Sep 24, 2026 · Free Guide

“BLUEPRINT” Build One AI Skill, Sell It to 1,000 Businesses

Greg Isenberg just laid out how he'd make $10 million with AI agents. Buried in the middle is one line most people scrolled right past: a text file that teaches an AI how to do bookkeeping for dental offices. That's not a side note. That's a business. Build it once, sell it a thousand times, like a book. Here's the full blueprint, step by step, no coding required.

Credit where it's due. This breaks down the "industry skill libraries" idea from Greg Isenberg's article How I'd make $10 million with AI agents. Go read the whole thing. This guide is the how-to for that one play.
THE BIG IDEA

Claude Code is the new operating system. Skills are the apps.

Think about what an operating system does. Runs programs. Holds memory. Plugs into the world. Has config files. Has apps. Now look at Claude Code:

A NORMAL OSCLAUDE CODE
Runs programsThe AI model
Holds memoryThe context window
Plugs into the worldTools and connectors (MCP)
Config filesMarkdown files
AppsSkills

On this operating system, an app is a text file. A skill is a markdown file written in plain English. Drop it in a folder and the AI can suddenly do a job it couldn't do before. No code. No app store review. No deploy.

And almost nobody is building them for real businesses yet. Greg points out there are a few thousand MCP servers (the connectors that let AI plug into software) and 30 million businesses in America. Almost none of them are connected. Most businesses only exist in the owner's head. Nothing is written down, so no AI can run any of it.

An industry skill is you writing down the playbook for a whole industry. Once. A good one that saves a business time and money sells for around $200, one time or every month.

THE PLAY

One industry. One very specific job.

Not "AI for dentists." Too vague. Nobody buys vague. Bookkeeping for dental offices. One job, one industry.

You write a skill that teaches the AI exactly how dental offices do their books. The insurance stuff. The categories. The monthly reports. Now the office manager or their bookkeeper does the monthly close in a fraction of the time.

“The customer's alternative is hiring a human.”
Greg Isenberg, on industry skill libraries

Here's why this is different from "go sell voice AI agents to plumbers." Both make money. Only one scales like a book.

CUSTOM AI AGENCY
Custom setup for every client
You babysit every deployment
Income capped by your hours
Client #50 is as much work as client #1
INDUSTRY SKILL
Same file for every customer
The customer runs it themselves
Income capped by your reach
Update the file once, everyone gets it
THE MATH

$200 a month. Count the offices.

There are tens of thousands of dental offices in the US alone. You don't need most of them. Watch what happens at each step:

10 offices × $200
$2,000/mo
Your first real proof
100 offices × $200
$20,000/mo
A real business
1,000 offices × $200
$200,000/mo
The reel math

Real talk: 1,000 is the ceiling, not the starting line. Your first goal is 10. Ten offices proves the thing works and pays for itself. And office #500 costs you almost nothing extra to serve. That's the whole point.

WHAT YOU'RE SELLING

What a skill actually looks like

A skill is a folder. Inside is one main file called SKILL.md (the playbook) plus any reference files the AI should read when it needs them. The same format also works in other AI agents, not just Claude. Here's what the dental bookkeeping one could look like:

dental-monthly-close/
├── SKILL.mdthe playbook (below)
├── references/
│ ├── chart-of-accounts.mddental-specific categories
│ ├── vendor-rules.mdHenry Schein → Supplies, etc.
│ └── insurance-deposits.mdhow EFTs and write-offs work
├── templates/
│ └── monthly-report.mdwhat the owner actually sees
└── examples/
└── approved-closes/real past months, names removed

And here's a stripped-down version of the main file. Read it. It's just instructions, like you'd give a new hire.

Example SKILL.md
---
name: dental-monthly-close
description: Monthly bookkeeping close for US dental practices. Use when the user uploads bank or card exports, a production and collections report, or asks to close the month, categorize transactions, or build the monthly report for a dental office.
---

# Dental Practice Monthly Close

## Who this is for
An office manager or outside bookkeeper closing the books for a general dental practice. You draft. A human approves. You never post anything to the books yourself.

## What you need before starting
1. Bank and credit card exports for the month (CSV)
2. The production and collections report from the practice software (Dentrix, Eaglesoft, Open Dental, etc.)
3. The practice's chart of accounts (references/chart-of-accounts.md)
4. Last month's approved close, if there is one
If anything is missing, stop and ask for it.

## Step 1: Categorize transactions
- Use references/chart-of-accounts.md and references/vendor-rules.md
- Dental supply vendors go to Dental Supplies. Dental lab invoices go to Lab Fees. Never mix the two.
- Unknown vendor? Do not guess. Put it on the Review List.

## Step 2: Match deposits
- Insurance payments often land as one EFT covering many patients. Match each one to the insurance payment totals in the collections report, not to single patients.
- Card payments land as batches from the processor, minus fees. Record the fees as Merchant Fees.
- Any deposit you can't match goes on the Review List with the date and amount.

## Step 3: Tie production to collections
- Pull gross production, insurance write-offs (contractual adjustments), net production, and collections from the report.
- Collections should tie to bank deposits, give or take timing. Show the difference and explain it.

## Step 4: Build the monthly report
Use templates/monthly-report.md. Include:
- Gross production, adjustments, net production, collections
- Collection rate (collections / net production)
- Supplies and lab fees as a % of collections, vs last month
- The Review List

## Rules
- Draft only. Never change the books directly.
- Never give tax advice. Flag tax questions for the practice's CPA.
- Remove patient names from everything you output.
- When unsure, flag it. A flag is fine. A confident wrong answer is not.

The top part between the dashes is the name and description. That's how the AI knows when to use the skill. Everything below is how to do the job.

Heads up: this is an example of the structure. A real one gets written with an actual dental bookkeeper, not copied from a guide. The reference files, the vendor rules, and the weird cases are where the value lives.

STEP 1

Pick your industry and your job

This is the most important decision you'll make. Run every idea through these six filters. If it fails one, keep looking.

01
It's text and document work
Reading, sorting, categorizing, matching, summarizing, drafting. That's what AI is reliable at today. Not physical work, not gut-feel sales calls.
02
It repeats with the same shape
Every week or every month, same inputs, same output. That's what makes one file work for a thousand businesses.
03
It needs insider knowledge
Rules a generic AI doesn't know. How dental insurance deposits work. How a property manager triages a leak. That knowledge is your moat.
04
Someone already pays for it
Either they pay a person to do it, or the owner burns nights on it. If nobody's paying in time or money, nobody pays you.
05
Thousands of near-identical businesses
And you can find them. Associations, Facebook groups, trade shows, subreddits, local directories.
06
You have the expertise, or can borrow it
Never worked in dental? Partner with someone who has. A bookkeeper, an office manager, a consultant. Split the revenue.

10 ideas that pass the filters. Every one is text and document work, and every one keeps a human in the loop.

IDEA #1
Dental offices
Monthly bookkeeping close: categorize, match insurance deposits, build the owner report.
Human check: Bookkeeper approves every entry
IDEA #2
Property managers
Triage maintenance requests: urgency, which vendor, draft the tenant reply.
Human check: Manager approves before anything is sent
IDEA #3
Roofers, HVAC, plumbers
Turn job notes into a written estimate and scope of work using their price sheet.
Human check: Owner reviews the number before it goes out
IDEA #4
Vet clinics
Turn the vet's shorthand into clean discharge instructions and follow-up emails for pet owners.
Human check: Vet signs off on every set of instructions
IDEA #5
Small law firms
Turn intake forms and call notes into a one-page case summary with missing info flagged.
Human check: Attorney reviews. Zero legal advice from the AI
IDEA #6
Insurance agencies
Compare last year's policy to the renewal, list what changed, draft the client email.
Human check: Agent checks coverage before sending
IDEA #7
Restaurants
Turn a week of vendor invoices into a food cost report and flag price jumps.
Human check: Owner or chef reviews the report
IDEA #8
Mortgage brokers
Read what the borrower sent, list what's missing, draft the follow-up request.
Human check: Broker confirms the checklist
IDEA #9
Nonprofits
Turn program data and staff notes into the quarterly grant report in the funder's format.
Human check: Director reviews before submitting
IDEA #10
Real estate agents
Pull every deadline out of a signed purchase agreement into a transaction checklist.
Human check: Agent or coordinator verifies every date

Best move? Pick an industry you've worked in, or one where your spouse, your cousin, or your old coworker has. Insider knowledge is the product.

STEP 2

Design it with Greg's 7 questions

This is where most people fumble. They start writing before they know what the thing is supposed to do. Answer these seven first. Here they are, filled in for the dental example:

QUESTION 1
What wakes the agent up?
The bookkeeper types “close the month” on the 1st. Later, it can kick off when the bank export lands in a folder.
QUESTION 2
What context does it need?
Bank and card exports, the production and collections report, the chart of accounts, last month's approved close.
QUESTION 3
What tools can it call?
Start with reading files and building spreadsheets. Connecting to their accounting software comes later, once it's earned trust.
QUESTION 4
What can it do on its own?
Categorize routine transactions from known vendors. Match deposits. Draft the report.
QUESTION 5
What needs approval before it goes out?
Everything that touches the actual books. New vendors. Anything on the Review List.
QUESTION 6
When does it escalate to a human?
Deposits it can't match. Numbers that don't tie. Anything that smells like a tax question or fraud. That goes to the bookkeeper or the CPA.
QUESTION 7
How do you know it worked?
The books tie to the bank to the penny, and the bookkeeper's edits go down month after month.

Then build the smallest useful version. Greg calls it the minimum viable agent: draft and approve. The AI drafts, a human hits approve.

Everyone wants to ship the fully automatic robot employee on day one. It demos great and breaks in real life. Your customer might be buying their first AI tool ever. Earn the autonomy.

STEP 3

Write it, then prove it on 50 real examples

Writing it. You don't write the skill from scratch. You interview someone who does the job for a living, record the call, and let Claude turn the transcript into the skill. Ask them to walk through a real month, step by step, including every "oh, except when..." moment. Those exceptions are the gold.

Prompt: interview → skill
I'm building a Claude skill that does [JOB] for [INDUSTRY].
Attached is a transcript of an interview with [NAME], who has done this job for [X] years.

Turn it into a skill folder:
- SKILL.md with a clear name and description (when Claude should use this skill), the step by step process, and the rules
- references/ for anything industry specific (categories, codes, vendor names, terms)
- templates/ for the finished output the customer sees

Write it in plain English. Every time the expert said "it depends," turn that into a rule or a flag for human review.

Before you write anything, ask me every question where the transcript is unclear.

Testing it. This is the step that separates a real product from a cool demo. Greg's rule: take 50 real examples of the job, run the skill on all of them, and score every one.

An example is one real piece of the job where you already know the right answer. For dental books, that might be 50 past months from a handful of offices, or 50 of the tricky deposits that trip bookkeepers up. Names removed. Right answers written down by the person who actually did the work.

EXAMPLE SCORECARD · 50 RUNS
42
Right
Matched the answer key
6
Flagged
Asked a human. That's fine.
2
Wrong
Confidently wrong. Fix these.

A flag is the AI saying "I'm not sure, you check this." That's the system working. The wrongs are what kill you. Fix the skill, run all 50 again, and do it every time you change anything. Greg calls it your gym. Paste this into Claude Code with your examples in a /tests folder:

Prompt: run the 50
Run the dental-monthly-close skill on every folder in /tests.
Each folder has one real example (names removed) and an answer-key.md written by the person who actually did the job.

Score each one against the answer key:
RIGHT = matches the answer key
FLAGGED = you put it on the Review List instead of guessing
WRONG = you gave a confident answer that doesn't match

Give me a table: example name, score, and what went wrong on every WRONG.
Then tell me the one change to SKILL.md that would fix the most WRONGs. Don't make the change yet.

Here's the bonus nobody talks about. That scorecard is your best sales asset. "We ran this on 50 real months of dental books, here's exactly what happened" beats any sales pitch you could ever write.

STEP 4

Free testers, then referrals, then sales

Don't try to sell it on day one. Prove it on real work first. Here's the order:

  1. Find 5 testers. Offer it free for 60 days. Look in industry Facebook groups, LinkedIn, local offices, and your own network. Bonus target: bookkeepers who specialize in dental. One bookkeeper can serve a dozen offices.
  2. Run it side by side. They keep doing it the old way while the skill runs next to it. Compare. Every difference is a fix.
  3. Ask for referrals once it works. Not before. A happy tester is the best salesperson you'll ever have.
  4. Sell to the referrals. Offer to run it free on their last month. Seeing it work on their own numbers closes the deal.
DM: finding testers
Hey [name], quick one. I'm building an AI tool that does the monthly bookkeeping close for dental offices. Categorizing, matching insurance deposits, the monthly report. Your bookkeeper still approves everything.

I'm looking for 5 offices to test it free for 60 days. We set it up on a 30 minute call and run it on last month's numbers, right on your computer. Then you tell me honestly where it screwed up.

No cost, no contract. If it saves you real time, all I'd ask is an intro to one or two offices you know. Open to it?
Ask: referrals
Glad it's saving you time. Quick ask: who are two other offices (or a bookkeeper who handles a few) that would want this?

I'll run it on their last month for free so they can see it work before they pay a cent.
Sales line: use your real scorecard
We ran this on [50] real months of dental books. It got [44] right, flagged [5] for a human to check, and missed [1]. Here's the full scorecard.

Want me to run it on your last month so you can see it on your own numbers?

Only use your real numbers in that last one. If your scorecard is 38 out of 50, say 38. Keep testing until it's a number you're proud of.

STEP 5

Price it and deliver it

You can sell it for around $200 one time, or $200 a month for access and updates. Go monthly. Here's why.

The file itself can be copied. A text file doesn't have a lock on it. So you're not really selling the file. You're selling:

  • The tested version. The one with the scorecard behind it.
  • Updates. Insurance rules change. Software exports change. AI models change. You keep it working.
  • Setup and support. Someone to call when it does something weird.

How to justify the price: ask how many hours the job takes them every month, multiply by what that person costs per hour. If the skill saves more than $200 of time, the math sells itself. Use their numbers, not yours.

DELIVERY 1
Package it
The skill folder, zipped. Or a private GitHub repo you add them to. Same thing, different door.
DELIVERY 2
They install it
In the regular Claude app: Customize, then Skills, then upload the zip. Anthropic's help center says Free, Pro, and Max plans can all upload their own skills. Claude Code users just drop the folder in their skills folder.
DELIVERY 3
30 minute setup call
You install it with them and run it on last month's numbers together. They watch it work on their own data. That call is where they fall in love with it.
DELIVERY 4
Monthly update
New version, plus a short note on what changed and why. This is what they're paying for every month.
REAL TALK

The honest caveats

I'm not going to pretend this is a push-button money machine. Read these before you start.

This is early
There's no App Store for skills yet. A few small marketplaces exist, nothing big. You are the distribution. That's the bad news and the opportunity.
The file can be copied
Assume someone will forward it. That's why you sell updates, testing, and support, not the file.
Accuracy is your job, forever
Rerun your 50 examples every time you change the skill, and every time a new AI model comes out. A model update can change how it behaves.
No unsupervised judgment calls
Money, medical, legal: the AI drafts, a qualified human decides. Permanently, not just at launch. You're not their CPA. Say that in writing.
Their data is sensitive
Dental offices have patient data and HIPAA rules. Remove names from anything you touch, have them run it on their own Claude account, and ask what they're allowed to share before you see anything.
Most attempts won't hit $200K
Greg says it himself: most startups fail, AI doesn't change that. It just makes finding out cheap. 10 offices at $200 is still $2,000 a month you didn't have.
YOUR FIRST 30 DAYS

The plan

WEEK 1
Pick and interview
Pick one industry and one job. Find your insider. Record a 60 minute call where they walk you through exactly how they do it, every weird case included.
WEEK 2
Write it and collect 50
Run the interview prompt. Gather 50 real examples with known answers, names removed.
WEEK 3
Test, fix, retest
Run the 50. Fix the wrongs. Run it again. Repeat until the wrongs are near zero.
WEEK 4
Five free testers
Send the DM. Set up 5 businesses. Run it next to what they already do and compare.
MONTH 2
Referrals, then revenue
Ask every happy tester for two intros. Run it free on the referral's numbers. Then charge.

This is the iPhone app store in 2009. The shelves are empty. The people who write down how an industry actually works, test it properly, and show up with proof are going to own those shelves.

Pick one industry. Pick one job. Start this week.

The Next Step

Guides show you what. AI Builders shows you how.

Inside the community, I walk through every build live, including the stuff that doesn't make it into guides. Regular people (not developers) figuring out AI together, shipping real projects, asking me anything. No fluff, no theory, just the actual work.

Join AI Builders→

skool.com/ten-fold