10 Practical Uses of AI: Bill Audits, Doctor Prep, Scheduled Agents
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I walk through ten uses of AI that I found from real people, and how to do each one this week. It's for anyone who already has Claude or ChatGPT and wants to put it to work on bills, doctor visits, rent, apartment hunting, and daily chores. One rule covers all ten: AI reads and drafts, and you check its work before anything gets sent.
What I cover in this video
- 0:46 Auditing a $195,000 hospital bill line by line
- 2:19 A five minute subscription sweep with Claude Code
- 3:04 Prepping for a doctor visit
- 4:35 Two tenant letters with opposite endings
- 6:07 An apartment hunter that runs twice a day
- 8:24 The scheduled morning brief
- 10:40 A private AI box at home and DIY AI glasses
- 12:59 Bonus: Claude Code reads a 1994 Jeep
Audit a bill and sweep your subscriptions
A man on Threads says his brother-in-law died of a heart attack two months after his insurance lapsed, and the hospital bill came to $195,000. He fought for the itemized bill with the billing codes, called CPT codes, and handed it to Claude on a $20 plan.
Medicare has a rule book for which codes can be billed together. According to the reports, the hospital charged for a main procedure, then again for the smaller procedures already inside it. He wrote a letter with Claude threatening legal action and bad press. The hospital came back at $37,000, and he says they settled around $33,000.
A reporter who covers medical bills says he hasn't verified any of it, and a lot of people think the lawyer threat did as much as the AI. To try it on your own bill:
- Call billing and ask for the itemized bill with CPT codes.
- Take your name and account numbers off.
- Ask Claude to flag duplicates and bundled codes, with the rule for each one.
- Check every flag yourself before any letter goes out.
The subscription sweep takes about five minutes. I dropped a fake bank export of about 300 transactions into a folder, opened Claude Code there, and asked it to find every recurring charge, group it by merchant, show the monthly and yearly cost, and flag anything that went up. It flagged the forgotten trial and the price bump I had buried in there, then wrote a cancel script and a lower rate phone script for the top five. You still make the calls.
Doctor visit prep and tenant letters
Lauren Bannon told Fox News that doctors said she had rheumatoid arthritis even though she tested negative for it. She asked ChatGPT what can look like rheumatoid arthritis. It said she might have Hashimoto's disease and to ask for a thyroid antibody test called TPO. Her doctor didn't want to, she insisted, and the test confirmed Hashimoto's. A follow-up ultrasound then found two small lumps, and those were thyroid cancer. ChatGPT gave her a better question, and the test and the scan did the rest.
To prep for a 15 minute appointment, write a dated timeline of your symptoms, your meds, and any lab results. Ask for a five line summary, what else can look like this, which tests would rule those in or out, and five questions to ask. Tell it not to diagnose you.
In 2023, a New York renter whose washers and dryers had been broken for over two years had ChatGPT write a letter citing the state's rent stabilization rules and retaliation. She says the machines got fixed that month. Compare that with a man who just sued OpenAI, alleging ChatGPT told him he could legally stop paying rent and stay put. He says he was evicted in June. Those are his claims, and a court hasn't ruled on them. The difference is who checked the law. Take photos, write down the dates you reported the problem, and ask for a firm, polite repair letter that doesn't threaten anything you can't legally do. Then check your local rules or call a free tenant hotline before you send it.
Scheduled agents that run while you sleep
A user on the Claude subreddit set up an apartment hunter with /schedule. Twice a day on Anthropic's cloud, it checks the main listing sites, filters by area and commute, tags every listing high, medium, or low, and emails him a digest with an outreach draft for each one. Never let it sign or pay.
He also keeps a folder for each part of life, each with a CLAUDE.md note Claude reads first, so it stops asking who everyone is.
For a morning brief, use Claude on a paid plan, connect Gmail and Google Calendar, type this, and hit schedule:
Every weekday at 6:45, send me a morning brief with today's calendar, emails that need a reply, and the weather.
In Claude Code it's /schedule, and it can't run more often than once an hour. If you'd rather go open source, OpenClaw has its own scheduler that posts into a chat app. These scheduled AI agents use your plan's usage, so they aren't free.
Claude in Chrome can record a chore once, like downloading this month's utility bill, and rerun it monthly. Anthropic says itself it's still risky, so I wouldn't let it touch payments on its own.
Local AI hardware and a 1994 Jeep
Ollama runs open models on your own machine. This command downloads a 2 gig model, and after that it doesn't need the internet:
ollama run llama3.2
I show it answering on my Mac with the Wi-Fi off. That box can be the brain for OpenGlass, an open-source project that turns any glasses into smart glasses with less than $25 of parts. The repo isn't maintained anymore, so treat it as a hacker project.
The bonus is a 1994 Jeep Cherokee from before the standard diagnostic port. The owner asked Claude Code how to read its computer, and it walked him through wiring in a $3 Arduino. I wouldn't copy the part where he changed the engine software, because that can break emissions law.
Key takeaways
- AI reads the boring stuff you'd skip, like the bill, the bank file, and the listings.
- It drafts the hard message: the letter, the phone script, the questions for your doctor.
- Anything legal, medical, or financial, check it before you send it.
- If you only try one tonight, I'd do the bill audit.
FAQ
Can AI help lower a hospital bill?
One family says a $195,000 hospital bill came down to about $33,000 after Claude read the itemized billing codes and flagged charges, though a reporter who covers medical bills hasn't verified it. Ask billing for the itemized bill with CPT codes, take your name and account numbers off, have Claude flag duplicates and bundled codes with the rule for each one, and check every flag before any letter goes out.
How do I find forgotten subscriptions with Claude Code?
Drop a bank export into a folder, open Claude Code there, and ask it to find every recurring charge, group it by merchant, show the monthly and yearly cost, and flag anything that went up. It can also write a cancel script and a lower rate phone script, but you still make the calls.
How do I set up a scheduled morning brief with Claude?
Use Claude on a paid plan, connect Gmail and Google Calendar, type a prompt asking for a weekday 6:45 brief with your calendar, emails that need a reply, and the weather, then hit schedule. In Claude Code the command is /schedule and it can't run more often than once an hour. These runs use your plan's usage.
Should I use ChatGPT or Claude for medical questions?
Use it to prepare for the appointment, not to diagnose. Give it a dated timeline of symptoms, meds, and lab results, ask what else can look like this, which tests would rule those in or out, and five questions to ask, then let your doctor decide the test.
Chapters
Transcript
Read the full transcript
0:00A hospital build a grieving family $195,000. And after Claude read the codes, the family says they paid 33. A woman told it was reflux asked chat GPT one question, pushed for a test, and that led to the scan that caught her cancer. And a $19 board can power DIY AI glasses. But the AI isn't the hero. The person who checked its work is. In this video, I'm covering 10 and how to do each one this week, plus a bonus from today where Claude gets inside a 32year-old Jeep. I run my business on AI agents. Let's jump right into it. Number one, audit any bill line by line. A man on threads says his brother-in-law died
0:46of a heart attack after about 4 hours in a hospital. His insurance had lapsed 2 months before. Then the bill came. $195,000. So, he fought for the itemized bill with the billing codes called CPT codes and handed it to Claude on a $20 plan. This is the bit that's easy to miss. Medicare has an actual rule book for which codes can be build together. So, think of one main procedure code as a box that already has the small procedures inside it. According to the reports, the hospital charged for the box, then for everything inside it. Again, Claude also flagged inpatientonly codes on a man who was never admitted and a ventilator build on the same day as critical care. So, he wrote a letter with Claude threatening legal action and bad press.
1:32The hospital came back at 37,000 and he says they settled around 33. That's an 83% cut. Now, a reporter who covers medical bills said he hasn't verified any of it himself. And a lot of people think the lawyer threat did as much as the AI. Fair. But fewer than one in four people who pay bills use AI for it at all. So call billing and ask for the itemized bill with CPT codes. Take your name and account numbers off. Ask Claude to flag duplicates and bundled codes with the rule for each one. Then you check every flag before any letter goes out. AI finds it. You check it. So that one took a fight. Hack number two takes about 5 minutes. And I'll just show you on fake data. I dropped a fake bank export about 300 transactions into a
2:19folder. Open Claude code there and ask it to find every recurring charge. Group it by merchant. Show the monthly and yearly cost and flag anything that went up. Same merchant, same gap between charges. That's a subscription. I buried a forgotten trial and a price bump in there. And you can see both get flagged. Then it writes a cancel script and a lower my rate phone script for the top five. You still make the calls. If you're in the US on Pro, Max, or Ultra, Claude's personal finance tab can do this readonly through Plaid. I'd rather use the CSV. Nothing gets linked. Still juggling Claude code, codecs, and cursor in five terminal windows. Beckin puts every coding agent on one canvas, side by side. Press a key and just talk to them. Every idea gets its own workspace
3:04with memory that carries over. Becken. Hey, Beckin. So, bills are one thing. Health is where it gets serious. And one in five US adults already use chat bots for medical advice. Number three, the doctor visit prep. Lauren Bannon told Fox News her story. She was 40. Doctors told her she had rheumatoid arthritis even though she tested negative for it. Then stomach pain and a 14-lb weight loss got put down to acid reflux. So what did she do? She asked Chat GPT what can look like rheumatoid arthritis. It said she might have Hashimoto's disease and to ask her doctor for a thyroid antibbody test called TPO. Her doctor didn't want to. She insisted. The test confirmed
3:49Hashimoto's. Then a follow-up ultrasound found two small lumps, and those were thyroid cancer. So, the chatbot didn't find the cancer. It gave her a better question. That question got her a test. The test got her a scan, and the scan found it. Her advice, if it gives you something to look into, ask your doctors to test you. A doctor in that same story said AI is not a replacement for human medical expertise. And most people agree. Menllo Ventures found 63% want a human for a diagnosis. Only 19% pick AI. So, here's the prep. Write a dated timeline of your symptoms, your meds, and any lab results. Ask it to get you ready for a 15-minute appointment. A five-line summary. What else can look
4:35like this? Which tests would rule those in or out? And five questions to ask. And tell it, don't diagnose me. AI helps you ask better questions. Your doctor decides the test. Hack number four is the tenant letter, and it comes with a warning label. Back in 2023, a renter in New York had washers and dryers that were broken for over 2 years. days after she asked for lower rent, her landlord raised it from 1389 to 1395. So, she had chat GPT write a letter citing the state's rent stabilization rules and retaliation. She says the machines got fixed that month. Now, put that next to this one. A 60-year-old Danish man named Klaus Elmfeld just sued OpenAI in San Francisco. He alleges Chat GPT told him
5:22he could legally stop paying rent and stay put. Wrote the letter for him and told him, quote, "You will not end up on the street. Legally impossible." He says he was evicted in June. A court hasn't ruled on any of that. Those are his claims. So, this is where I'd slow down. Same tool, same kind of letter. One path asks for a repair and documents everything. The other path stops paying rent. And in New York, the attorney general's own tenant guide says withholding rent can get you sued for non-payment. So, the difference is who checked the law. Take photos, write down the dates you reported it, and ask for a firm, polite repair letter that doesn't threaten anything you can't legally do. Then, check your local rules, or call a free tenant hotline before you send it.
6:07AI drafts the letter, you check the law, and maybe you're done with that landlord and you need a new place. Number five, the apartment hunter that runs twice a day. A user on the Claude subreddit set one up with slash schedule. So twice a day on Anthropics Cloud, it checks the main listing sites, filters by area and commute, tags every listing high, medium, or low, and emails him a digest with an outreach draft for each one. And my favorite line from his post, "When I sign a lease, I disable the trigger." So is this new? Not really. In 2023, a coder in Berlin named Daniel Dipold was 4 months into his search. Chad GPT helped him write a scraper. He built a list of over a 100 property managers, emailed them, and got 35 replies in one day. About 2 weeks in, he had two
6:53accepted applications. Business Insider checked his screenshots. He also said a lot of its ideas were bad, like a chatbot talking to landlords for him or signing a lease without seeing the place. So, you keep the human parts human. So, start simple. Ask for 20 unusual ways to find a place in your city. Then throw out the bad ones. When that works, put it on a schedule and swap the listings for used cars or contractors. It's the same engine. Just never let it sign or pay. So, that same Reddit user had a bigger setup and it's my favorite one on this list. Hack number six, your life in folders. He keeps one for each part of life. Health, finance, travel, home, and personal admin. Each one has its own claude.md,
7:38which is just a note Claude reads before you type anything. So, when he opens the health folder and asks for a care summary for a specialist, he says it already knows the family context. In his words, that's what made it useful day-to-day rather than a clever toy. That thread got 730 up votes. And another person in it says they use Claude 65% for life admin, 35 for work. That one's through Claude projects, which is the simpler way to do this. Why folders? Because Claude stops asking who everyone is. Here's my version on fake files. Five folders in health. A claude.md that says who's in the family, where the files live, and one rule, never send anything without asking me. So, you open claude code in that folder, ask for a one-page summary for a
8:24cardiology appointment, and it pulls from the notes already sitting there. No explaining your whole life again. So, folders are step one. Number seven is where it starts working without you, the morning brief. One parent in that thread gets a brief every morning on type 1 diabetes news that matters for their son. Another person's brief reads out their calendars and now the day's cleaning jobs, too. And this is basically how my channel runs. Claude Code jobs on a schedule. This video started as one of them. A job kicked off at 4:00 a.m. Researched it, wrote it, and edited it while I was asleep, and the voice you're hearing is my clone. So, yeah, I'm a little biased. So, how does that work? A clock hits 6:45. A session starts on Anthropics Cloud, even with your laptop closed. It reaches into
9:09your calendar, your inbox, and the weather, folds it into one message, and your phone lights up. So, to set it up, use Claude on a paid plan. Connect Gmail and Google Calendar and type every weekday at 6:45, send me a morning brief with today's calendar, emails that need a reply, and the weather. Then, hit schedule. In Claude Code, it's SLC schedule and it can't run more often than once an hour. If you'd rather go open source, OpenClaw has its own scheduler that posts into a chat app. One heads up, these runs use your plan's usage. They're not free. That brief reads your stuff. Number eight actually clicks through it for you. Record a chore once, then put it on a schedule. So in Claude and Chrome, you click the record icon and do the chore one time.
9:55Say, log into your utility site, download this month's bill, and rename it. Stop and save it as a shortcut. Type a slash to run it again. Then click the clock icon, set it to monthly, and that bill shows up every month, ready for the finance folder from hack 6. So, in practice, it goes like this. Hit record, grab the bill, rename it, stop, save it as a shortcut, and set the clock to monthly. A few limits, though. It's paid plans only, desktop Chrome only, and recording works in the classic side panel, not when the panel runs as a co-work session. Chrome has to be open for the schedule to run. And Anthropic says itself it's still risky. So, I still wouldn't let it touch payments on its own. And that's the last hack that
10:40lives only in software. Now, let's build some hardware. About a third of Americans have a smart speaker. So, what if the AI brain lived in your house, too? Number nine, a private AI box at home. So, Olama runs open models on your own machine. You type Olama run Llama 3.2, it downloads a 2 gig model, and after that, it doesn't need the internet. Here's my Mac with the Wi-Fi turned off, and it still answers. So, how slow is a Pi? Jeff Gearling benchmarked this exact model. On a Raspberry Pi 5, about five tokens a second. Slow, but it works. on an M4 Mac Mini about 41. So roughly eight times faster. And I almost missed this. The 16 gig Pi 5 just went up to $385.
11:28It launched at $120. Raspberry Pi blames the price of memory. So the Pi isn't the cheap way to do this anymore. Start with the Mac you already own. If you want it to talk, Home Assistant sells a $69 open speaker that can use your local model as its brain. They call that part experimental. So, go in knowing that. And that box is about to matter because the last one, hack number 10, is DIY AI glasses. Open Glass is an open-source project that turns any glasses into smart glasses with less than $25 of parts. An 1899 Seed Shia board with a camera and a mic, a small battery, and a 3D printed mount. It can send what it sees to a cloud model or to your own box from Hack 9 running a small vision
12:13model. So, Hack 9 is the brain for Hack 10. Want a pin instead? A maker called Techisms built an AI pin on the same board with a tiny speaker that talks back, though that one runs through a cloud service. And a project called PMSG has open modular glasses on that board with a starter kit coming for around 110. So, build or buy. Brilliant Labs Halo is $3.99 and shipping now. They say the AI runs on the glasses chip and your phone, though there's also a cloud agent and the free features have daily caps. So, DIY is under 25 bucks and a weekend. The battery is short and the open glass repo isn't maintained anymore. It's a hacker project. Very cool hacker project though. And one bonus because it showed
12:59up today. A guy on X has a 1994 Jeep Cherokee that was failing the California smog test, and he says every retest cost him about $100. The car is from before the standard diagnostic port, so no scanner plugs into it. So, he asked Claude Code how to read its computer. Claude worked out the old protocol the car speaks and walked him through wiring a $3 Arduino into it. Then, it logged his drives and went through a copy of the car's own program. It found that the California version trims fuel at the exact load the smog test uses, and he thinks that's why it kept failing. At one point, he let Claude switch on the fuel pump, and he says that's when he started to freak out. And he went further than that. He says Codex got its own code running on the car's computer,
13:44and the Jeep passed with no mechanic and no dealer tool. Now, I wouldn't copy that last part. Changing engine software to pass an emissions test can be against the law, so check your state's rules first. AI reads the car, you check the law, and the reading is the real hack. If something you own has a computer in it, an agent can probably help you read it. So, look at all 10 together. AI reads the boring stuff you'd skip, like the bill, the bank file, and the listings. It drafts the hard message, like the letter, the phone script, and the questions for your doctor. And it runs on a schedule while you sleep. One rule covers all of it. Anything legal, medical, or financial, check it before you send it. Remember those two tenant letters. Same tool, opposite endings.
14:31So, if you only try one tonight, I'd do the bill audit because it's the one most likely to pay for itself in an afternoon. Okay, your turn. Which of these 10 are you setting up first? And what's the one bill you'd want audited? Drop it in the comments. And if this saved you some time, tap the like button so it reaches more people. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Code Sub Agents for Beginners: Building My First One
I asked Claude Code to look at my codebase and suggest three sub agents for easyflip.ai, then built the first one, an AI accuracy optimization agent. Getting it to read brand and size from photos took a lot of pasting browser console logs back into Claude Code.
Claude Code Examples: 20 Real Projects People Are Building
I count down 20 real things people built with Claude and Claude Code: games, hardware, apps, and agents that run a business. Number one is Dave Jones of EEVblog, who says Claude Code rebuilt his 121GW multimeter firmware in plain C from a single prompt in under 24 hours.
Claude Code Video Editing: My Full Automated YouTube Pipeline
I make my YouTube videos and Shorts with Claude Code. 15 skills handle research, scripting, B-roll from X, coded animations, my ElevenLabs voice clone and the final cut. Scheduled jobs build it all overnight while I sleep. My repos video hit 82,000+ views without me filming anything.
