← Adam Garceau

I Built My Own Check Printing Software With Claude Code

October 12, 2026 · The written version of my video. Reading time: about 5 minutes.

I was about to buy check printing software. Instead, I asked Claude Code to build it. Here is what that took, including the part that went wrong.

I mean the software, not Claude itself. Claude is the tool I use to make it.

The job

I do the books for a small business. One day the owner walked me through how he pays people. He had a box of blank check stock with the check printed on top and the stubs below. I only need to print checks a few times a month, so I had a choice: buy the software or ask Claude to build it.

I'm not a coder, so I used five rules to decide and to build. They'd work for any small tool you're considering.

Rule 1: Do the math before you build anything

Three questions:

• How often will I use it? A few times a month is the sweet spot. If it's every day, a paid tool with support is often worth the money.
• How simple is the job? Print a check on blank stock, with the check on top and two stubs below. That's one page with some text on it. A simple job is a good job for AI.
• What happens if it's wrong? For a check, a bad print costs me a sheet of paper, as long as I test before I use real stock. If a wrong answer could cost somebody real money and I can't catch it first, I don't build it.

Rule 2: Describe the job, not the code

I didn't tell Claude which language to use or how to structure anything. I described the job in plain words. Claude's answer was that it could write a small check printer itself for free, and it wrote the first version right there.

The first version came back in about six minutes. It filled in a form, wrote the amount out in words on the amount line, and showed a preview of exactly what would print.

Here's the part the Short left out. Getting a working first version was fast. Getting to something I trust took longer: about 22 minutes from the first message to a Mac app I trust, including the test page and turning it into a Mac app.

Rule 3: Ask for the guardrails a paid app would have

A paid app doesn't just print. It stops you from making expensive mistakes. So I asked for the same:

• A register of every check I've printed.
• The next check number assigned automatically.
• A refusal to print the same check number twice, unless I tick a box that says reprint, so I can't do it by accident.

None of that is fancy. It's the boring part of what you pay for, and it's the part I would have forgotten to ask for if I'd only asked for "a check printer."

Rule 4: Test with something cheap first

This is the part that went wrong. The AI filled in my bank's routing number from memory, and it was wrong. It was a different bank's number. Before anything printed, I had it check the number against a public routing number lookup, and that's what caught it.

The app also has a mode that prints on plain paper with fake blank numbers and a big TEST VOID across the middle. I ran that first, then held the page up to my real check stock against a window to see whether the lines matched. A wrong number on a real check would have wasted a check and maybe a bank fee. Checking first cost me nothing.

One limit to be upfront about. The bank line at the bottom of a check is normally printed in magnetic ink, and my printer doesn't use that. The app draws the line with a free font. Banks are allowed to process a check from an image of it, so a sharp, correctly shaped line usually reads..

Rule 5: Keep it yours

The app runs on my Mac and nowhere else. It listens only to my own machine, so nothing outside the computer can reach it. My bank numbers live in the Mac's own keychain and never leave the computer. To open it, the app emails me a one-time code, and it locks itself after 15 minutes.

There's no subscription and no company that can change the price or shut it down. It's a native Mac app of about 1,300 lines of code. That ownership is the real reason I built it. Someone in the comments of a video about canceling Claude Code put it well: they built a tool once and use it for the rest of their life, and it doesn't matter if they ever get cut off.

When you should not build

• Don't build payroll, tax filing, or anything that handles other people's private data. The check printer is a printer. It doesn't do my accounting, and it holds no one's information except my own bank line.
• If you'd spend more hours building and fixing than the software costs, buy it. A paid app comes with support and updates from a company that checks its formats against banks. I don't have any of that. I have a tool that does one job for me and that I understand.

Quick checklist

1. Use it a few times a month, not daily.
2. The job fits on one page.
3. A wrong answer is cheap, and you can catch it before it counts.
4. You asked for a register, auto-numbering, and a duplicate block.
5. You tested on plain paper before real stock, and verified every number against a public source.

If you want the safety side, how to stop an AI from wrecking your files while it helps you build things like this, read Claude Code Deleted 48,000 Files.

I'm a marketer who builds his own tools. More at adamgarceau.com.