June's AI Community of Practice session was about the step most people skip when they try to get something useful out of AI: giving it context. Small group, real conversation. If you couldn't make it, here's what came out of it.
Watch the recorded portion of the session on YouTube. A transcript is available on the video page.
Why your AI feels mediocre
If you've ever typed a five-word request into an AI tool, gotten a generic answer back, and decided the whole thing was overhyped: this one's for you. Because that's almost never the tool. It's the context.
Out of the box, AI doesn't know what you value. It doesn't know your community, your forms, or how your team actually works. So it gives you the average answer for the average user, which is exactly the bland, obviously-AI-written draft that makes you cringe. Context is what closes that gap, and it's the step people skip the most.
"AI works best when you bring it into your workflow with the right context."
Treat AI like a brilliant intern
The metaphor we keep coming back to: AI is a brilliant intern. It has a PhD in everything, but it's its first day, every day. Eager, fast, desperate to please, and it will happily hand you the answer it thinks you want.
But it's never set foot in your county. It doesn't know your board, your forms, or that the river floods every spring and your commissioner cares a lot more about septic reports than the thing you're working on. You would never hand that intern a task and walk away. You'd sit them down, brief them, and give them the binder of SOPs first. That's exactly how to treat AI, because your community, your documents, your voice, and your priorities are what make the answer useful.
Three kinds of context you can add
- Documents. Give AI a PDF whenever you can: a protocol, a brief, a style guide. The model has more to work from, and the answer gets noticeably better.
- A role. Tell it who to be. "Act as our communicable disease coordinator at a department of 80,000." "Act as a resident with a sixth-grade reading level reviewing this brochure." "Act as a skeptical reporter. What are the three things I'll have to answer for tomorrow morning?" Role-setting changes everything, and almost everyone skips it.
- Constraints. Tell it what not to do, and where the edges of the box are. Sixth-grade reading level. Under 200 words. Cite your sources. Don't reference case specifics in the public version. Constraints are how you stop AI from doing the thing that makes you cringe.
The context that only lives in your head
Documents and roles only get you so far, because a lot of what makes your work yours never made it onto paper. It's in your head and your colleagues' heads: what the commissioner actually cares about, what the real priority is this quarter, why a particular draft feels wrong.
The fix is simpler than it sounds: just say it. When you read an AI draft you hate, the running commentary in your head ("this reads like it's for a commissioner when it's actually for the public") is the context. Say it out loud as feedback and it goes straight into the work.
And you don't have to type it. Voice memos are one of the most useful habits we talked about. Your phone transcribes for free, and so do most note-takers and meeting tools. Two minutes of talking in the car between meetings can beat twenty minutes of typing a briefing. One of our colleagues records every stray thought into a folder (over a thousand of them now), and instead of scrolling back to find what she was thinking about a project in March, she just asks the system. Those memos become their own running context.
How AI actually works (and why it matters)
Here's a simplified but useful way to picture it. An AI model predicts the next word, and a chat re-reads the entire conversation every single time you ask it something. That has two practical consequences worth knowing.
- Early uploads fade. The protocol you attached at the top of a long chat is now competing with everything you've said since. The longer the conversation, the less weight it carries.
- Long chats hallucinate more. The more the model is juggling, the easier it is for it to miss a detail or fill a gap with something that just sounds right. Check the answers near the end of long chats a little harder, and don't be shy about asking it to cite its sources or "did you make this up?"
Chat vs. knowledge base
There are two ways to hand AI your documents. One is to drop them into a chat for that one conversation. The other is a knowledge base: documents stored outside the chat that the tool can always reach. That's RAG (Retrieval-Augmented Generation): the model chunks your documents and pulls back only the relevant pieces when your question calls for them. It shows up as Custom GPTs, Agents, and similar features depending on the tool you use.
The rule of thumb: a chat is a conversation; a knowledge base is a library with a librarian. For the documents and tasks you come back to over and over, build the knowledge base.
Where PH360 fits
Here's the honest version. Everything in this post is doable by hand. You can gather the documents, write the role, set the constraints, build the knowledge base, and keep all of it current. It works. It's also a second job, and most public health teams don't have a spare person sitting around to run it.
That's the whole reason we built PH360™. Remember the chat-versus-knowledge-base point from earlier? PH360 is the knowledge base, already built. It's an AI platform made for local public health, and the context is in the box from day one. The FDA Food Code is loaded. CDC communicable disease guidance is loaded. The role, the constraints, and the guardrails that keep it from quietly inventing a citation are built in and maintained for you. You're not opening a blank chatbot and explaining what a health department is for the hundredth time. You're starting from a tool that already knows the work.
So the job changes shape. Instead of briefing AI from scratch every time, you bring the two things only you have: your jurisdiction and your judgment. PH360 drafts the boil-water notice, the grant narrative, the inspection summary. You read it, fix what's yours to fix, and sign your name on it. That last part is the point. The signature stays human.
If your department is weighing AI right now, the fastest way to judge it is to watch it work on something off your own desk. Book a walkthrough →
The AI Community of Practice runs through the year, with a break over the summer. It's a free, open conversation for people working in or adjacent to governmental public health who are thinking seriously about AI. Learn more and join the community →
