---
id: "workshop:using-generative-ai-to-improve-everyday-workflows-for-public-professionals"
type: "InnovateUS workshop"
title: "Using Generative AI to Improve Everyday Workflows for Public Professionals"
url: "https://innovate-us.org/using-generative-ai-to-improve-everyday-workflows-for-public-professionals"
date: "2026-01-23"
duration: "60 minutes"
series: "The Prompting Lab: Integrating AI into Your Public Sector Work"
instructors: ["Deborah Stine"]
recording: "https://www.youtube.com/watch?v=uTzlnDq3SCE"
transcript: "automatic, from the recording"
source: "InnovateUS (innovate-us.org)"
licence: "CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org"
---

# Using Generative AI to Improve Everyday Workflows for Public Professionals

InnovateUS workshop · 23 January 2026 · 60 minutes

Using Generative AI to Improve Everyday Workflows for Public Professionals

Generative AI  is becoming central to how public agencies handle the routine, time-consuming tasks that shape everyday government.  This workshop provides public professionals with GenAI methods that go beyond basic prompting to more sophisticated methods that streamline routine tasks, such as distilling complex guidance, translating technical material, and organizing data into usable structures. Participants will also explore how GenAI can function as a strategic thinking partner. 

This session is designed for public professionals who manage basic and complex information flows and want to use GenAI responsibly to improve the effectiveness, efficiency, clarity, and quality of their regular work.

## Learning goals

Learning Goals:

• Understand how GenAI can streamline routine and complex workflows by transforming dense information into clear, usable outputs for their own use, colleagues, decision-makers, and the public.

• Apply structured prompting techniques that support higher-quality GenAI outputs, including methods for developing reusable prompts tailored to everyday government tasks.

• Recognize how GenAI can serve as a strategic thinking partner, extending staff capacity while maintaining responsible use practices within public-sector constraints.

## Speakers

- **Deborah Stine**, Senior Fellow, Innovate US

Part of the series [The Prompting Lab: Integrating AI into Your Public Sector Work](https://innovate-us.org/workshop-series/the-prompting-lab-series).

Recording: https://www.youtube.com/watch?v=uTzlnDq3SCE

## Transcript

Automatic transcript of the recording: names and terms can be misheard, and the times can be a few seconds off. Each time links to that moment of the video. The transcript alone: https://innovate-us.org/using-generative-ai-to-improve-everyday-workflows-for-public-professionals.txt (plain text) and https://innovate-us.org/using-generative-ai-to-improve-everyday-workflows-for-public-professionals.json (every segment with its start and end second).

**Deborah Stine**

[0:00](https://youtu.be/uTzlnDq3SCE?t=0) And thanks again. I do appreciate when folks introduce themselves in chat, you know, so that I know, have some idea of who is out there. So today we're going to talk about using generative AI to improve everyday workflows. And I just as a preface, I want to say that, you know, I did this, I did a workshop on prompting, let me think, I think it was like prompts unleashed. I did it, I think about a year ago, and a lot has changed in this past year. And I think it really emphasizes the importance. So I'm going to try to take you up in this presentation to, you know, like one level up from what I was able to do a year ago. And the other thing I always like to mention is that I, you know, there's a lot of great speakers on Innovate US that are enterprise level speakers. I'm the person kind of on the front line, the person who's actually implementing, I use AI tools every single day. In fact, this particular presentation is on Gamma.

[1:08](https://youtu.be/uTzlnDq3SCE?t=68) And the way I generated it is I went and I had some notion of what to do, I went into a dialogue with ChatGPT and then also with Perplexity. Those are the two tools that I use the most. And then I copied it into Gamma, which is another tool to help you generate presentations. And the reason I use it is because I'm terrible at graphics. It is, you know, might be an engineer, but I would definitely never make it as an artist. And so I need help on graphics and developing them. And Gamma is an AI tool that helps me do that on presentations. And I just got tired of using PowerPoint. So that is the tool that I'm using today. So we're going to cover these five areas. It's a lot to cover in basically 45 minutes. So I'm going to go pretty fast. If you have questions, enter them in chat or the Q&A function. I should say the Q&A function as we go along, and we'll get to as many as we can at the end. The good news is this session also has some office hours that we'll be doing in about a month.

[2:10](https://youtu.be/uTzlnDq3SCE?t=130) And that will also be a chance for you to kind of play with these tools. And I would say playing is the main thing to do. The first time you do this, don't do it on something super important. Play with something, you know, even personal, like your kid's soccer schedule or something before you really try to implement it. So we're going to do how and why to use Gen AI tools, brainstorming, research, summarizing documents, drafting inputs. And this is all based, these topics were based on a previous poll. So please make sure you fill out the poll that you have. Maybe it's not yet closed yet on the screen, but we can look at it later. Nope. Okay. Hold on. I guess now it's not working. Okay. No. Oh, there we go. Okay. So the first thing I always like to emphasize is there are a lot of tools out there.

[3:17](https://youtu.be/uTzlnDq3SCE?t=197) These are the five big ones. And different tools are good for different aspects of it. So if I'm writing something, like if I'm, say, writing something for a public policy person, I will use ChatGPT. That's what I like the most. If I'm doing analysis for somebody, then I will use Perplexity. And I like Perplexity because it links from sentence to source and I can check all the information. If I want to just kind of find out like what's out there, like on today, I wanted to, for this presentation, a lot of what I do for Innovate US, I wanted to find out examples of where this has been done at the state or local level. I will use Gemini because it will give, I find the best information. A lot of people like Cloud. A lot of people like Copilot. I used to like Cloud, but then it started like refusing to answer things. And so I stopped using Cloud, but I've heard it's improved, but I don't use Cloud very often. And I'm not going to include examples from Cloud, but most of the things I'm talking about for the top three can also be used in Cloud. And Copilot, I know a lot of government agencies have Copilot because they have something going on with Microsoft. And it is good, you know, for certain things, but for most of the work that I do, I prefer these other top three tools as opposed to the others, but they're constantly changing. I mean, even like something that I was able to do like a week or two ago, I can no longer do today or something or vice versa. So I just want to emphasize that these tools are like a moment in time. So one of the previous talk in this series, this topic came up. So I thought I would share this information, which is like, where are people finding AI most useful? And this is a study that's from the Federal Reserve Board of St. Louis. And just going from the top, you'll see, this is figure seven, writing communications is most useful than administrative tax, summarizing, searching for

[5:23](https://youtu.be/uTzlnDq3SCE?t=323) facts, generating ideas and so forth. So you can sort of see how people are using it. And then the other thing is, is how are people saving time? So people like myself who use it like every day, you'll see that we're saving, according to this analysis, most people like 40% are saving four or more hours or three hours. So like when you add those two together, it's over 50%. The less you use it, of course, the less time it's going to save you. So if you only like use it a day a week, then still about 10%, well, 30% are saving hours. So it is a real efficiency saver. But I also want to emphasize, and I will throughout this presentation, it also helps on strategic thinking. And that is just as important as everything else. And I apologize for any squeaking you hear, I have a new puppy and she's playing with her toy. And at some point, I'll introduce her. Okay, this is another analysis. This is from Pennsylvania. So I think that there's the speaker at Pennsylvania was on the series last week. And their study, which I found very interesting, also kind of talked about how their use cases in PA, and I'm based in Pittsburgh, by the way. So this is of interest to me. And you'll see, again, very similar data that is writing assistance, text generation, drafting emails was the biggest, then research idea exploration, learning new topics was second, and then summarizing was third. And according to their analysis, people spent about 35 minutes each day using like ChatGPT, and they saved 95 minutes.

[7:13](https://youtu.be/uTzlnDq3SCE?t=433) So overall, you know, we're seeing good results, you know, from these, these tools. So this is the what I mentioned earlier, this is the survey from last week. And I organized like the order of these based on what you guys wanted to know the most. And we're going to be starting with this first one, which is brainstorming. And then just we'll cover all four topics. I thought I'd also give you some real world examples, because that also came up as something that people wanted from the survey. So these are just a few. This is an article, whenever it says click here, by the way, when you guys get the PDF of this, you can click and go to that article. Washington City officials are using ChatGPT for government work. And so you'll see here prompt, using the mayor's voice, can you rewrite this letter to be a little more collaborative and less aggressive in tone? I do a lot on tone. I also ask things I tend, because I was a former professor at Carnegie Mellon. I also use the three C's, I ask, I write what I want. And then I say, make this more clear, concise and coherent. Because that is my thing that I do. Then you'll see on the bottom left there, it was a letter of support for grant funding. That was sent by the mayor that was written by ChatGPT. And then here's like an Instagram post that somebody wrote.

[8:42](https://youtu.be/uTzlnDq3SCE?t=522) This one I thought was interesting. This is basically where there's AI targeted digital advertising to try to recoup money. So there was in this Wilmington, Delaware, they had $24 million in unpaid water bills, and they were able to recover 1.1 million in bills. So obviously, it'd be better if it was closer to 20 than one, but still it kind of shows things like this, how AI tools can be used. But I did want to say, here's an example where things did not go so well. So this is a Bellingham, Washington. And in this case, a staffer used ChatGPT to try to do a vendor request, and he biased his results. And so at least according to the story, and I'm sure there's probably somebody from Washington State who knows more about this, they did it in such a way that the way they designed the request for this proposal favored one company vertex over this other company origin, and origin was like about a million dollars cheaper. So it also emphasizes how important it is to do what's called human in the loop, that somebody somewhere along the line needs to check what other people do to make sure things are not biased or that different perspectives are missing.

[10:25](https://youtu.be/uTzlnDq3SCE?t=625) Okay, hold on. I'm going to go to this option. Okay. So now we're going to go into brainstorming, planning and prototyping. So the main perspective that has changed since I was working on this is rather than kind of using these tools as getting information, like we're used to using, say, Google, like, you know, we type in information and we say, hey, you know, what's happening on Gen AI related to, I don't know, garbage collection or something, and then information comes back.

[11:43](https://youtu.be/uTzlnDq3SCE?t=703) That is kind of like basic level prompting. But what's really useful is if you use it to thinking AI is more as a strategic thinking partner. And as somebody who's like a solo entrepreneur, you know, I don't have colleagues per se, other than those who I ask, you know, to do these things. I don't have somebody next door to ask. And so this is a way to you know, think about to think through an issue before you actually work. And so I want to show you different levels of prompting relative to those tasks. So this is like a level one. This is a I'm just calling this a structuring prompt. So it's a standard prompt, it would have been what I say taught, like last year, your policy aid in this in this agency, based on this description, brainstorm three implementation options, with pros and cons, key stakeholders, equity, resource requirements, label any assumption or missing from missing information you're making.

[12:51](https://youtu.be/uTzlnDq3SCE?t=771) And this is there's nothing wrong with this prompt. And sometimes there's a good time to use it when you're trying to get certain information and sometimes not. So let's just take this, I just basically took that prompt, and I gave it an example of the state unemployment insurance agency. And to brainstorm these three implementation options. And then this is what I got. And you'll see here that it is, you know, gives everything kind of in bullet form description, pros and cons, the stakeholders and all these different things. So it does give me the information I want. And that information can be useful. Well, I'd like to go up though another level, and that is to incorporate community engagement. And so in this case, rather than me just saying, okay, this is like repetitive task, I'm going to first ask the staff to list their most painful repetitive tasks that take a great deal of time and need little human help. So if, for example, a prompt, like a prompt that I might give to ask this was to say current state employees have identified these things that take great deal of time, need a little human help for each task, briefly describe how you as an AI tool could help and what information you need. Okay, so I'm going to stop here. And I'm going to keep on talking. But what I want you to do is like answer this question for yourself. Like when you think about all the activities that you engage in, what is a painful repetitive task that takes a lot of time and needs some human help?

[14:40](https://youtu.be/uTzlnDq3SCE?t=880) Because by identifying that, that will help you figure out how to, you can use these tools. So here you have a state unemployment office. So you can do these two ways. You can create a survey and you can use Use Chet GPT or another AI tool to help you create a survey to ask this question in a Because I didn't have anybody to ask, what I did was I just used perplexity, again that's nice neutral way. what I use for research and analysis, and I said, you know, go out there, find me studies to find out what are the most painful things that state unemployment offices face. And it came up with this list of six things. The first is explaining standard decisions. And so this was the prompt that I used after that, the next level.

[15:43](https://youtu.be/uTzlnDq3SCE?t=943) So I went beyond it and I said, well, current state employees have identified these painful repetitive tasks that take a great deal of time and need a little human help. For each task, briefly describe how you as an AI tool could help and what information And then I just sort of gave the list of things. you need. And then this is, I'm only doing the first one on the list, but it gave information for all of them to say how AI can help, the information it needed and the confidence level. So the reason it's giving confidence level is because I have in my chat GPT general guidance, something that says for any information that you provide to me, tell me the confidence level you have in that analysis. And it will tell me low or medium or high. And that's really important because if it's low, then I look at it one way.

[16:46](https://youtu.be/uTzlnDq3SCE?t=1006) If it's high, then I look at it another way. So in this case about explaining standard decisions to claimants, AI can generate plain language explanations, tailored tone and reading level, and ensure consistency. And then the information it needs is some sort of approved decision template, legal language, the catalog of common decision reasons and triggers, and plain language standards, accessibility requirements, and prohibited phrasing. So if I was still working for government, which I did at one time, then I would find out this information and then come back and say, okay, I now have this information and I'm ready to try to figure out how AI can help. And let's start first with doing some plain language explanations.

[17:49](https://youtu.be/uTzlnDq3SCE?t=1069) Okay, so this is the sort of the next level, okay? And this is more the dialogue, okay? So this, I will say I do this a lot. It takes a lot of time, but I do really think that it helps me a lot of analysis. Remember, this is the brainstorming section. And so this is where you really want to have some sort of dialogue, some sort of interaction. Even if, you know, like the outcome for this might be, say, like a whole list of options that you then discuss at a staff meeting or something like that. So again, focusing on that first thing, you know, so we have an unemployment office. We asked what are the biggest problems? It was the first one on that list was explaining standard decisions to claimants. And then what I then ask is provide me a list of the questions you need to answer in order to help me meet this goal. And you can see this is, I'm going to go to another page, but on this first one, this is what it needs to know.

[18:52](https://youtu.be/uTzlnDq3SCE?t=1132) What benefit programs, claim types, and decision makers are in the scope? What decisions are truly standard versus discretionary or appeal prone? What are explicitly out of scope for automated explanations? And so here's more questions that it needs. What kind of tone do you want? You know, where must explanations be identical across cases, this kind of experience? So these are hard questions to answer, right? But the first thing is to generate the list of questions, maybe discuss them again with Again, we're talking about humans in the loop. these humans. How would we respond to this and develop the content that the tool needs to help you? So in the end, you're going to get a better product. And what I do is I ask it to then ask me each question like one by one, and then I will respond to them. So you'll see here for foundational scope and constraints, ask me each question one by one.

[19:57](https://youtu.be/uTzlnDq3SCE?t=1197) I will provide the answers, provide your summary of my response, which I may correct. Do not synthesize the results until I ask you to do so, because there's a tendency with these tools that they really kind of want to get you the answer. Like sometimes they're kind of like lazy, they're like, oh yeah, here's the answer. Don't bug me. Go away. And then other times they keep saying, do you want to do this or do you want to do that? And so you have to kind of learn like, like with my puppy here, you're the boss. And you know, sometimes you're going to say, stop, don't, I don't want to do this. And sometimes you're going to say, hey, this isn't good enough. I want, I want more. And so I like to think about it, you know, like I've had a million interns, fellows and things like that in the day. And it's learning, it's learning who you are and all sorts of other things. So in this case, you'll see here that the response after I put this prompt in is like the very first question, which was just that one, that one line, what benefit programs, claim types and decision categories are in the scope.

[21:04](https://youtu.be/uTzlnDq3SCE?t=1264) But then it's adding additional information in terms of what it needs to respond. So examples of dimensions to consider for your thinking only, not as a structure, program boundaries, claim types, decision categories, and so forth. The one thing I wanted to also mention is that sometimes like I'm, I'm somebody who likes to type, but for other people, they like to speak. And so like with chat GPT and some of the other tools, there's a little microphone there and you can just give them an answer and you can actually like talk and basically talking back and forth where, between you verbally saying your responses to these and maybe asking additional questions and providing additional context and it then providing written responses. Okay. So the next category is research or information synthesis.

[22:07](https://youtu.be/uTzlnDq3SCE?t=1327) So there's kind of like three categories here. So one is what I would call general research focus AI, and I will then use perplexity, But sometimes you want a literature review, you just want to find out what information you know, for that. is out there. So, and I've worked on a report over the past couple of months, it's going to be released in a few weeks about how to use waste heat from AI data centers for some sort of public And one of the things, you know, I did use perplexity, but I also want to see what was benefit. out there in say the scholarly literature. And here you might use a specialized tool as opposed to those more general tools like perplexity, like Semantic Scholar that has keyword matching, ELLICIT, which has research briefs, or Research Rabbit, which has visualizations, which I'll show in a second. Sometimes you want some synthesis, okay, then you might use a tool, a specialized tool like consensus that shows where peer reviewed studies agree or disagree, S-Cite, which is paper citations or ELLICIT, which is structured evidence tables.

[23:20](https://youtu.be/uTzlnDq3SCE?t=1400) Of these, I think ELLICIT is, you can see is on both charts. It's the most common one that people use. But this is Research Rabbit. And I think it's good, it generates this chart, so you can kind of see the interconnections among all the research. And then you can like look at different ones. Sometimes it's good to upload, like, you know, like, if you find like, in my case, there's like two or three papers that were exactly what I wanted. You can upload those to give context. And that as a starting point for your work, like who cited them, and then that brings you to more things. So these tools are getting more visual, they're getting much better. I tried ELLICIT, I don't know, a year or two ago. I was not impressed. ELLICIT today, much better than it was in the past. And I asked things like, do not give me, you know, I'll say only give me like peer-reviewed literature or literature that's from think tanks, and so forth.

[24:22](https://youtu.be/uTzlnDq3SCE?t=1462) And you can restrict what you have. So here's the example here. You're a county program manager, you have this messy data, scattered case notes, unstructured survey responses, some mixed feedback from stakeholders. And you want to transform that chaos into clarity, some tables, themes, actionable insights. And this, again, is a level one prompt. Basically, here's the data, turn this into a list, turn this list into a table. Here's the columns I want. Do not, it's good to tell what not to do, do not infer new categories without flagging them as assumptions. You can also do things like how to visualize it. So I've been using the visualizing tool quite a bit lately. And what I will do is I'll take the concept for the visualization, and then I'll put it into another tool, which right now I'm experimenting with Notebook LM, and have it go do visualization there or this Nanopro banana, and I put it there.

[25:31](https://youtu.be/uTzlnDq3SCE?t=1531) But generally speaking, for chat GPT, I don't do it for anything, like I don't like their slide decks, I don't like the visualizations, but it does help me brainstorm what the visualization should be, which then I put in a better tool for visualizations. Okay, so this is a level higher. So now I'm not just telling it what to do, I'm giving it more context. So I'm a policy analyst specializing in community feedback. I've collected 500 comments about broadband, summarize these across these categories, and presented as information with these three columns, and don't introduce new categories, and ensure there's a neutral tone, but accurately reflect the sentiment of the comments. So it's a little bit higher, because I'm giving it a lot more information as to the purpose and why I'm doing this. The next thing is the dialogue prompt, which is, again, where I'm going into the situation.

[26:36](https://youtu.be/uTzlnDq3SCE?t=1596) So I still give, I'm a program analyst, I'm trying to make sense of this data. I have a lot of disorganized information. But I will say, instead of jumping straight to organizing it, ask me questions to understand the context, right? So I'm not assuming what context it needs to know. Instead I'm asking it to ask me to provide that context, like what the data is about, what I want to learn from it, how I plan to use the results. That's really important. For example, how I plan to use the results. Like sometimes I might be presenting just internally to my internal team. Sometimes I'm presenting it to a board of some kind, an advisory committee. Other times I'm doing it for a report that will eventually become public. And how I organize the data and the degree of content I put in will really depend on who my audience is.

[27:42](https://youtu.be/uTzlnDq3SCE?t=1662) And so it will ask questions like this. And then, again, I have these sort of same rules. Ask me one question at a time. Form your understanding of each answer before asking the next. I do that because oftentimes it does not get the answer right. But I would say, I don't know, 75% it gets it right, 25% it does not. That's a pretty big percentage. Flag any assumptions. Do not create categories or synthesize until I change it per se. Because again, it's always rushing to try to do this. And then your goal is to help me co-develop a logical, useful structure for analyzing this messy program data through an interactive dialogue. And then I would go through that dialogue and answer lots of questions. Things that AI does not know that you might need to provide are things like the local political context, the relationship with stakeholders. For example, I was talking to somebody in Washington state, and they told me we never use the word stakeholder in Washington state because of concerns with the local tribal communities.

[28:52](https://youtu.be/uTzlnDq3SCE?t=1732) We instead use interested parties. Well, that's an important nuance from my local community that I want to put into it and make sure that I don't use words that are of concern to that group. Anything, again, more context, historical attempts and lessons learned in a community like if wind turbines, we had them and then one blew up or something. Well, you want to put that information in because AI does not know what's happening in your state or your community. Budget realities, equity impacts, all these things are important to give. Sometimes I find that the best way is not just to tell it, but to identify, say, articles and things like that and provide that as a, you know, I want you to consider this context from this article, and I will either upload the article or provide a link depending on what's allowable for that AI tool.

[29:54](https://youtu.be/uTzlnDq3SCE?t=1794) Okay, so the next category is on summarizing documents or public input. So again, this is sort of the level one prompt. Your role in this agency, your task is to summarize this, describe your audience, identify three to five key actions, flag uncertainties, use all of the text provided, do not create new policy or facts. That's important because of course we have this challenge every once in a while where the tools will make things up. I will say that that is much, much less of a problem today. Then in the past, and also it can be helped when you take those things, where do not create new policy or facts, and is that if you have it in your general instructions to your AI tool, that's another important aspect. So I want to give this as an example of like what I do when I'm doing something new. So somebody told me something I did not know at a workshop I'm doing on applied AI.

[30:59](https://youtu.be/uTzlnDq3SCE?t=1859) Well, this is like I'm learning from somebody who knows even more about this stuff than I do. And he said, hey, you can go to Google Sheets and just like you put equals A1 plus B1 plus C, you can say equals AI and use an AI prompt in the Google Sheet. I thought, oh, that's cool. I've never done that. So I thought I would give it a try. And so what I did is I went to Gemini, that's Google's version of an AI tool. And I say, hey, how do you do this? And give me an example. So it gave me this example, categorizing 311 service requests. Imagine you have a sheet with hundreds of entries from citizens, instead of a staff member reading each one to assign it to a department, you can use an AI function to do it automatically. Okay, so I thought, well, that's cool. So here we have the setup, date of the comment, the citizen comment saying, and then in column C, you'd have this kind of AI tool.

[32:08](https://youtu.be/uTzlnDq3SCE?t=1928) And it's again, a way to quickly sort of pivot from one thing to another. And so here is an example of what the AI tool might do. You can say equals AI, is this public comment positive, negative or neutral? Or summarizes 500 word testimony into two bullet points or extract the street address from this messy text. I was like, oh, this is all useful kinds of things. So then I said, provide an example with fake data in a Google Sheet that I can use as a demonstration for a workshop. So this is going back to what I said before, is don't just like do this with real data right away, just use some fake data so you can figure out how to use it and you can check. So you'll see here that it created this raw data, an A and B with the date and a citizen comment. And then part two is this AI magic where you have categorization with this formula.

[33:10](https://youtu.be/uTzlnDq3SCE?t=1990) So like classify this issue into one of these categories, road maintenance, utilities, parks and public safety, and then have an urgency score where this on a scale of one to five, rate this and then just only provide the number. So again, I'll remind you of this human in the loop. So you're having AI do sort of the first cut, like you might have say an intern do the first cut, but then you need to go back and take a look at the ones, the ones that particularly ones that are most urgent and say, okay, do I agree this is urgent or not urgent? So I did this and I copied the data. Well, actually even better, they gave me a Google sheet with this data. And then I said, and then I copied just to demonstrate it. I said, okay, well, I'm gonna do this. I'm gonna put the formulas just in the sheet I copied them. so I can use it. I create a category called categorize as a next step.

[34:10](https://youtu.be/uTzlnDq3SCE?t=2050) I put equal and I copied this prompt in and then look, it gave me this data. It gave me this urgency score. And I say, well, that's cool. I can see how I might use that. And so if I was in applying it, I would again, take a small data set, like if I have 500, I might take 25 and practice these cause I want to refine this. So like, I'm thinking, okay, well, I would like to color code this. So like things that are urgent are in red, things that are in like low, I might be yellow or something. Can I create more like equal AIs that are gonna go to that issue? Okay, so the next thing is drafting or editing public communications. So this is like, the hard part is not doing this. The hard part is like making sure you have a proper process to do it.

[35:15](https://youtu.be/uTzlnDq3SCE?t=2115) So first you wanna set the roles constraints. So you're saying as context, I'm a policy analyst in a city housing department, draft only from this text, do not add new facts, like anything that seems inconsistent. So sometimes these tools will try to like add in more information than I want them to add. And so it's really important to tell it not to do that because again, as you note here, you don't wanna have hallucinating where it's inventing things. And you also want to mirror how government works in reality. You also wanna think about who is your audience? I sort of mentioned this earlier. One might be a brief for your supervisor, an executive summary that they're gonna be using for decision-making. Another might be plain language bullets. So you're talking to a committee advisory board. So it's gonna take out the jargon and technical terms. So you wanna think about who the audience is.

[36:17](https://youtu.be/uTzlnDq3SCE?t=2177) And so lots of times I produce things for like several different audiences, even from the same exact content. And I just need to write that out and then have that influence the instructions that I give. I might start with the most intense level one, and then go to the next step board let's say the supervisor, and next for like a community meeting. It all sort of makes sense. And then you want it to reformat it depending on what it is. Like sometimes it might be an email. Sometimes it might be an implementation table. You need to think about what you're communicating to and the ultimate goal in terms of what are the next steps? What action are you expecting out of that activity? And you wanna check and correct. So you can ask these tools to sort of like themselves do a check. Ask them to identify its own errors.

[37:18](https://youtu.be/uTzlnDq3SCE?t=2238) Ask for it to say, okay, what are some nuances? What are some things I should think about? And that's another way to do these tools. Okay, so here's some like customizations you might do. So like make this content at an eighth grade reading level for a community newsletter. Adapt the key points for a social media thread for several different posts. Create talking points for the agency head. Translate everything into Spanish. So again, I go back to this dialogue mode. Rather than just saying like a prompt like this, you can say, I'm creating talking points for the head of my agency. What information do you need to know to help me do this? Ask me each question one by one. Do not synthesize until I tell you to do so. Because the head of a agency, they're gonna differ. Like sometimes they differ. Like sometimes they want, like some people I work with, they only want like something on a slide deck.

[38:20](https://youtu.be/uTzlnDq3SCE?t=2300) And sometimes they want an entire in-depth memo. So you wanna give that context for whoever it is you're speaking with relative to them as an individual. Okay, so I'm doing pretty well here and making it all the way through. So wrap up. Okay, so this is again, something I would like you guys to type into chat. What is one concrete workflow you will pilot in the next two weeks? So here's some ideas. You could summarize an agenda packet for a board meeting or public distribution, organize public comments, draft constituent emails, create tables of visualizations. So I want you to just kind of like take a moment to kind of think about that as I continue talking. The next thing is to name your guardrail. So somebody, again, there needs to be a human in the loop. Who, how are you going to do that?

[39:22](https://youtu.be/uTzlnDq3SCE?t=2362) So one might be having AI outputs reviewed by a manager before external use. No PII, that means no personally identifiable information. So that's something you always want to think about. For if you have attachments, where you're asking them to say, summarize a document, you want to always keep the original source text. That's important because I have found, particularly in recent times, that some of that source text disappears. And so I am glad that I kept this CDC document from a couple of years ago that I was using for a workshop I was doing, or I kept this document from USAID because things do disappear. And you want to make sure that you're able to access it. And then whenever you are doing things with numbers or dates or citations, you want to check those as well. One way sometimes I will do this is I will take my content from tool number one.

[40:28](https://youtu.be/uTzlnDq3SCE?t=2428) Like, let's just say I did something in perplexity and I will copy it over to ChatGPT and say, hey, take a look at this output. What are its strengths? Please check any facts to see if they're correct. What are its weaknesses? So actually if I was facts verification, I would do it the other way. I'd go from ChatGPT to perplexity. But, you know, I often will use three or four of these tools for the same thing just to make sure I'm not like missing anything. And it'll always come up with something. You know, nothing's going to be 100% correct. Here are two places you can go for more information. So GovTech has this prompt list and you can go there and there's all sorts of tools that are for state and local government tasks. I think it's like about 20 or 50 or something. And then Urban Institute, I actually heard a presentation a couple of days ago. I mean, it was yesterday from their chief information officer.

[41:31](https://youtu.be/uTzlnDq3SCE?t=2491) They're doing a lot of great things there that are related to AI implementation and I think becoming a good model for like an organization that's accepting of AI and having to do a lot of things that a lot of government agencies do as well as Urban Institute. So you want to build your practice. Again, you want to start small. Like first, you know, you want to play with it with again, data that are just maybe personal just so you can get used to it where you kind of know the context. So you want to pilot this one low risk workflow and with whatever guardrail that you have, try to figure out how much time it took and the quality of the outputs. So that way you can figure out whether or not you want to use it for this task or not. If the answer is no, don't give up at that point. Try to get in a couple of months because again, like Gemini, you know, before that was Bard, I just stopped using Bard because it was like totally useless in my opinion. And then Gemini came along and it was like a million times better and I'm now using Gemini a lot more.

[42:36](https://youtu.be/uTzlnDq3SCE?t=2556) Claude, like I think at one point gave me citations Same thing with Claude. and then it took citations away. So I stopped using it. So now at some point I would go back and look at Claude. They seem to be doing some interesting things. Then you want to refine your prompts. You know, so I talked about different levels, right? So maybe start with a basic prompt and then say, okay, well, what about this? What about that? And then you kind of expand your use, go to a second workflow and then share your knowledge. A lot of what I learned and the reason I am myself taking this class with Applied AI is for colleagues who are kind of at my same level to learn new things. And, you know, we all learn from each other. And it's really, really important in this world to share knowledge because everything is changing so fast on an almost daily basis. So remember the fundamentals. You're still accountable. Always have to verify these things and protect privacy. Those three things, no matter what it is that you do, you always have to keep in mind those fundamentals.

[43:43](https://youtu.be/uTzlnDq3SCE?t=2623) So here's some things you can do in the next couple of days. Find out what your agency's AI use policy is. And if there's not one, ask to start. There's a lot of like what we're calling shadow AI being used, even if there's no official policy. So it's good for you to get together with other people using AI so you can share information, tips, tricks and things like that. Identify your first pilot workflow. Set up your chosen AI tool with appropriate settings. I told you like with chat GPT and some of these other ones you can put restrictions in as to what it will do. There's also a button, by the way, that you can change so that your information is not shared with everybody in the world. And, you know, just spend like 30 minutes just trying to figure out that application. Then share your learning, you know, try to get a buddy, keep track of resources and then document it and kind of work with other folks.

[44:49](https://youtu.be/uTzlnDq3SCE?t=2689) So that is it. I am very pleased that I made it through all that content in 45 minutes, which was definitely a challenge. You have some practical workflows, examples, and, you know, ultimately. Our goal is to serve the public interest, and that means making government more efficient, effective, equitable. But always, these are my four Es, keeping in mind that ease of political acceptability, that people are supportive of how you're using these tools. You want to verify things, protect privacy, and maintain accountability. So on February 27th, I will be having office hours, where all I will do is answer questions. And if you have any comments, or the way you want to interact with me, LinkedIn is typically the best place. You can send me a message on there. So that is it.

**Agueda Quiroga**

[45:47](https://youtu.be/uTzlnDq3SCE?t=2747) Okay, Deborah, thank you so much. There are many and many questions.

**Deborah Stine**

[45:53](https://youtu.be/uTzlnDq3SCE?t=2753) Okay, so let's see how many we can cover in 15 minutes.

**Agueda Quiroga**

[45:56](https://youtu.be/uTzlnDq3SCE?t=2756) Yeah, I mean, the Q&A feature today wasn't working, and I promised everybody that I checked before the session, but who knows. So we have the questions for you into a Google document, so we can start reading now. Okay, so one question that has been repeated in several, I mean, formulations, but it's the same question. So any thoughts on crafting better prompts versus having a back-and-forth exploratory conversation versus hybrid, for example, back-and-forth, then a distillation-type prompt? I think this is about the process of getting a better prompt. Do you have any suggestions?

**Deborah Stine**

[46:43](https://youtu.be/uTzlnDq3SCE?t=2803) Yeah, so I think it really depends on the task. Okay, so for some tasks, you know, like if I'm just, say, looking for examples for, like, I talk to many, many different audiences, everything from high school students, to people like you, to people in manufacturing, to PhD students. I mean, I really cover the full gambit of people in the talks I do. And I, this is not, I mean, this is just one of, like, 40 different kinds of workshops I do. And if you look, many of them have been done here. So if I'm, if I just want, like, an example, like, hey, as I mentioned earlier, you know, like, I'm talking to Innovate US, I know that they want state and local government examples. I would just go there and ask the example very simply. But I'm not just saying, give me an example of the application of AI. That would be, that's too basic, right, for, like, how to synthesize data. I'll say, give me an example for state and local government. That's adding context, right? So context is, to me, the most important principle of your core thing is, you know, like, that's why I was showing you the different levels, right? So you can just, you know, if you just already know everything you want to be in a table or whatever, and you can put it in and do it, you can use a simple prompt. But if you wanted to, if you're going to have a higher level activity, then it's going to be different, right? It's going to be, you're going to need more of that context. And you might, if it's something really major, really have to enter that dialogue. So it all depends on what you're doing.

**Agueda Quiroga**

[48:28](https://youtu.be/uTzlnDq3SCE?t=2908) Thank you so much. I have another question that, in a sense, relates to a future work, I know you will give it to us, but we have not announced it yet. The question is, what is the trade-off between level of depth and detail in the prompting and taking up too

**Deborah Stine**

[48:49](https://youtu.be/uTzlnDq3SCE?t=2929) much space or AI processing tokens? Are you talking about, like, time? Is that

**Agueda Quiroga**

[48:53](https://youtu.be/uTzlnDq3SCE?t=2933) what's getting into? It's time and resources. I mean, which is like, you know, when to take and go manual and when to insist.

**Deborah Stine**

[49:02](https://youtu.be/uTzlnDq3SCE?t=2942) Yeah. Like I said, it really depends on the importance of the task and the level of risk that you have. I mean, I think that's, this is true. I also teach things like on risk analysis. And, you know, if it's just something, you know, that is, again, it's just for your internal team. If somebody makes a mistake, it's like not a big deal, then you don't need to spend a lot of time on it. But, you know, if you're presenting something to the senior leadership, like a board, or if you're going out to the public, if it's anything external, then you really, really want to spend more time on it. And so, you know, that benefit risk, you know, benefit cost ratio, which is also something as benefit cost analysis, that is kind of based on your personal time and what it is that you do. And it's something I think about every day. You know, like I told you, I was using this new gamma tool for PowerPoint. Well, it's got some good and bad things. It's good in that it will generate images for me. But I will tell you, every time it generates an image, I have to say, make the people in this more diverse, because it's all young white men, right?

[50:13](https://youtu.be/uTzlnDq3SCE?t=3013) This is very irritating to me. But, you know, that's the way it is. And I still think overall, I'm getting a better quality presentation from gamma than I am from PowerPoint, for the kind of work that I do. But if I'm just doing a presentation I've done a million times before, you know, then I'll just stick with my PowerPoint slide deck, even though it's kind of

**Agueda Quiroga**

[50:39](https://youtu.be/uTzlnDq3SCE?t=3039) boring. Okay, so there's a question exactly about the safety thing, which is always a concern among work. And give me a second, I need to locate. Okay, here we are. Sorry. This question has two parts. The first one is, this is careful comment. But to ask again, these are all unpaid versions. And how do we make sure that the data is confidential, and secure and not leaked? And then I have another question about free tools that could be used. Okay, well, okay. So first,

**Deborah Stine**

[51:25](https://youtu.be/uTzlnDq3SCE?t=3085) I would say that if you, if you are, you know, if you are at a state local government agency, and you're doing anything with anything that's data, then you should be kind of like an enterprise, an enterprise system where everything is kept within that organization. And your organization should pay or do whatever it is to do that activity. For most of the tools, like when you're experimenting, you can use like an unpaid version. But in exchange, like you'll always get something, right? You'll, you know, they want something from you. And if you don't want to share your information, then you just got to pay the money. Now, the, the good news is, is at the moment, a lot of these tools are, you know, you are cheap, right? I mean, that's how I can afford them. You know, there may be 20 $25 a month, things like that. And that's not probably going to last forever. But like, you know, if you go to like, just Google, you can just click a tool and like, go to AI mode, and you can ask Gemini questions. And it's really easy. And it's free. Okay. But if you have any concerns about confidentiality, whatever, then you're going to have to just make a decision. Either I'm either going to pay the money, you know, to protect that confidentiality, or I'm not going to use an AI tool, and I'm just going to do it the old fashioned way.

**Agueda Quiroga**

[53:02](https://youtu.be/uTzlnDq3SCE?t=3182) Okay, thank you so much. There's a very technical practical questions that says, how mechanically can a PDF document file say 200 similar letters be input to AI to generate a spreadsheet or data set of all database storage of content?

**Deborah Stine**

[53:23](https://youtu.be/uTzlnDq3SCE?t=3203) So the question is like, in terms of like uploading a PDF or whatever? Yeah, I'm not quite sure. It's really, again, if, if you have the paid version of say chat GPT, which I use, because that's, you know, the most common one. It is, I mean, I upload PDFs of documents like all the time, you know, and I will, I'll upload, like when I'm writing original analysis, I'll upload the PDF of my own document, and say, summarize this for me. And you know why? Because somebody else is going to summarize that document using, using an AI tool. So on this one, I was talking about, about WasteHeat from data centers, and I'm, this is for an organization called Reimagine Appalachia, that I'm, that I worked on for several months is going to come out in a couple weeks. Well, I'm going to go to an AI tool and say, hey, summarize this for me. And if I don't agree with that summary, then that means I'm not, I'm doing something wrong on my side. And as an originator of new content, something, something is going wrong. So, you know, I'll always go back and try to figure out like what's happening, you know, to, to what that is. But, you know, you can take three or four different sources. You can also, it's also good, you can do output that's different sources, like you can have it, you know, like produce an Excel file, or a Word document or a Google doc or whatever it is that you want.

**Agueda Quiroga**

[54:53](https://youtu.be/uTzlnDq3SCE?t=3293) Hey, do we have time for a final question? Maybe? Yeah. Yeah. One more. Yeah, here we go. Okay. This question says, do you have recommendations on prompts to use AI to develop SOPs from training materials and transcripts? For example, is there a persona you'd recommend? Question and sequence, question and sequence, etc?

**Deborah Stine**

[55:16](https://youtu.be/uTzlnDq3SCE?t=3316) Yeah, I mean, the standard sequence would work just fine in that case. Like what I, if I'm like doing, like, well, I did a workshop at the University of Virginia a couple of weeks ago. This is primarily for PhD students at universities in Virginia, on how to do policy analysis, and so forth. But I did a workshop at the University of Virginia on how to do policy analysis, and so forth. And, you know, then, in those cases, like, Chad GPT, or the other tools are really, really very useful, you know, in terms of doing stuff. So I will, like, say, upload my content, and I will say, create worksheets for me that you can use. I mean, I used to come up with those worksheets and things myself. But now I can ask Chad GPT, do it? And do I always change stuff? Sure. But as a first start, it makes sense. Then I will also, like, use it to say, okay, well, I want to come up with some interesting activities to do. Okay. So it is very easy to, sorry, my puppy is talking about stuff, I'm throwing a toy for her. So first, for standard operating procedures, things like that, what I find works best as a prompt is to put up something that I like, right? Get a model. Again, this goes back to context. So here's an example of a standard operating procedure that, like, maybe we're using in our organization, or maybe what we found online that you like. Here is something, here's some content. Sorry.

[56:58](https://youtu.be/uTzlnDq3SCE?t=3418) Okay. You get to meet the dog. So you stop sparking. So then I would say, okay, here's the content that I have, you know, for, like, whatever it is I want standard operating procedure from. Then here is an example of one, generate me something like this. Okay. And that will do, that will do it. This, by the way, her name, his name is Einstein Cubed. We just got him this weekend. He's about 10 weeks old. He's a Westie, a West Highland Terrier. So always give it context. That's the kind of the basic principle that you need to always let it know this is what I want. And if it doesn't give you what you want, then say, hey, this is not what I want. It's a very important vibe with my puppy here to tell it no, to set up guardrails. We have things all around the house. So the puppy does not get in trouble. Right? So that I think is a

**Agueda Quiroga**

[58:04](https://youtu.be/uTzlnDq3SCE?t=3484) good note to end on. Okay. So for our participants, there are more questions, but we do have a follow up ask me questions office hours on February 27th. So please make sure that you bring that questions and also other questions on your experiencing and practicing on workflows. So we can address there and learn from each other. It's gotten a pleasure as usual, Deborah. And thank you so much for all the thanks, everybody for attending till next time. Presentation and recording to all participants. And hope to see you soon. Next session is next Friday with David Smith from Colorado, who will host the first in the U.S. Prompting Lab opening office hours. Bring your questions on. So hope to see you soon. And thank you so much for attending. And thank you so much. Bye-bye.

[59:06](https://youtu.be/uTzlnDq3SCE?t=3546) Bye.

* * *

Original page: https://innovate-us.org/using-generative-ai-to-improve-everyday-workflows-for-public-professionals · Licence: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/), as stated on innovate-us.org
