From Prompts to Powerhouses: Building AI Workflows & Skills for Research Administration

Webinar series introduction (0:36)

Janior Valle: Okay, I think we can go ahead and get going here. So hello, welcome everybody. This is the second series of our webinar where we’re doing a little bit of knowledge transfer and guidance in terms of how to use these different AI systems, how they work; they’re all very different, even between ChatGPT, Claude, Gemini, they all have their quirks. The first session last month covered part one of the series where we shared a prompt template and covered some best practices in terms of how to put those prompts together. You were able to leave with those prompt templates and use them in your day-to-day. For part two, we have a QR code up front. I’m going to also go ahead and paste a link in the chat. You can scan it or you can click on the link that I’m going to put in chat here. That will give you the template that we’re going to follow today for creating the reusable workflows and skills, and you’ll be able to follow along on what we’ll be doing and take this back with you and use it yourself. Okay, so let’s go ahead and get started. My name is Janior Valle. I’m the chief technology officer here at Streamlyne. My background is as an engineer building a lot of these AI products, AI workflows, a lot of skills, which is what we’ll be covering today. With me today is Hector Segarra. He is part of Streamlyne as well. Hector, do you want to just give a brief intro to yourself?

Hector Segarra: Sure. Thank you, Jay. I am Hector Segarra. I’m a business analyst at Streamlyne. Prior to that, I worked in research administration in various areas. I started in grant writing, moved on to proposal review and submission, and then worked in various compliance areas. I am by nature a thinker. I enjoy working with technology. Even before we entered this AI age, I enjoyed creating little Python scripts to process information or putting things on Google and making it better to analyze. This is a really exciting time to be a research administrator, and we’ll get into that as we go into this webinar. Thank you, Jay. Back to you.

Janior Valle: Okay, thank you, Hector. As I mentioned the last time we had this session last month, this is a monthly series that we’re trying to have. After the session today, we’ll be announcing the next webinar in the series for August, which will include Trisha Callahan from Emory University in a better prompting session. Today we’re going to cover reusable workflows and skills. The goal here is the same as the first webinar. We want to give you some insights into how to use the systems, how to create these reusable workflows, and it’s going to be a practical hands-on session. I’ll cover a few things related to some knowledge transfer: how are these reusable workflows, what makes something a reusable workflow, how do you create these skills? These skills have different names depending on the AI product that you’re using. We’re going to be walking through this together. The goal is for you to leave here today with that template that you can put into use. A quick note on questions: if you have any questions, please put them in the chat. We have folks monitoring the chat and we’ll try to answer your question as best we can. We will have a dedicated 15-minute live Q&A section at the end of the webinar. If there are questions that come up during that time related to the topic at hand, I’ll try to answer those as we go through today. Let’s start off with some knowledge transfer and then we’ll dive into the hands-on portion. The first thing is a workflow is very different from an automation. When we talk about a workflow today, we mean a set of tasks that an AI model can take and use as a guide to perform some sort of task. With automation, you’re automating a workflow that has a repeatable set of steps and the output is always going to be the same. Unfortunately, with many of us, those outputs are not always the same; they do require a lot of human judgment before you turn in that piece of work, submit a proposal, or do anything. Today we’ll talk specifically about workflows. These are things that can be repeatable somewhat; they may have decisions you have to make, but the core of the workflow is a repeatable action that can potentially be reusable and includes some human element to review that output. One thing you’ll constantly hear me say in these sessions is you very much want to be in the loop: human in the loop. You don’t want to accept AI-generated outputs as is because very often these AI models don’t have the whole context about the task at hand, and so they will make up facts in some cases or give you wrong information. Be cognizant of that. Now, why do we want to build these workflows? I’ll give you my experience. Some of you may have run into this or not. I use AI on a pretty much daily basis throughout the course of my work. I’m an engineer, so I work with these systems very often. One thing I noticed as I worked with AI is I’m constantly repeating my prompts; I’m typing in the same information, the same prompts, the same workflow, the same steps, and eventually I asked myself why am I doing this again? I started saving those into a little prompt template notebook; I had an Obsidian app where I stored markdown files and I saved all my prompts in there. Eventually I started turning those into skills, which is what we’re covering today. Copying and pasting your prompt templates or writing those prompts over again takes a lot of time, especially if you want to do it right. That’s where these skills come into play. They let you natively incorporate that into your AI models so you don’t have to consistently copy and paste or pull and tweak them. We’ll show you what that looks like in a little bit. I like to do a quick poll in the room as a temperature check to see, is anybody out there using skills in the AI ecosystem? I think Gemini may call them gems, Copilot may call them custom agents. Feel free to type in the chat if you do. I’m just curious. All right, so the agenda: somebody said yes, we use gems. Great. We’re going to cover a lot today: a quick knowledge transfer on what’s the template for the prompt versus a template for the workflow skills, how to design and put this together, and then we’re going to do it live. I have four live examples that Hector will cover with us. We want to give that portion a lot of time, which may mean we only cover one use case. The templates we provide can be used for your future use cases or even your own personal use cases. Again, the concept is to learn that reusable…

Reusable skill template (10:47)

Janior Valle: So you have the template. If you didn’t scan it, let me just pull this up real quick so we can take a look at it on the screen. We had a QR code in the very beginning; if you didn’t scan it, it’s okay. It’ll come again at the very end and I’ll paste it in the chat one more time because I know a few folks joined us after that slide went through. So this is the template here. And honestly, it’s really short because most of the content is going to be dictated by your use case. So what’s in here? Let’s cover this briefly. First of all, we have a name.

Janior Valle: So this would be the name for the skill. A lot of these reusable workflows—that’s what the AI products call them, their skills—so that’s why I’m using that terminology as a heads up. The name you want to give it is actually really important. The name gives it contextual use so the model knows, okay, I have this skill available to me. Maybe my skill is the budget justification drafter. It’s pretty clear what that skill does. You very much want to use something that is contextually related to what that skill or workflow is doing.

Janior Valle: Now, the trigger might be a different concept to a lot of you, so I want to take some time to explain this. Prompts are the thing that you’re typing into that chat box. You pull up ChatGPT, it has a prompt, you type into that box, ChatGPT processes it and responds back to you. Skills can be invoked by the AI model directly without your guidance. What I mean by that is you can specify a trigger for when these skills should kick off. For example, using that budget justification draft skill name as an example, we can say use this skill whenever we’re talking about drafting a budget justification. It’s pretty clear exactly what that is and when it should trigger.

Janior Valle: Any time you talk to your AI model—whether that’s Copilot, Claude, ChatGPT, whatever the case is—when it detects that you’re talking to it about a budget justification, it will invoke the skill and go through the workflow. You can think of this like: “Hey, Chat, I’m working on a budget justification. Here’s all my documents. Can you help me out with this?” It just invokes the skill automatically and starts processing all your documents. You’re going to see this live with Hector pretty soon. This part is very important because often a lot of these AI products will only load the information that I’ve selected on the screen.

Janior Valle: It doesn’t know anything about the steps, what the review process is, or what the output is. It just knows initially what that skill is and what the trigger is. At the time that you’re talking to it about a budget justification, it’s like, oh, I have this skill here for budget justification. Let me trigger that, and then it gets all the instructions. This also applies for saving memories. A lot of these models can only retain so much in memory before you have to start a new chat or it loses past discussions.

Janior Valle: Inputs are going to be the artifacts or documents you provide to an AI model. For budget justification, for example, maybe you want to give information on your rates, on your proposal, who the people on the project are. You want to give it as much information as possible so it can do the task you’re asking of it. The process is how it’s going to do the work, and you’ll want to write this as a step-by-step: one, here’s the first step; two, here’s the second step. Output: indicate the final output. What is the artifact the model should provide? Is it an Excel spreadsheet? Is it a PDF? If it is a PDF, how do you want that PDF output? What’s the format? Do you want a summary at the very beginning? Do you want a source list of all the materials it pulled from at the end? You want to dictate what that output should look like.

Janior Valle: Then the review process, which is arguably the most important for us as humans: at what point is the AI model done and at what point do you take over for review? For example, one of my reviews is I tell my AI model to create my Excel spreadsheet or PDF as much as possible according to the skill and then provide me the list of sources it used in terms of the inputs, note if it ran into any issues during processing, and call out any additional information I should know about as part of my review process.

Janior Valle: All right. So again, that’s the template. Hector is going to do that demonstration live here shortly.

Janior Valle: I skipped around; I skipped the previous slide because we already covered it. In general, pick tasks that have a stable core. It may not be 100 percent repeatable, but maybe you can get 80 or 90 percent repeatable and then take over that last 10 or 20 percent. The main steps of a related workflow or skill—if repeatable—make it a good candidate for creating workflows or skills. One common question I get is, what’s a skill? Really, it’s just a standard operating procedure. Many SOPs have conditionals and branching and follow very similar logic to what a skill is.

Janior Valle: Hector is going to take over in about the next five minutes. He’s going to pull up the first use case, which I believe is the budget justification. Correct me if I’m wrong, Hector. He’s going to cover four distinct tools: ChatGPT, Copilot, Gemini, and Claude. The way skills work with a lot of these platforms is different. In the interest of time, we’ll do our best live, but know that these skills will work for any of these platforms. The way they interact with the skill may be a little different, but ultimately it’s a skill or a gem, whatever they call it in that platform, and you’ll be able to follow the template to build those reusable skills on those platforms.

Janior Valle: All right, Hector, I’m going to turn it over to you for that first live use case.

Live demonstration setup (19:28)

Hector Segarra: Thank you, Jay. Let me set up here real quick.

Hector Segarra: So let’s start by looking at that prompt. This is the program that Jay mentioned: it’s Obsidian, but you can do this in any text editor.

Hector Segarra: Before getting fully into it, I wanted to mention that I feel research administrators are among knowledge workers uniquely positioned to make good use of AI and particularly skills to assist them because the profession by nature requires systems thinking. A lot of what you do is taking complex regulatory and institutional requirements and turning that into auditable, repeatable steps. You have checklists and SOPs, and those are the building blocks from which you can draft those skills. Another thing you can do is have your AI agent, whichever one you’re using, collaborate with you in building that skill. You can give it the general outline and have it fill in the blanks, walk you through steps or requirements.

Hector Segarra: So let’s take a quick look at what would be the budget justification drafter skill that we’re working with. And of course, every institution is different. So you would make adjustments and personalize it for what you are intending to do. But the idea here is to give you a taste and a general idea of what the process is. As Jay mentioned, we have triggers. If you’re providing a project budget to the AI or when you use a phrase such as “draft a justification.” Just as important is when not to use it if you don’t want to trigger that template. For inputs, every modern AI is pretty good at realizing when it is missing a required file. You want that to be an initial evaluation and have a hard stop if it’s missing an important file or context. Describe the process in terms of how you want the AI to go about it. How would a human do this? Start with reading the budget, making sure that subtotals and totals are calculated correctly. Sometimes people mess with the Excel cells and the formulas aren’t auto-calculating, so that’s an important step. Then have it look at the template you’ll be using. Draft category by category, step by step, showing the calculation for every figure and citing our rates by class. An important element of prompting to reduce the possibility of hallucination is to ask the AI to refer back to its source documents. You probably don’t want to ask an AI to evaluate its confidence from 1 to 100; that’s not very reliable. But having it refer back to sections and page numbers, having it cite where it’s getting the information, increases your success rate. Look at the template, have it match the tone, preserve figures, and decide on rounding depending on how you want to calculate your budget. If there is an element not supported by the sources you provide, have the AI highlight that. Never fill in the blanks. Provide an excerpt of the sponsor guidelines so it can check the budget against them. If there is a conflict, flag it and identify it. Specify what the AI is going to produce. You can provide your files in their real-life formats like Word, Excel, and PDFs, and it can output files in the same format, but all of that adds overhead. Extracting text from a PDF is a lot more complicated for the AI than reading a text file. Take that into account: modern state-of-the-art AI will still produce the file, but it will take longer. There’s an element of speed and cost, and most people are working with subscriptions. Then there’s the review element, where the AI only makes a recommendation and you have the human in the loop to deal with that.

Skill structure options (26:47)

Janior Valle: What

Hector Segarra: Okay.

Janior Valle: One thing I want to mention,

Hector Segarra: Yeah.

Janior Valle: Hector, if we can go back just very briefly.

Hector Segarra: Sure.

Janior Valle: Sorry, I’m going to interject a lot today.

Hector Segarra: Feel free.

Janior Valle: One thing that I want to mention, because it is a common question that I get very often, is: do I have to have the name, the trigger, the inputs, the process, the output, and the review? Do I have to have those sections? No, you do not. This is what we recommend based on our use cases and using the AI models and working with a lot of different systems. If there’s a section you want to remove because it doesn’t apply to you or you want to replace it or add to it, by all means do that. AI products very much require that personalized touch. Think about the ways in which you yourself work, and incorporate that into these skills. I just wanted to throw that out there.

Hector Segarra: Thank you. All right. So let’s start with Copilot. As Jay mentioned, Copilot doesn’t name its functionality “skills” per se. You can have agents with a set of instructions, which is their way of structuring that work. Go to create a new agent. It will offer an interactive conversation to build it; I’ll skip that. You would add a name; let’s say “patient.” You always have to add a description. In the instructions is where you would put your skill file or skill information instructions. I noticed when working with this agent builder that it is constantly updating and making suggestions. Some of them are useful, some less so. It will try to take over your process if you let it. It also has a section on knowledge sources.

Janior Valle: Well, you pulling that up Hector so something

Hector Segarra: Mm-hmm.

Janior Valle: These are very different systems: ChatGPT, Copilot, Claude, Gemini. They all call these skills different things and interact with them differently. If you have access to several models, play around with them and the skills. My experience is that depending on the work, different models excel at different tasks. For example, I have some skills where I’ve asked the AI to learn my preferred text generation for emails and communication. I had my AI go through years of emails I’ve written so it could lock down my tone and how I communicate. Now I have a skill that whenever I need a draft communication, it can create a draft in the way I would normally write. Of course, you always need to tweak; you never want to accept it out of the box. That tends to work better for me with Claude from Anthropic.

AI model preferences (31:38)

Janior Valle: I don’t really care. I don’t care so much about how ChatGPT or Copilot write it. But when it comes to, for example, anything visual that I generate like PowerPoint or visualizations or things like that, ChatGPT doesn’t really do that well. So I use Claude. Where I do use ChatGPT is, for example, if I have to do a lot of research, like if I’m looking into how something works or I’m working on a new engineering task and I have to do heavy research into some sort of task or something, I like ChatGPT for that. I very much like the way that it does that research. The reason I’m telling you this is because not every AI model is the same and you really want to try out—if you have access to multiple models—try out the different models, see what works best for your task, and I think you’ll find that there’s probably one that’ll be your favorite for any given task. Just wanted to throw that out there.

Hector Segarra: Thank you, Jay. One item here that it notices right away is that the prompting mentioned skills and it wants it to be agent focused. So just a heads up: when you’re working with Copilot specifically, it likes its particular phrasing. What I wanted to show you was an anatomy of different elements that can go into a skill. You have your prompt here, which is the saved skill markdown file. You can have reference documents, assets, etc. One of the limitations, at least with Copilot on the web in this subscription, is that its knowledge sources are URLs. It can only take web addresses. We’ll work around that by providing those files in the conversation itself. We do want it to be able to create documents. We don’t need the image creation. It’s a good idea to turn it off if you don’t need it. It reduces what’s getting transmitted to the agent as a tool or in the system prompt that you don’t see. Then it has a section with suggested prompts that you can create that’s just for one click, like “draft a budget for me” or something like that. See it. That should save it. If it passes the checks, then you can create it. That’s a private agent. You can start chatting with it. All right. So here I’m going to upload our files and let me try something real quick that I know is going to error out but I want to show you. If you try to add more than three files, it gives you an error. We’re running up against that limitation as well, at least the way that this skill is designed. It uses additional sources, but we can see guidelines, information, template at these and it’s going to tell you that it’s missing the budget itself. As we did that, we provided the instructions, we provided a template that it would be basing its response on, an excerpt of the sponsor guidelines, and your institutional facts document that has your current rates and other related information. So here is the text output. We’ll be seeing this or versions of this repeatedly here, but you have your senior personnel section, other personnel, undergraduate fringe, equipment; it marked correction required. Highlighting that sponsor requirements and institutional policy define equipment, that changed, but this is $5,000 so this should actually be materials and supplies. We get to the flag section. We instructed that if there are any issues to add them as flags. So you have, again, the equipment—there should be materials. The participant supports were incorrectly included. Missing information on department. Missing a title. Additional input for the narrative justifications. You’ll see that it’s citing back to the source document that was provided. That’s how Copilot does things in general. We could ask it to output this as a Word document. Let’s see if it doesn’t take too long, otherwise we’ll move on to the next one. Let’s peek behind the scenes at what it’s doing. It’s using Python to create that Word document. There it is. Copilot has its limitations. All right, so moving on to Gemini.

Gemini and gems (40:07)

Hector Segarra: Gemini again has its own skill system, so to speak. It uses gems. We would go to gems here. If this doesn’t show up in your menu, you can Google to find the URL for gems. It didn’t show up for me until I created the first one. When you’re here, skip over the Beaker icon of the new GEM; that’s their labs functionality. That’s where Google is testing different features that might or might not become full-fledged functionality. What they’re testing out here is a tool to build web apps by conversation with an AI, which is kind of nifty, but it doesn’t produce the output that we want. It does have a more structured interface so it’ll tell you, “Upload your budget now, upload your institutional information,” etc., but down here is the traditional gems function. So let’s see.

Hector Segarra: And I’m adding version two because I already had an initial one preparing for our webinar. All right. Seeing the description as well, then let’s pull up the instructions. And you’ll notice that in this suggestion, it asks you to give the AI a role. In my experience and from what I’ve read about recent AI models, this field changes every three months, but it’s not necessary. You’ll notice it’s not in our template prompt. You can do it. It would likely not hurt, but it’s not really required. None of these really apply, but if you are creating a skill to do a different function, these are the sort of add-ons that are available. And then Gemini does allow you to add proper reference files. So let’s add the institutional information, the template. Pretty much. So you’ll want to add to its knowledge base whatever documents are not going to be changing on a per-conversation basis. Your budget and your sponsor guidelines would be changing. All right. Let me save that. Mm-hmm. So again, if your institution has an agreement with Google, if you’re using the educational version, then I know that there are some exclusions so it doesn’t use your information for the training data. But if it’s your own personal Gmail account, it’s pretty hard to opt out of that. What Google does generally is that it says I will let you opt out, but then I’m going to remove functionality from the product that you’re using. I haven’t checked recently, but for Gemini, at least I know that if you opt out, then it would not keep your conversation history, and that sort of makes it more of a pain to use it.

Janior Valle: Yeah, we always say just double check with your IT or legal just to be safe.

Hector Segarra: Definitely. All right, so here’s Gemini, let’s upload my budget, my guidelines.

Janior Valle: Mm-hmm.

Hector Segarra: All right, so it’s processing that. Let’s see if it starts. Google also has a nifty trick here that I want to show you. Let me see if it starts. You can see here in the show code that it’s building a Python script to read the budget and it’s also going to use that to check the calculations. A problem: if you keep up with this field, and particularly its use in research administration, a few years ago the evaluation was that AI was not quite there in terms of the tooling. Part of it is that at that point, AI was really bad at math. But as you can see, there’s been improvements, and as the agent is provided additional tools, it can use programmatic development options to review those files.

ChatGPT, Claude, and desktop agents (47:50)

Hector Segarra: So a problem a few years ago was AI being bad at math, but it can now write you a script or an Excel file that will do the calculation and you want to use that instead of letting it use its discretion. Okay, so we have our budget justification. Let’s look at the output here first. It identified the equipment threshold violation. It identified the F&A base miscalculation. And then let’s take a look at this document. All right. We have here highlighted some fields that this was expected. Copilot kind of missed this one: this PI input for role description, again with the graduate research assistant. It used the institutional information provided. It added a note on the misclassification here, additional input, and then the flags: needs department and title for the header, missing information on roles, travel justification, a bitmap classification and F&A-based calculation. So it caught all the expected errors that were included in the budget, which, by the way, before moving forward, let me mention, when you’re building a skill you have your process where you break it down into steps, you write it, you upload it to your agent, whichever agent you’re using, and then it’s good to have a few test cases, especially where you know what you expected to find to calibrate and see if there’s anything that you need to go back and change. All right, so that’s Gemini. Then we have moving on to ChatGPT and Claude. These work a little differently and they’re the ones that are at the forefront of skill usage and agentic use. While we’re working here with the web versions, both of these companies have desktop applications, and in general as Jay mentioned, you use what is authorized and check with your IT, but Codex and Claude Cowork are more capable in the sense that they reside in your computer, they have access to your files, and the interface for creating and using skills is a bit more upfront and highlighted.

Janior Valle: I just want to clarify that when Hector said codecs, that’s the name of the app, but they have two different apps now. It’s codecs and then the ChatGPT app. I think they’re trying to combine them. So really when we say codecs, it’s just the ChatGPT desktop application. There are a lot more feature rich, so highly recommend.

Hector Segarra: Definitely. All right, so let’s hear.

Adding AI skills (52:04)

Hector Segarra: So this is what ChatGPT uses. It calls it the skill creator workflow. So that is a skill in and of itself that it has, and you can invoke that and create your skill. It’ll give you a brief overview of what it should have. Let’s just add our files. Let me see. So we’re selecting the institutional information, our template, and the instructions.

Janior Valle: And while this is running, one thing I want to mention is you can get very creative with these skills. So what I’ve seen in the past, and I’ll use a budget justification example, just because that’s what we have on the screen. I’ve seen before where research offices will share skills between each other, amongst our team. And you can get really creative with that. What I mean by that is right now, Hector is providing the documents up front with the skill, right? But maybe you want to create a skill and you want to share that with a group or your team. You can include in the skill a kind of Q&A portion. So you’re not going to know what the budget is or the proposal, the PI or all that information up front, and so you can create the skill in a way where you adopt the best practices within your research unit. Think about the things that are required for a budget justification. You can include that in your skill. And then as part of the Q&A, maybe when a member of the team says, hey, I want to create a budget justification, it kicks off that Q&A process where the model can say, OK, I see you want to create a budget justification. Do you have what supporting documentation do you have? Who’s the PI? And then you can include what are the triage steps or the intake steps that you need, that the model needs in order to perform that unit of work.

Hector Segarra: Right. So it’s working its way through it. One thing you’ll notice, depending on the model that you choose, is that ChatGPT and Claude can take longer than Copilot and Gemini, but it is a more advanced model. So one thing, it’s fine that there’s already a skill there that I ran through. I had provided an example Python script that it could use for this and it went a bit deeper and found that there were some items that were hard-coded onto that script. Again, you can just conversationally guide it to make it more flexible, share the revised script, and when it’s finished, it installs its skill so you can use it on the web app as well. Let me see if this one finished. So while that works, let me hop to Claude over for a moment. So another example of how you can add a skill, if you’re using it in another AI agent or you wrote it, put it together, you can turn that into a zip file. If you go here to settings. Skills. Writing settings. Skills. Add. You can either add it conversationally, write the instructions, or upload that package. Let me find that here in my desktop. And that’s a new skill for Claude. You would, do you said you type a forward slash and it lists the skills it has? Budget justification. Budget guidelines, it should already have the institutional information and the templates in the skill. So, ChatGPT Soul is still taking a while there. So while we’re waiting there, Jay, I don’t know if you want to maybe tell them a bit about the different models and when it’s a good idea to pick one over the other.

Choosing AI models (58:26)

Janior Valle: Yeah. So there’s kind of what we talked about in the first session. And just to recap, there’s different models. They call them modalities. So you have your image models that can produce images. You have your text models that we’re all familiar with, those are ChatGPT, Claude, Copilot, those are your text generation models. There’s also vision models that can actually see images. In general, there’s a lot of different AIs that perform or excel at related tasks. I mentioned this at the beginning, but going back to the examples that I gave, when it comes to anything that has to do with communication or emails or anything that somebody is going to read, I very much like Claude. I just like its approach a little bit better. But that’s just me. Someone else may prefer ChatGPT. When it comes to anything like image generation, ChatGPT is actually really good now for image generation and so a lot of these models are now starting to become or support multiple modalities so ChatGPT can not just do text generation anymore, it can do image generation and it can also do vision so it can actually see the image. If you give it a screenshot and say look at the top right corner what’s there it can actually see what’s in that image. Same thing with Claude. It’s a multimodality agent, so it can support text, image, vision. I take that back. Image with Claude is actually horrible. I recommend if you’re going to generate any images, I’ve generated some images of my dog in the last session. Those are from ChatGPT. They look very good. They look almost like someone actually took them. Claude does not do great at images, but he can do vision and he can do text. I don’t think we’re doing any coding here, but ChatGPT and Claude both do really well with that. Honestly, in general, a lot of this is very subjective, except maybe the Claude image generation one, it really is terrible. But that’s why I say if you have access to multiple models, definitely try them out and try them out for different tasks because you’re going to realize which one you like a little bit more. My chat GPT is my daily driver. I use it for everything. I have it linked to AI meeting notes, my emails. I have it doing research for me, that sort of thing, so yeah, just definitely recommend trying it out and seeing what works best for you.

Hector Segarra: Yeah, so let me close this out. Let me show you a sample output. It’s essentially the same process, except that I completed prior to that. Ideally, you could queue up a few of these and leave them running to work in parallel to save time so you don’t have to stare at the screen and wait. You’ll notice here with Claude, I did not specify that I wanted a Word document and it used Markdown for its output, which again you can download, you can copy and paste this into a Word document. One thing that happens here is that it shaved time off the total it took to create it because it was just adding text to a file rather than dealing with creating the Word file and updating it.

Hector Segarra: So you’ll see in general it was more detailed in terms of what it flagged. It found all the items that we expected, but it had a more step-by-step itemized breakdown of it. After that, I asked it to write that same content as a Word file. I do like that Claude and ChatGPT have these built-in previews, which you don’t get with the other models. Here’s the output. The formatting is a bit more elaborate. It added different colors. It added a call-out box and a separate page with the flagged items. That’s Claude. ChatGPT is still going to take a while. It does pretty thorough testing of the skill before it makes it available, which is a good thing if you have the time for it. In general, when it’s done, you provide the files and call the skill by the name you gave it. It took six minutes. I think this was Luna, which is a smaller, faster model. And there we go.

Budget review and templates (1:04:07)

Hector Segarra: To see it structured a bit differently — this is all a matter of how you want the template to look as you provide it. One thing I liked is that it called out the specific cells where the items or miscalculations are, so that is helpful and would certainly save time when going over and reviewing it. And that’s budget justifications in general. Back to you, Jay.

Janior Valle: Thank you, Hector. One thing that I’ll mention while you’re still there, Hector, is that it’s actually a relatively new feature. I tried to find it while you were talking on the web versions, but I could only find it on the desktop versions. If you need a little bit of help in terms of creating these skills, I definitely recommend the ChatGPT app if you have ChatGPT. If you have Anthropic or Claude, there’s also a Claude desktop app. Copilot, I don’t know that it has this feature, but for Claude and ChatGPT, they have a record-your-own-skill feature where you can start that recording and walk through what the task is, then stop the recording. Based on that, it’ll generate a skill for you that you can then take and tweak. So I highly recommend that. If you have any of those subscriptions or licenses available, the desktop app is probably a better bet than the web version because it has more features and functionality.

Hector Segarra: If you’re able to use it, I would certainly recommend it. Before you continue, Jay, this skill creation finished and there was a little pop-up that said if you have the same account on your phone, you’ll get a notification there. That’s helpful if you need to leave it running and walk away. It’ll let you know when it’s done. Please go ahead, Jay.

Janior Valle: Yeah, no worries. This is a good segue into our Q&A. There’s already a question in the chat, Hector. You’re probably better equipped for this one: What are other research admin skills that people are using?

Hector Segarra: Oh, sure. Let me pull that up. Let me share my screen again. So that actually was put together as a sort of motivation when we finished and what else you could do. We went through the literature so you don’t have to. Here are some ideas from different institutions and companies that have been applying AI to research administration. Pre-award: funding opportunity announcement snapshot, justification drafter, funding digests. If you have a routing form and you’re doing that manually, give us a call, but otherwise you can have it put that together. Initial triage is interesting: proposal drafts against screening questions to flag. Reward, funding opportunities: try to identify triggers and flags that would kick off compliance requirements, build your submission timeline, subaward package review, read award letters and agreements and pull the different fields into your setup form, contract or agreement redlining. Digest what reports you have coming up and draft the email reminders to send to the responsible party. Flag expenditure review items and put together progress reports. In compliance, you can have a quick check against required elements or IRB and IACUC items. You can provide a list into the skill, a checklist of common protocol errors at your institution, to check specifically for those and flag them. Draft informed consent forms — this is interesting because this was actually a study where the AI-assisted writing ranked higher in legibility and understandability than the regular consent form. In export controls, there are advancements in restricted party screening, gathering evidence on possible false positives, automatic reminders, and running regulatory changes against your SOPs and institutional documents so you start from a briefing and not from scratch. In general: policy and SOP Q&A, training material drafting, working on your templates. I found this one interesting because of what the AI thinks is the internal deadline compliance, flagging anything under 60 days. Draft communications and prepare reports, draft reports and briefings for your leadership.

PDF editing use case (1:11:02)

Janior Valle: One that I see commonly, from interacting with different institutions, is that AI is pretty good at editing PDFs. It still struggles with reading them sometimes, which is kind of funny, but it’s really good at editing them. In research administration we have a lot of forms, so you can create a skill to build templated forms with your letterhead, your logos, your branding, and create the forms with editable PDF fields. I’ve seen that before and wanted to share it.

Janior Valle: In general, one thing I recommend is that the folks who adopt these workflows best are not the ones who try to create all these skills up front. They inject them into their normal day-to-day workflow. It does require you to step away and think about what you did during the day: what are the tedious pain points that are taking a long time and are manual? Those are the candidates for skills, and you identify them during the course of your day-to-day.

Janior Valle: Realize that and start automating those things via skills. Over time you’ll notice that one day you have 20 or 30 skills that you’re using on a day-to-day basis. At least that’s been my experience and that’s been the experience I’ve seen at other institutions as well. So I just wanted to share. We are wrapping up. We have about one minute left. Are there any other questions or comments or anything? Please feel free to put them in the chat. While we wait for that, I will say we are going to share the recording. After the call ends, we’ll let the recording buffer and then we’ll distribute it. So last call for any questions. All right. Well, I want to thank everyone for taking the time to attend today. I very much appreciate you joining us. One more reminder: we’re going to have another session. It’s going to be August 26. It will be at the same time as this webinar. We’ll have a guest speaker. Her name is Trisha Callahan. I’m sure a lot of you may recognize that name from Emory University and Kira, of course. We will send out those details on the listserv and join the normal communication channels. Thank you so much.

Hector Segarra: Thank you.

Janior Valle: Take care.