look for the unclaimed property and then

2:11 submit the forms. And for each of these use cases, I'm going to provide you with a Grockbot template that you can actually just click into and you can use my exact template for yourself. And then

2:23 I'm also going to give you a prompt that you can use to set it up in case you're using something else than Grockbot. And by the way, I'm going to put all of these prompts into a document hosted on

2:32 Box. I'll drop the Box link down below. Shout out to Box for helping out with that. Okay. Next, I actually control my car through my Grockbot. So, I just got a Tesla Model Y. I told my Grockbot to

2:44 connect to my Tesla. It downloaded all of the API documentation. I got an API key. I gave it to it. And I can tell it things like, "Hey, get the car ready for me." And it'll make sure that the

2:55 climate is on, my seats are heated, for example, and it even sets my destination of where I want to go. It can even open my garage door for me. And soon hopefully it'll even be able to pull the

3:06 car out of the garage just having it waiting and ready for me. For the Tesla bot, here's the prompt. You are my Tesla assistant. First, figure out what you need from my Tesla account through

3:16 Tesla's official fleet API and my connected calendar. So, which Tesla's on the account, usual climate and seat heat settings, frequent destinations like work or school drop off, etc. Only ask

3:29 me if something truly cannot be inferred. setup that only I can do like signing into my Tesla account, two factor codes, and pairing the virtual key on my phone. I will handle myself.

3:39 Job: Control my Tesla through the official fleet API, climate and defrost, seat and steering wheel heat, lock and unlock frunk and trunk, charge port and charging limits, and sending navigation

3:52 destinations. When I say get the car ready, precondition the cabin and seats, open the garage if that's set up, and send navigation to where I'm going next on my calendar. And again, I'll drop the

4:03 Grockbot template down below. All right, next is my computer cleanup bot. And I am frequently running out of room on my hard drive because I just store a lot of videos and I do a lot of screen

4:14 recordings. And this is a bot that runs every Monday and looks for cached files, stale files, anything to save me room on my hard drive and will automatically put it in the trash for me. Now, it doesn't

4:27 delete it, but it will find everything and list it all. So, here are some examples. So, here, this actually just ran today. About 50 GB on the Mac Studio looks safe to clear. You've got 371 GB

4:39 free, so no rush. Nothing was deleted. All right. So, 13.7 GB of camera clips. These are old. We have codecs, 44 GB of dated output. We have OBS, 11 GB of recordings that it says I can drop the

4:56 MKV versions of. Great. And yeah, it just found a bunch. Sometimes it's just cache, but it will automatically find everything and just clean it up for me. And here's the prompt for the computer

5:07 cleanup bot. You are my computer cleanup assistant. First, figure out what you need from this machine. For example, free disk space, large folders and downloads and desktop, cache locations

5:18 for my OS and installers or disk images that look stale. Only ask me if something truly cannot be inferred. Job: Scan weekly for stale, duplicate, and oversized files such as old downloads,

5:31 caches, and installers. Suggest what to delete with estimated space freed by risk level. That's really important. So, if it's a high-risk deletion, it will tell you. Prefer obvious caches and

5:43 leftovers before anything that might be personal documents. Rules: Never delete empty trash or move files without my Okay. Show me the full path and why you flagged each item. Skip system

5:53 directories you are not sure about. This is what recording videos used to be like for me. I have struggled with insomnia and poor sleep for years and that's until I became a customer of

6:13 eight and I purchased the eightle pod about 2 years ago. I first heard about it from Brian Johnson and Andrew Huberman. So the pod's smart mattress cover cools my bed before I get in it.

6:24 And then I recently got the pillow cover which also keeps my pillow nice and cool before bed. And then rather than being jolted awake by an alarm in the morning, my bed gently warms up to the point

6:36 where my body naturally just wakes up. And then when I snore, the base adjusts its height just slightly and I stop snoring. Autopilot uses the bed's AI to adjust the temperature of the bed

6:49 throughout the night, optimizing for deep sleep. And now they've released their newest generation Pod 6, which is smaller, more powerful, more accurate, all while dropping their starting price

7:01 to the lowest ever yet. Go to eightsleep.com/matthew Berman. Get $200 off with code Matthew Berman on the brand new Pod6. All right, next is my daily briefbot,

7:14 and I have separate ones for work and for family. This is a daily brief that gets created for me every morning. And it will look at my calendar, my emails, my to-do list, everything, and then put

7:24 together a daily brief. And that includes things like the things I need to get done, the events for the day, any context that I need, any items that it can get done for me. It's all included.

7:35 And because there's a bunch of personal information in both my family brief and my work brief, I'm actually just going to show you the prompt I use instead. You're my daily brief assistant. First,

7:46 figure out what you need from my connected email and calendar. Only ask me questions if it cannot be inferred. Your job. Every weekday morning, send me a short brief. Lead with anything that

7:56 needs action. Include my urgent emails, then today's meetings with context on each one. Who's attending recent emails with them and notes from past meetings when available? Keep it scannable under

8:07 a minute. Rules: Never send, archive, or delete email on my behalf without my okay. read only unless I explicitly ask you to act. Stay quiet on days with nothing urgent and no meetings unless I

8:19 ask for an empty check-in. Okay, this next one is probably one of the biggest times savers of all the use cases I have. I have what's called a family bot and it manages everything to do with my

8:32 family. And so if you have kids, you're probably very familiar with this. You get emails all the time from their school, from their sports, and they're not short emails, and they're not

8:41 infrequent. They come all the time, and they're usually paragraphs and paragraphs long, much of which can be ignored. So, I have my family bot looking at every single email that comes

8:51 in, looking for specific ones that are either from my kids' schools or from their sports. I have it summarize all of those emails, and give me the most concise summary possible. I also have a

9:04 Telegram group chat with my family and I have it send those summaries to that group chat. Then it autoarchchives the email for me. It's also directly connected to my family calendar. So if

9:16 it sees any events in those emails, it will automatically put it on my calendar and invite my wife to that event. And Grockbot Muse dot, they are all extremely proactive. So, now that it

9:29 understands all of the different sports and the school activities and my calendar, it has been an incredible timesaver. Here's an example. My son had a soccer game this weekend and the day

9:40 before the soccer game, my family bought messaged me saying, "He has a soccer game at this time at this location. He's the home team, so make sure to wear the red jersey and remember you are on snack

9:53 duty for this weekend. Do you want me to order it from Door Dash for you?" And so it takes all of these different parts of my Grockbot experience or whatever your personal AI experience is and really

10:05 puts it together in such a holistic way. And again, that probably saved me like 30 minutes of time right there. So if you have kids, I highly recommend setting this up. And here's the prompt

10:18 for the family bot. You are my family logistics assistant. Your job is to keep our household on top of everything coming from school activities and other family commitments. First, ask me one

10:28 question at a time who is in my family and their ages. Of course, you can just tell it rather than having it ask you questions. But if you just want to copy paste, this is the way. Which schools,

10:38 teams, clubs, or activities each person is involved in. It can usually infer that by your emails. Which email accounts those messages arrive in, which calendars we share, and who else should

10:47 be involved. Read new emails from schools, teachers, coaches, and activity organizers. Summarize only what matters: dates, deadlines, forms, payments, things to bring, and schedule changes.

10:58 Skip the fluff. Add every event deadline and early dismissal to our shared calendar with the location and details. And invite the other adults who should know. Never create duplicates. Send one

11:09 short daily digest to our chosen channel grouped by family member. The day before any event, send a heads up with everything needed to show up prepared. Where and when, what to wear and bring,

11:21 any assigned duties, carpool, snacks, volunteering, and anything to sign or pay, flag anything urgent right away. And next is my leadbot. So, I get a lot of DMs, whether it's from Facebook or

11:34 LinkedIn or X or Tik Tok, Instagram, and I was not checking them or at least not often enough. And I would get a lot of companies reaching out to want to work with me or even just like cool people

11:46 who wanted to show me awesome projects that they were building. And I just couldn't keep up with all of it. And so I made the leadbot and it goes out and it checks each one of these sources. It

11:58 reads the DMs, will categorize them by whether it's a company wanting to sponsor one of the videos or somebody cool showing me a project. It will automatically tell me about the

12:08 important ones. It'll archive the spam and then allow me to reply directly from the chat with my bot. So, I'm not having to go into each individual platform to write my replies manually. And the cool

12:21 thing is it'll actually draft emails for me to sponsors who look potentially awesome to work with. And again, all I have to do from within Grockbot is just click send. Here's the prompt for the

12:32 leadbot. You are my lead finding assistant. First, figure out what you need for my connected Instagram, LinkedIn, and X accounts. for example, which inboxes to watch, what a serious

12:41 sponsor asks looks like from past conversations, and how often to check during my active hours. Your job, monitor those DMs for sponsorships, brand deal, and partnership leads. Vet

12:51 each one for a real company, a clear ask, and a budget or timing signal. Report only the legit ones with a short summary and suggested next step. Ignore spam, vague pitches, and engagement

13:03 bait. Never reply, accept, or negotiate as me without my okay. draft replies for my approval. Stay quiet when there are no real leads. Today we are launching the Frontier Pass. It is a membersonly

13:16 community that gives you access to exclusive videos and interviews, a private Slack community where you can come chat with me and my team, monthly AMAs, and each month we're giving away

13:27 to one person $200 to help pay for your Max AI subscription. And if you're already a YouTube member, you don't need to pay again. you can go claim access to the Frontier Pass immediately. So, go

13:40 check it out. I hope to see you in Slack. All right. And next, again, maybe one of the more valuable bots that I have. This is the meeting summary bot. Every single meeting that I join has my

13:50 Fathom Notetaker. The Fathom Notetaker is recording the meeting, writing a transcript of everything we talk about, and then saving that transcript. After every meeting, my meeting summary bot

14:03 will actually first of all to know I just had a meeting will look at my calendar and look for those meetings. Then it will ping the Fathom API, grab the transcript, and then write up a

14:13 summary of the meeting that I just had. But it takes it a step further than that. It'll look for action items from the meeting transcript. So, for example, if I say, "Hey, okay, I'll get you that

14:23 proposal by Monday." It will actually know that. it will flag that for me and will even add it to my to-do list. And it even takes it a step further. If it can do something for me, it will

14:36 automatically do it and just say, "Hey, you said you were going to send this person a proposal. Do you want me to draft that proposal and get the email ready to send to them?" And it's all so

14:46 easy and automated. It is awesome. And the cool thing is I know my Fathom is recording my meetings. So during the meeting, I might say something like, "Oh yeah, and remember to send the

14:57 proposal." Just by saying it in the meeting, my meeting summary bot will know to look for that. Here is the meeting summary prompt. You are my meeting summary assistant. Figure out

15:08 what you need for my connected meeting recorder and calendar. So I'm using Fathom, but you can input anything, and it might just ask you what you use to record your meetings. Your job after

15:18 each recorded meeting, pull the transcript and send me a short summary with key decisions, open questions, and action items. Who owns what and any dates mentioned. Keep it tight enough to

15:28 skim before my next meeting. Rules: Never send the summary to other attendees. Post it publicly or create tasks in other tools without my okay unless I already told you to do so. Stay

15:39 quiet when there are no new recordings. And so, here's how it works. I have a meeting. My Fathom agent, which is separate from Grockbot, is in each meeting recording it. It will actually

15:50 write up a transcript right here. My bot will look for that transcript, will ingest it, understand what that meeting was about, will spit out a summary right here, and it'll send the summary to

16:05 Slack. And within the meeting summary bot, it will also, since it has access to my to-do list, which by the way, I use to-do list, it'll look for items in my to-do list that either got completed

16:17 based on the meeting or that it needs to add new to-do items to. And so, it is really neat. It is basically always staying on top of what I need to get done. And so, one thing that I told my

16:28 Grockbot to do generally is to always reference my to-do list. And every time it does anything, whether that's look at a meeting transcript, read my emails, look at my calendar, it is always

16:39 looking and referencing against my to-do list to see if it should mark something off the list or add something to the list. So, it's a very dynamic to-do list that I'm working with in collaboration

16:50 with my AI assistant. So, I use this very simple to-do list app called To-Doist. I'm not paid by them. I just have been using them for a long time and I really like it. And it also has an API

17:01 and a direct integration with Grockbot. So I basically just point my Grockbot at to-d doist and now it knows everything I need to get done. Now Muse has its own built-in to-do list, which is quite

17:14 nice. But my primary is Grockbot, so this is no problem. I just point it at my to-do list. And here's the prompt for the to-do manager. You are my to-do manager. Figure out what you need from

17:25 my connected to-do app. Email and calendar. So, look for title tasks, look for default due date patterns, where they live, and only ask me if something truly cannot be inferred. Job: Manage my

17:40 list directly in my to-do app. Add tasks from my emails, meetings, and messages when something actionable appears. Keep tasks organized with clear titles, projects, and due dates. Mark items done

17:51 when I say I finish them, and periodically scan for completed or stale tasks. Tell me what you changed in a short receipt. Rules: Never delete a task permanently without my okay. Prefer

18:03 completing or rescheduling over deleting. Do not create duplicate tasks for the same ask. Stay quiet when there's nothing to update. And again, a huge time-saving bot for me is my email

18:14 triage bot. I have three separate email addresses. One for personal, one old work email, and then my new forward future email. And I have my bot scanning all three of them. It first looks for

18:29 emails that it can easily archive. Things like Amazon delivery notifications, receipts, and other invoice type items and anything where it doesn't require a

18:41 reply from me. And it will automatically archive those. It'll tell me about it. It'll give me a very short like single sentence summary or receipt of the emails that it archived. And then it

18:53 starts going through the kind of higher risk ones. So anything it thinks that I need to just give a short reply to, it'll go through it one by one with me and say, "Hey, do you want to reply?"

19:03 And it will actually draft a reply. Usually I have to change one or two things and then I just let it send. And then there are the kind of high-risk ones where it requires me actually

19:14 thinking through a question or problem. I have to put some kind of document together or something where I have to send something. And for those, my Grockbot tries to put together whatever

19:25 document or email draft or anything that I need automatically for me. And I'm usually changing a few things and then sending. Here is the prompt for my email triage bot. You are my email assistant.

19:37 First, figure out what you need for my email and calendar. So things like which accounts I use, who is always high priority from recent threads, what looks safe to autoarchchive, such as shipping

19:49 notifications and receipts and my usual work hours. Your job summarize only my important emails, pulling key dates and ass out of long newsletters. Autoarchchive low-risk mail like

20:02 shipping notices, receipts, and obvious promo noise. And tell me what you archived in a short grouped receipt. Then walk me through the rest, one email at a time, with a short summary and a

20:12 suggested reply or action, including by voice while I drive, which I use Grockbot in my car. Of course, you can use Muse or Dot in your car if you have Android Auto or Apple CarPlay. Flag

20:23 contracts and e signature requests immediately and keep them visible until they are signed. Never send, permanently delete, or buy anything without my okay. Draft first and let me approve. Never

20:33 autoarchive contracts, legal mail, or anything that needs my signature. Stay quiet when there's nothing new. Okay, this next one saved me a thousand dollars a year. You should do this

20:45 today. This is my PG&E bot. And PG& is specific, I believe, to California, but it basically applies to whatever your electricity or utility providers are. So, what it did was it actually logged

20:59 into my PG& account, looked at the last 12 months of electricity usage at my house, looked at all of the patterns, and then compared electricity plans, and it actually found one that saved me

21:13 $1,000 per year, and it not only suggested the plan, but when I analyzed the comparison and made the decision that I did want to change, it changed for me. It was so simple. So, here's the

21:25 prompt for that. You are my utility bill assistant. Figure out what you need from my email and any utility account I can sign into from the browser. So, this would work for gas, electric, solar,

21:37 whatever. Review about a year of electricity usage. Compare every available rate plan I am eligible for and tell me which one saves the most and how much per year. Walk me through the

21:49 switch steps only after I choose a plan. Use my utilities official site or app. PG&E style plan comparison is a fine example pattern for other utilities too. Never switch plans, change account

22:02 settings, or share account credentials beyond the login I completed myself. Recommend first switch only with my explicit. Okay. All right. The next use case is setting up a calendar assistant.

22:13 Again, just like I have multiple email addresses, I have multiple calendars. whether that's my work calendar, my personal calendar, my family calendar, and I synced all of them to a single

22:24 bot, so it knows exactly what I'm doing at any time of the day or date, which is really nice because then if I have my email bot responding to somebody who's asking for a meeting with me, I can say,

22:37 "Okay, go ahead and schedule them for next week at whatever time I have available." and it will automatically know the specific times I have available and propose it and negotiate the actual

22:49 schedule with that person. It will also notify me of any conflicts. If I accidentally have my son's soccer game as well as a movie I'm supposed to go see, it will automatically notify me

23:01 about the conflict and ask me if I wanted to reschedule it for me. Super super useful. So, here's the prompt for my calendar agent. You're my calendar assistant. Figure out what you need from

23:13 my connected calendars, which calendar is primary, my time zone, typical working hours, buffers I already leave between meetings, and any recurring personal blocks. Your job is always lead

23:24 with today's agenda when I ask about my schedule. Watch for conflicts and tell me before they bite. Preferring to propose a reschedule over declining when that is possible. Schedule events when I

23:36 ask, checking for duplicates first. Book appointments for me, for example, a dentist, doctor, or haircut by finding open times that fit my hours. Then confirming the time, place, and details

23:47 with me before anything is submitted. Never create, move, decline, or book anything without my okay. Label times in my time zone. Tell me what changed after every edit. All right, so for these next

23:59 two, I'm just going to go through them very quickly because they're essentially kind of the same thing. I have an Uber bot and a Door Dashbot, and it is exactly what it sounds like. I can

24:09 literally tell my bot to order me a car. I need to get to San Francisco by 2 p.m. Or I can say, "Hey, can you reorder the last order I got from Mendescino Farms and have it delivered by 100 p.m. And it

24:24 will go through and do those things for me. It is dead simple. And because I created bots for them, every other agent that I have working also can talk to the Door Dashbot or the Uberbot and schedule

24:38 things for me. So you can start to see how all of these different use cases start to play together. So for example, on the daily briefbot, it can say, "Hey, you have a bunch of meetings, but I see

24:50 you have an hour gap, and usually on Mondays, you like to order food. Do you want me to put in that order so it arrives between your meetings?" And it's just really cool because with all of the

25:00 context, it can be extremely personal and also get real world things done for you. Now, that's most of the use cases that I'm using my personal AI assistant for today. Now, again, I'm going to drop

25:13 a document down below with all of the prompts for each of these that'll be hosted on Box. Shout out to Box. And if you use Grockbot like me, I'll drop all of the Grockbot templates down below so

25:23 you can one-click copy the exact use cases that I'm using. I love talking about practical AI, things that you can actually get done with AI, and that's exactly what I did in my Jev use cases

25:36 video. So, go check that out right here.

Frontier News · by Hyperjump Technology