When I took over as treasurer for a small nonprofit, I inherited a problem I didn't know about yet.
The previous tax preparer hadn't filed the organization's annual return in three years. Three years is the IRS's line. Miss three consecutive filings and your tax-exempt status is revoked automatically. There's no appeal process. The only way back is to apply for reinstatement.
So I applied. The IRS wrote back. They needed more information.
If you've ever opened a letter from a government agency, you know the feeling. It's the only mail I read standing up in the driveway. Dense language, reference numbers, a deadline, and a list of requests you could read two or three different ways.
I promised connections this week, meaning AI hooked up to your email and calendar. Then a letter from the IRS jumped the line. Connections are still coming.
Last issue I called a skill your version of a technical order: write down how you want something done once, and it runs to the same standard every time. This issue is what that looks like on a real task I had to get done.
What I actually did
I blacked out our tax ID number and uploaded the letter to ChatGPT. My instructions were plain: review this letter, tell me what they're asking for, ask me whatever questions you need to fill in the gaps, then draft the response.
Document review, then a drafted response. That's the whole job.
ChatGPT walked through the letter item by item. The IRS wanted a signed declaration, a description of what the organization does, and financial information. Then it asked me questions, one at a time, for the things only I could answer.
Once it had what it needed, it produced a finished response organized to match the letter. Their item one, our answer one. Anyone reading it could check that every request was covered.
Then I made it a skill
Quick refresher from last issue: every job on an F-15 had a technical order behind it, the same procedure every time, wherever the jet was parked. A skill is your version of that. You write it once, and you stop re-explaining yourself.
When the response was done, I asked ChatGPT to turn the whole process into a skill. It wrote it, I saved it, and it now lives in my account as "eric-doc-review."
The rule at the top is one sentence: do what the task requires, no more and no less. I wrote that line during my own review of the first draft, when I caught something the IRS never asked for.
A few of the other rules it now follows every time:
• Read the document first. Figure out exactly what it's asking for before doing anything else.
• Only ask questions whose answers would change the response.
• Once there's enough to finish, stop asking and write it.
• Organize the answer the same way the request is organized.
• Before handing it back, check two things: did I answer everything, and did I add anything nobody asked for?
The next letter I get probably won't be from the IRS. It might be from the county, an insurance company, or a vendor with a contract. Different senders, same steps: read it, figure out what they want, ask me only what's missing, draft the answer, check it.
That's the test for whether something should be a skill. Do you run the same steps every time, no matter what document lands on your desk? If so, write the steps down once and stop re-explaining them.
What I didn't do
I didn't take the draft and mail it to the IRS.
I read every line. I checked it against the letter. I made a couple of small adjustments, then filed it.
One example: the draft included details about our spring fundraiser. The IRS hadn't asked for any of that. It made the response longer without making it stronger, so I removed it.
That edit is what drove the core rule in the skill. Stick to the facts about what was asked for and stop there.
Back to the technical order. Even with the TO in hand, a crew chief still signs for the work. You don't sign off on a jet because the paperwork says it's ready.
The AI did the heavy lifting on the first draft. My name is on the response. So the final check is mine.
The number I took off the letter
Before I uploaded anything, I blacked out our nonprofit's tax ID number.
Honestly, it probably didn't matter. A nonprofit's tax ID isn't a secret. The IRS publishes the name and ID number of every organization that loses its status this way, right on its own website.
I removed it anyway, because the habit is what protects you later.
The next letter might have a Social Security number on it. Or a bank account number, or a date of birth. ChatGPT didn't need our tax ID to understand what the IRS was asking. It won't need yours either. You get the same output without sharing that information.
The low-tech way works fine: cover the number with a sticky note before you scan it or snap the photo.
Before you upload any document, cover the numbers that identify you: Social Security, tax ID, account numbers. The AI can read the letter without them.
Where it stands
The IRS hasn't responded yet. When they do, I'll tell you how it went, good or bad. That's the honest measure of whether this process worked.
Try this
Next time an official letter shows up (insurance, the county, a benefits office, a contractor), try this before you start writing a reply:
1. Cover or black out any ID numbers.
2. Upload the letter and paste this:
Review this letter. Tell me in plain language what they are asking me to do and by when. Then ask me only the questions you need answered to respond, one at a time. When you have enough, draft the complete response, organized to match their requests. Don't add anything they didn't ask for.
3. Read every line before you send it.
If it works and you know you'll do it again, use the same line from last issue: "Turn what we just did into something you'll remember, so I don't have to explain it again next time."
Hit reply and tell me what kind of letter you'd try this on first. I'll share a few in a future issue, with the numbers blacked out. I’d love to connect on LinkedIn as well https://www.linkedin.com/in/eric-skelton-ai-over-50-newsletter/
— Eric
