 Prompt Local
Issue No. 9 · Tuesday, September 29, 2026 · By Scott Olsen | ||||
|
PULL UP A CHAIR I heard from a reader last week, and it got me thinking I skipped something. Prompting should have been Issue 1 or Issue 2 of this newsletter, not Issue 9. Here's why. One of my favorite interview questions, back when I was hiring, was always some version of: a user calls and says they can't print, how do you troubleshoot it. You'd get the most elaborate answers. People would walk me through print management, removing, reinstalling, and updating drivers, troubleshooting the network to make sure the printer and the computer were still talking to each other. What I was actually listening for was much simpler. Is the machine powered on? This issue is that same question, aimed at Prompt Local. It should have been one of the first things I covered, not the ninth. So, to the reader who sent it in: thank you, genuinely, for taking the time. I hope more of you do the same as this thing grows. Enjoy your week. If there's something you want to see covered, hit reply and let's talk about it. THIS WEEK'S PICK Writing prompts that actually workA reader's note last week got me thinking about something I skipped: what actually goes into a good prompt. So this week, no new tool. Just the handful of things that actually change what comes back, pulled straight from what OpenAI, Google, Microsoft, and Anthropic tell their own users, a couple of formatting standards a reader flagged that none of those vendor guides mention, and the one thing none of them will tell you, because it has nothing to do with the AI.
One more worth flagging as new rather than settled: OpenAI has reportedly walked back some of its own older advice to write long, step-by-step prompts, now leaning toward a couple of clear constraints and more room for the model to work the problem. I could not get my hands on OpenAI's own current consumer help article to confirm the exact wording, so treat that one shift as "reportedly," not gospel, until someone reads it directly. One of our readers sent in something that goes further than any of those vendor guides do: a target reading level, and a short list of formatting choices none of them mention. Plain-language and accessibility standards commonly call for an 8th grade reading level or lower, 5th grade where you can manage it, since more than half of American adults read below a sixth grade level. On formatting: skip italics and underlining, both genuinely harder to read for people with dyslexia and low vision, skip all caps, use bold for emphasis instead, and keep text left-justified rather than centered. None of that is a personal preference, it is standard plain-language and accessibility guidance, and it is worth pasting into a prompt as instructions right alongside the technique tips above.
THE PART NO PROMPT GUIDE MENTIONS Your prompt can be a public record tooNone of the four vendor guides above mention this, because it has nothing to do with getting a better draft and everything to do with what you are allowed to type. As covered back in Issue 6, an AI prompt can be a public record the same as an email, both what you typed and what you got back. North Carolina's own state IT department turns that into a plain rule: "Never enter personally identifiable or confidential information into publicly available generative AI tools." Information you type into a public AI tool is treated as "released to the public," and may be subject to a records request whether you meant it to be or not. This is not hypothetical. A reporter in Washington state filed public records requests across several cities specifically asking for officials' ChatGPT logs. In Bellingham, one of those logs showed an employee drafting an RFP for utility billing software, and asking ChatGPT to write requirements that would exclude one vendor "without explicitly saying so," and separately to write requirements that would favor another. That second vendor won a five-year, $2.7 million contract. Nobody had to hack anything or leak anything. Somebody just asked for the chat logs, and the city had to hand them over. The rule that actually matters: before you type anything into an AI tool, ask whether you would be comfortable with that exact text showing up in response to a records request next month. If the answer is no, that is your answer about whether to type it. HOW TO ACTUALLY USE IT 1. Before you type anything, run it past the records-request test above. Strip out names, addresses, case numbers, anything you would not want surfaced with your name attached. 2. Then build the prompt: who it's acting as, the background, the format and audience, a reading-level or formatting rule if your office has one, and a real example to match if you have one. 3. Read the first draft, then tell it what is still wrong in one or two sentences, don't start over. Every vendor's guidance agrees on this: the second try, not the first, is where the real draft shows up. 4. Same rule as every issue before this one: read it against what's actually true, and send it through whatever review your office already runs. THE RULEBOOK · TWO RULES WORTH A GLANCE What changed that affects how your office can use AI. Short, plain, and sourced. The part that keeps you out of the newspaper. Tagged Local for a city or county action, State for a change at the state level, Federal for a change at the federal level that reaches your desk.
ONE MORE QUICK WIN Start a shared doc of prompts that actually workedThe next time a prompt gets you a draft you barely had to touch, copy it into a shared doc before you move on. A running list of "prompts that worked for us" saves the next person in your office from starting at zero, and it is a lot more useful than any generic guide, this one included, because it is built from your own actual documents. | ||||
|
Know someone at another city or county who could use this? Forward it over.
Until next week, Scott Olsen, CGCIO Former CIO with eleven years in Florida local government Practical guidance for local government, not legal or procurement advice. |