Check Its Work, Then Let It Run: How to Trust AI and Put It on a ScheduleA real AI class I teach my team on how to use AI in our business: the six mistakes to stop, one ready-to-use job for every team, and how to make it run without you
This is a real class. I teach my team at Inspired Solutions how to use AI in our business. This one was recorded live on Friday, October 16, 2026, and it ran about 37 minutes. My team was on the call with me, and a few of them spoke up, so you will hear a real conversation, not a script.
Like always, I used AI to create the slides about AI, so by definition the class is meta.
Each section below starts with What I said to my team in the class. Then My thoughts gives you the part to take with you. Claude is the AI assistant I use. Everything here works the same way with other AI tools.
Don't ever just assume that Claude or any AI did everything right. I am using AI all day to do a bunch of different things. When I tell it to do something, I ask it what it changed. Show me the difference. Show me this versus that.
I use something called an IDE, an integrated development environment. It is a program that lets me see the old file and the new file side by side. If you are asking AI to do something, ask questions and verify.
Here is one I have caught more than once. It made up somebody's name or email address. I said, "Where did you get that from? That's not their email address." Now I am wasting time sending an email that is going to fail, and I am doing more work, not less.
Sometimes artificial intelligence is not that intelligent. You have to check. And then you have to put some systems in place to make sure that you correct it moving forward.
If you are just asking AI to do something and you are not reading it, shame on you. Read your emails. Read your proposals. Read the documents before you turn them in.
2.
The six mistakes and the one line that fixes each
What I said
Here is what Claude did to me, and the line I now use to stop it. Copy the line that fits your job.
Mistake 1. It made up an email address from a company's naming pattern.
The fix: "Use only email addresses from my mail or contacts. If you can't find one, leave the To line blank and tell me."
Mistake 2. It said an email never came. It was there.
The fix: "If you find nothing, search a second way (name, subject, date range) before you tell me it is not there."
Mistake 3. I asked it to go through my emails and consolidate every job title I had submitted in past proposals into one long list. It said it found all of them and handed me a list that was "ready to go." I opened it and said, wait a minute, there's not a lot here. It had 36. After I pushed back, it added them up and the real number was 119. It is like talking to your kids sometimes.
The fix: "Give me the full list the first time. Tell me what you left out and why."
Mistake 4. It told me something was expired, reading an old file. The system we use to track the real dates had the right ones. I said no, it is not expired. I said no, it is not expired, let's move forward. It kept coming back with, but are you sure?
The fix: "Our tracking system is the source. If a file disagrees with it, the system wins, and tell me the file is out of date."
Mistake 5. It scored every YouTube title 100 out of 100, even nonsense.
The fix: "Test it with junk first. If junk scores well, the tool is broken."
Mistake 6. It wrote a reply that wiped out the email it was answering.
The fix: "Make it a real reply. Keep their message quoted underneath mine."
3.
A real example, and the footer you must remove
What I said
I will give you a real-life example that happened to a friend of mine this week. She is a professional. She graduated from college 40 years ago and has been a senior vice president and an executive vice president at big firms.
She was interviewing for a big role. They asked her to submit her resume, so, like everybody else, she asked AI to update it. She has an associate's degree. She never completed a bachelor's, but she has a bunch of other certifications, so it did not really matter. She is an executive.
But the AI changed the degree on her resume from an associate's to a bachelor's. She did not see it, and she submitted it. In the interviews, where she filled things out by hand, she wrote the associate's degree. She went through eight interviews and negotiated her salary. Finally, they told her she had the job.
Then HR did the research, and her college came back and said she only had an associate's. HR asked her about the bachelor's. She said, "I don't have a bachelor's." They said, "Your resume says that you have a bachelor's." She had to tell them AI must have done that, and apologize.
That cost her four days of agony. She still got the job, thank the Lord. But the point is: don't do that. Make sure you read this stuff.
One more thing. When you copy and paste from AI, the reminder at the bottom ("AI can make mistakes, review your work") can come with it. Remove it. Pay attention to that. Don't leave it at the bottom of something you send to a customer.
4.
Tell it once: put the rules in your project
What I said
If I want to make sure it does not do that again, I tell Cowork or Claude Code to update my CLAUDE.md file, or whatever file I am working with, with these rules. You give it rules so that it becomes a permanent change. Once it is in your CLAUDE.md or AGENTS.md, that's it. It's a permanent rule.
Cowork is where I work with Claude on email, documents, and slides. Put the rules in your project's instructions. Claude Code is the version that works inside a folder of files. Put the rules in the CLAUDE.md file in that folder. Every chat in that project reads them first. You type them once, not every Monday.
The seven rules I put in a project:
Use only email addresses from my mail or contacts. If you can't find one, leave it blank and tell me.
If you find nothing, search a second way before you tell me it is not there.
Give me the full list the first time. Tell me what you left out and why.
One source wins. If a file disagrees with it, the source wins, and tell me the file is out of date.
If every result looks perfect, test it with junk before you trust it.
Replies keep the original message quoted underneath.
Drafts only. Never send. Write Unavailable instead of guessing.
5.
Give it the files, the context, and the connectors
What I said
Say a contract is awarded. I take every file for that contract and put it in one folder: the performance work statement, the statement of work, our proposal, the slides from the kickoff call, and the meeting notes. Then I create a CLAUDE.md file inside that folder, so whenever I say I want to work on this project, Claude knows everything in the folder. If it does not have access to all the files, it is not going to be able to do the job.
You have to give it access to the information. You have to provide the context. You have to make sure you have connectors. And then you have to give it clear instructions on how you want the output to look.
A connector is how AI reaches the data inside a tool you already use. Companies built something called MCP, a model context protocol, so people can connect their own data in a tool to their own AI. We turned on the connectors for our project board, our office suite, and our meeting app, so our team has access.
Markdown is a plain text format that is easy for AI to read, and humans don't read it that well. So for my meeting notes, I make it create two copies. One is a markdown file, which goes in my database for AI to read. One is a Word document with our colors and our logo, so humans can read it. A few minutes after a meeting, my team had notes with our logo that we could send to everyone on the call.
6.
One ready-to-use job for every team
What I said
The concept is what I am trying to teach you: what you can do with AI. Whether you are on that team or not, it does not matter. Each job below has a prompt that tells Claude what to read, who to act as, what to make, what not to do, and what I will check.
Sales: quotes and proposals.
When a request for quote or proposal lands, Claude builds the quote worksheet, the submission checklist, and a first draft from what we already wrote. It leaves the price blank, because you set the price. It flags every line with no vendor quote. When an amendment comes, I ask what changed and have it update the worksheet and checklist.
You check: every vendor cost against the vendor's quote, line by line.
HR and payroll: check a pay run before it goes out.
Claude compares what is about to be paid against the approved timesheets and flags every mismatch, with a question for the manager on each flag.
You check: every flagged row before anything is submitted.
HR: state rules side by side.
We hire in many states, so Claude puts the rules next to each other, one row per rule and one column per state, with the official link and the date it checked. Over time you see where one state has a requirement nobody else has.
You check: every line against its official link. Anything new or unclear goes to our attorney before we act.
Project management: contract folder to project plan.
Claude reads the whole folder and gives you the plan, the equipment check, and a one-page risk list before kickoff. If the contract does not say, it writes "Ask the COR" (the contracting officer's representative) instead of guessing.
You check: the equipment check against the contract before anything is ordered.
Software licenses: renewals and license position.
Every month, know what renews in the next 90 days, where seats sit unused, and where a true-up bill might land, before the vendor tells you.
You check: the top five renewals against the actual contracts.
Inventory: make sense of the weekly reports.
Turn the weekly reports into one page: what changed, what to reorder, and what is wrong, such as received but not billed.
You check: three parts against the physical count.
Month-end report.
The same inventory numbers and the project board become a five-slide deck and a two-page PDF with charts, plus three lines for the cover email, saved to Drafts.
You check: every number against the report it came from.
Marketing: LinkedIn posts and blog posts.
This is how I do it. I start from a real job we did and write it the way I talk. Every week I ask Claude to go to our past performance and find something to write about. If I like it, I say cool. If I don't, I say pick another one until I like it. Then I say, now create an image, using the branding documents I already gave it: our logo, our colors, and our website. I did that in about five minutes. In the past, that cost a lot of money.
You check: every number and certification. I post it, not Claude.
Certification study.
Say you have to take an exam. Ask Claude to find the official exam objectives, or upload them yourself. Then have it act as an instructor who has taught the exam for ten years and build a study plan, a one-page cheat sheet for each domain, a practice test you can click through, and flash cards for what you miss. Ten minutes a day. No brain dumps and no real exam questions, because that can void the certification.
You check: the plan covers every objective on the official list.
7.
Make it run without you
What I said
A lot of people only use AI where they ask a question, get an answer, ask a question, get an answer. So you only use AI when you are touching the keyboard. You can tell it to do things on its own.
If you refine something until you get it the way you want it, you can say, do that for me every Monday at 7 AM. Do this for me every day at 6 AM. Do this on Tuesdays and Thursdays. You turn it into a repeatable process, you put it on a schedule, and it does it without you.
There are three steps:
Run it by hand until it is right twice. Fix the prompt each time it misses. Two clean runs in a row.
Save it so you can run it by name. Ask Claude to turn the chat into a saved skill.
Put it on a schedule, weekly or monthly. Claude runs it and the result is waiting for you.
You have to tell it where the result goes. A lot of what I have creates a file in a folder. Some of it sends me an email. The output could be a board, a file on your computer, a Google document, or a spreadsheet it updates every week. Whatever you decide, you have to tell it.
8.
Schedule smart: it costs money every time it runs
What I said
Four rules for anything on a schedule:
Weekly or monthly, not hourly. Hourly jobs burn through your allowance. Ask me first.
Medium effort. It is plenty for every job in this class.
Name it with your name, then the job, so we can tell whose job is running.
Every Friday, check your list. Turn off anything you did not use that week.
9.
Cowork or Claude Code: which one do I use?
What I said
Cowork is best for email, documents, slides, and our project board. Its rules live in the project instructions. Claude Code is best for a folder full of files and for jobs that must run the same way every time. Its rules live in the CLAUDE.md file in your folder.
10.
The rules, and who presses send
What I said
I have given my agents their own email addresses. I never let one send an email from our company address, because anything going out from the company, I need to hit send myself. I have a hard rule on that, because it can make a mistake. But I do let an agent send from its own address. When I say, send an email to someone on my team, it sends it from that email.
The rules are the same as always:
Drafts only. Claude writes the email. You press send. Scheduled jobs too.
You are responsible for it. It makes fewer mistakes, not zero. Check it.
No guessing. Every prompt says: write Unavailable instead of making something up.
Watch who sees what. Pricing, pay, and employee information only go to the right people.
Keep private information out. No CUI (controlled unclassified information), Social Security numbers, medical details, or passwords.
11.
Homework
Run your team's job this week. If it works twice, put it on a schedule.
Paste the seven rules into your project instructions today.
Change one line of your team's prompt so it fits how you really work. Save that version.
Next class: one minute each. Show what it made, and whether it ran without you.
The Word
The scripture behind it
"Only simpletons believe everything they're told! The prudent carefully consider their steps."
Proverbs 14:15 NLT
This is the class I teach my own team on how to use AI in our business, and now you can learn it the same way they do. Check its work. Then let it run.
If you are not a member of our Patreon, join today. Coaching and mentorship designed to maximize your potential, increase your personal productivity, and fulfill your life’s purpose. Weekly teaching, a private community, and one-on-one coaching calls, all on Patreon.
LIBRARY TIER
$25/month
Hundreds of videos and weekly teaching on life, faith, family, business, and relationships. Private Grace Life Discord, early access, and bonus episodes.
MONTHLY 1:1 MEETINGS
$100/month
Everything in Library, plus a monthly one-on-one phone call with Rick or Isabella (coaching, prayer, personal or business guidance) and direct Q&A access.
DIRECT COACHING
$500/month
Monthly one-on-one video coaching sessions with Rick or Isabella, direct Q&A access, and the full content library.