In short
How to choose the right AI approach—simple chat prompts, structured prompts, reusable “projects” (saved instructions + files), or autonomous “agents” (tool-using systems that act toward a goal). The episode emphasizes starting with the simplest option that solves the task and moving up only when needed.
Guests
Corey Knowles and Grant Harvey (hosts of Neuron Podcast). No other guests are mentioned.
Key claims
- Prompts are what you send to the AI; ambiguity in “fix this” requires more context (e.g., paste code or provide the document).
- Structured prompts (headings, labeled sections, role/task/context/output format) improve quality for multi-step or formatted outputs.
- Projects (folder-based) store instructions and uploaded files for repeatable workflows, improving consistency and speed.
- Agents provide autonomy and dynamic tool use but reduce control, can be expensive, and require detailed “definition of done” plus sufficient tool access to succeed reliably.
Notable examples
- Fixing an essay about the fall of the Roman Empire; fixing code from a screenshot.
- A “podcast helper” project generating show notes and drafting emails.
- Using agents for morning news headline gathering (scouring the last 12 hours).
- Agent use with connectors/MCP (e.g., Google Drive, Gmail, Dropbox; potentially generating an n8n workflow).
Written by AI. May contain mistakes. Listen to the episode to check what was said.
Chapters
Tap a time to open that second in VONavigating AI Options
0:23 to 1:10
Exploring the different approaches to using AI: prompts, projects, and agents.
“I'm Corey Knowles, and as always, joined by Grant Harvey.”
Understanding Basic Prompts
1:10 to 2:26
Defining and exemplifying simple prompts used in AI interactions.
“So to kick us off, Grant, how would you define prompt?”
The Role of Context in Prompts
2:26 to 3:55
Discussing the importance of context in crafting effective prompts for AI.
“And you paste the whole essay that you already wrote into it.”
Structured Prompts for Complex Tasks
3:55 to 7:21
Introducing structured prompts and how they enhance AI output for intricate tasks.
“You know, the truth is the vast majority of like, you know, most of my interactions with AI are on, you know, simple needs.”
Structured Prompts for Complex Tasks
7:25 to 8:23
Introducing structured prompts and how they enhance AI output for intricate tasks.
“When was the last time your AI app surprised you with a hallucination, policy violation, or just plain weird output.”
Advanced Prompt Structures and Outputs
8:31 to 14:00
Exploring advanced structures for prompts and specifying desired outputs in AI tasks.
“Tell them the neurons sent you, and start building some AI you can actually trust.”
Leveraging AI for Customized Reports
14:00 to 15:00
Learn how to instruct AI to generate reports that match your specific needs.
“And, you know, you don't have to go that in depth.”
Organizing Tasks with AI Projects
15:00 to 18:10
Discover the benefits of creating structured projects for repetitive tasks using AI.
“Or maybe you would do that with a project.”
Refining AI Prompts for Efficiency
18:10 to 26:00
Explore techniques for optimizing AI prompts to improve output quality and consistency.
“But then you can do this really handy thing called instructions.”
Understanding AI Agents and Their Usage
26:00 to 28:00
Gain insights into the functionality and decision-making capabilities of AI agents.
“Agents is what everyone is talking about these days.”
Show all 23 chapters
Deciding When to Use Agents
28:00 to 29:25
Learn how to assess when to use agents for tasks versus completing them yourself.
“And come back to me when it's finished and you have something for me to take.”
Using ChatGPT Agents for News Gathering
29:25 to 31:41
Discover how ChatGPT agents can enhance your news gathering process.
“Do you have anything else to add to that?”
Understanding Agent Limitations
31:41 to 34:19
Explore the limitations of agents, including access and hallucination risks.
“So let's say, you know, you have access to your Google Drive, you have access to all of your logins for all the websites you're logged into, you know, the Chachapiti agent.”
Enhancing Prompts for Agents
34:19 to 36:24
Learn how to structure prompts effectively for agent tasks.
“prompt structure a little bit when you use it that way.”
Connecting Tools and Agents
36:24 to 40:09
Understand how to connect various tools with agents for better task execution.
“are all great theoretically you could also tell it to drop its output there couldn't you What do you mean by that?”
Choosing the Right AI Tool
40:09 to 42:15
Find out how to select the appropriate AI tool for specific tasks.
“Every time you download a mobile game and pop in and connect it to your Google, that's a little different, but at its core, that's what you're looking to do here, is connect your Google to this.”
Choosing the Right Tool for the Job
42:15 to 45:46
Learn how to select the simplest AI tool to address your needs effectively.
“If a simple prompt gets you the results you need, don't go build an agent about it.”
Understanding Structured Prompts
45:46 to 50:52
Discover when to use structured prompts for complex tasks and projects.
“So, multi-part task would be you need the agent.”
Utilizing Agents for Dynamic Tasks
50:52 to 56:01
Explore how agents can handle autonomous tasks and the benefits of using them.
“Other examples we talked about last week in Chat2BT agent.”
Verifying AI Outputs
56:01 to 57:50
Learn how to verify the information provided by AI tools to ensure accuracy.
“And I think it's important to be just cognizant of that.”
Data Privacy in AI Usage
57:51 to 59:36
Understand the importance of data privacy when using AI tools and sharing information.
“Last thing related to that is always consider your data privacy and safety.”
Personal AI Usage Breakdown
59:37 to 1:02:43
Explore how to categorize your daily AI tasks between prompts, projects, and agents.
“I pitched this idea to you, Corey, which is essentially maybe we need to think of creating a clean room equivalent, but for AI.”
Managing AI Agent Credits
1:02:44 to 1:06:08
Discuss the management of AI agent usage and subscription credits for optimal efficiency.
“I would say for me, just day-to-day in-and-out prompting makes up probably a third.”
Transcript
Automatic transcript. May contain errors.0:00Should you just ask ChatGPT a question, write a structured prompt, build a project, build a project, or deploy an agent? We're going to help you figure out which one you need today.
0:22Welcome, humans, to this week's Neuron Podcast. I'm Corey Knowles, and as always, joined by Grant Harvey. Hey, Grant. What's up? So, over the past two and a half years, using AI has gotten a little more complicated. You have more options than ever, and they're actually quite different from one to the next. And we're not just talking about companies. We're talking about models and ways to approach models. Each have their strengths and weaknesses. So regardless of which company's tools you prefer, today we're going to explain how to pick between an agent, a project, a structured prompt, or just typing a bunch of words at it and hitting enter and moving on because each one of those kind of has its place.
1:10So to kick us off, Grant, how would you define prompt? Well, the prompt is what you input into the AI chat window. Or if you're working with an API, it's what you send over the wire into the cloud and what is then used to give you an output, simply put. Yeah. Well, then what about, then let's talk about how we would describe, say, a simple prompt. Okay. I think it's good to have these definitions clear up front. So as we go forward, it all makes sense to anyone who's listening. So the most basic possible prompt you could do in AI is basically asking it a question, like a single question with a single answer.
1:59Fix this. Well, yeah, yeah, exactly. But even simpler than that, right? Using it like a Google search, like saying, what is, please explain to me the fall of the Roman Empire. You know, you're not going to get a simple answer, but you know, that's a simple question, right? Yeah. Yeah. Now, what you just described is basically giving it context along with your prompt. So in the example that you said, let's say you have something you wrote, like your bad essay about the fall of the Roman Empire. And you say, fix this. And you paste the whole essay that you already wrote into it. Now, you're leaving a lot.
2:38It's very vague, isn't it? You're leaving a lot up to the AI to decide what fix this means. And they're pretty smart, but they're not mind readers. They're probabilistic predictors. So they're going to try and fix it based on what's in their training data and what they've seen is like a good essay on the fall of the Roman Empire. But let's say you were to take a screenshot of like an error in some code that you're writing and you say, fix this. Well, it's going to be able to use reasoning, depending on what model you use, which we'll get into that, to think about the error and say, okay, here's what I need to fix this error.
3:15Now, it probably needs to know your code so that it can actually fix the error in the screenshot. So you might also need to paste in the code. Or if you're working in a coding app like Cursor, which we can get into that, but that's not really the subject of this episode. It will already have that context. But that's essentially how it works. You give it a bunch of data, essentially, and a very specific goal at the top. And it then goes to work crunching the numbers to try and figure out, okay, let me try and solve this problem for this person. And for most of your day-to-day tasks, it's probably enough.
3:54It's true. You know, the truth is the vast majority of like, you know, most of my interactions with AI are on, you know, simple needs. Like, can you find me an article on this? Can you, you know, sort this out, copy edit this document for me, whatever the case is, you know. And that kind of stuff, it doesn't have to be anything fancy for it. It can be super simple. um but that brings us to our next question which is structured prompts which is a very different way of approaching you know when you get into you know and you'll think of this more with research models like when 01 and 03 dropped uh when claude starts dropping research tools at gemini um one of the things you start finding is like here are these prompt patterns that are that are really useful and they're very structured.
4:48Not just like to receive structured data, which is a phrase you hear a lot from technical people talking about AI. And this is actually more simple than that. But sometimes you have a problem that is actually a series of problems. Like instead of copy edit this document, it might be, hey, I need you to go through this document, edit it, check it for accuracy, find some good quotes to include in it, and then I want you to finish it all up, put it in a fresh doc file for me to download. When you have needs like that, you don't come at it and, I mean, you can absolutely just write that sentence in there.
5:30And there's a 6 to 7 out of 10 chance it'll get it and it'll be just fine. But, you know, when you really want to maximize the quality of what you're getting, when you really want to maximize the amount of compute that's being spent on solving your problem versus deciphering your words, because that's a thing, you know, that we should all know about. You know, you want to get into structured prompts. And these will use, they might use bullet points or numbered lists. You might use headings for multi-part questions if need be. essentially what you're doing is looking to cut back on the ambiguous nature of your words and put them into a clear format that the model can read and know exactly what you want at a glance versus having to translate corey ease as it sometimes is left to do and uh yep so you know some of these you know you'll see a lot of this in you'll see a lot of this in the work engineers do building out agents and stuff.
6:34Like, you know, typically their prompts, you'll see them. They love writing them in XML, for example, which is not something you have to do. Or markdown, yeah. But people have shown good results from that. For me, it's more about labeling things within the prompts. I'll be like, we're going to work through this problem together. I've got some examples. I've got these things below that you can go through. And then I'll go down and I'll type in all caps with a colon, the example. paste the example. The questions, type the questions. The document, the document, I'll stick it right there, just in a way that more than anything, it's about clear labeling of what you need, what you have for it, making sure you've provided ample context and a good structure, I would say.
7:22Hey listeners, it's Corey. Quick question. When was the last time your AI app surprised you with a hallucination, policy violation, or just plain weird output. If your answer is yesterday, you're not alone. That's why I've been digging into Galileo, the end-to-end platform that helps teams evaluate, observe, and frankly even guardrail their AI in real time. So think of Galileo as like a co-pilot for your LOMs. It stress tests every prompt, every model variation with more than 20 out-of-the-box metrics plus the option to add your own. Then its proprietary Luna models run those evaluations in under 200 milliseconds, catching trouble before your users ever even see it.
8:10The result? Companies like HP, Twilio, Comcast are already shipping AI features with confidence instead of crossed fingers. So if you're ready to move from it usually works to it always ships safely, Head to Galileo.ai, G-A-L-I-L-E-O.ai, and book a demo today. Tell them the neurons sent you, and start building some AI you can actually trust. Yeah, and there's actually prompts, I guess you could call them frameworks, for working with these structures. Like, if you look at the example we gave at the beginning, whether that's single question, no context, or request fix this with context below it.
8:51You can think of that prompt structure as task, fix this, context, your copy-pasted document, right? A more rigid structure you can follow, which a lot of people, as these models get smarter, you don't really need as much of this tooling, I think, unless you're doing something really complicated, which we'll get into later. But a really rigid structure you can follow is You start with a role where you focus the attention of the model's massive brain to focus on one specific section of content. So you would say, role, your role is you are a neuroscientist who is the best in your field. You read PubMed and you only answer in PubMed, you know, responses, like cited responses, something like that, where you give it a very, very specific role to focus its attention on that section of its data set.
9:43So you could do the role. Then you would do the task or goal. Let's do goal first. My goal is I want to look at this cancer data or look at this study data and figure out what it means for this cancer medicine that I'm researching. right yeah and then underneath that task your task is to go through you know first off read the data and analyze it second of all put all of your findings in a bullet point list and third of all then you know based on those findings um compare it to this other document and you know do some cross-examination task there yeah um and analysis you wouldn't want to do that with like a really basic model you'd want to do that with one of the thinking ones and we'll get into that in a minute.
10:29And then underneath that, you would have the context. And the context is, you know, all of the study data, which to your point, you would label, you know, like this is the study that you need to review. And then underneath that, you would say, this is, you know, the details on my cancer drug that I'm developing. This is a really complicated topic. Yeah. I don't think you're going to be doing this, but, you know, that's the idea, right? That's right. Yeah. Yeah. The whole idea there is just, you know, making sure that you're maximizing the odds of getting a good response for what you need. By focusing the attention, critically.
11:04By focusing the attention and being clear. And yeah, I mean, like, you know, if what you want is to, you know, I don't know, make a meme of an, you know, an apple and a dog, you don't need that. But if what you want is to, you know, generate something that's work-related, that's analyzing a report that's, you know, dealing in scientific data, like Grant's example there, you know, this is one of the tools in your toolkit that's really handy to whip out when you need to. Yeah. And speaking of tools, you could also distinguish when to use tools as part of your prompt as well. But that's when you're getting more into more complicated workflows.
11:44But to your point with the meme, let's say you want to call OpenAI's image generator or web search. Those are two tools that are built into the application. You could say, you know, help me generate a meme with image generator. Now, I think at this point, they're smart enough where you say, help me generate a meme. It'll automatically route you to the image generator, at least on chat GPT. But some other systems, like if you're working with Anthropic or Google, it's usually good to call out the tool that you want. So like for the example with Anthropic, You could say, hey, look, I need all the information on this news story today or I need all the latest news on, you know, cryptocurrency today and use Web Search and go out and give me the latest news as of, you know, today's date, July 29th, 2025.
12:32Now, I won't give you everything because they're limited, but it will it will then go out and it will use Web Search based on your request. That's so awesome. Yeah. And it's, you know, it's doable. It's not that hard. It requires no technical knowledge. All it requires you to do is just really kind of think through your problem and what your process would be if you were going to work on it and be aware of what might make sense to you and won't for a computer. yeah or another way that i've heard it put which makes the most sense to me is imagine you're trying to explain to a new intern or just anyone exactly how to do your job and do a task what would you need to say you would need to tell the person their role so you need to say okay this is your job you are an intern here you are going to research all this stuff so you are the best intern in the world we have the best faith in you like you're going to crush this research task okay your task is to go out and research these 50 companies we need to know everything about them um you know Our goal with this research is to try and figure out which of these companies to invest in.
13:37And here's all the context of what we know about them so far and what we do and don't like about investing. So here's all that. Keep that in mind as you do this research. And yeah, give us a report by the end of the day. That was one thing we forgot actually in the structure was it's also really good to determine the output format of what you want. Yes. Yes, that's a good point. Yeah. Yeah. Like I want this in a in a report with an executive summary and sections breaking down this and this and this. And, you know, you don't have to go that in depth. But and the other thing is, after you've done a few of these, like with Chagiput, it knows that's how you like a report.
14:15Yeah. Like, you know, now when I ask for a report, it has an executive summary. It has the sections that I want normally without me having to babysit it, you know, and tell it every time. You know, so that's pretty amazing. I personally, especially if you're using something else like Anthropics Claude, would not trust it to do that. Instead, what I would do is I would and we'll get into exactly how to do this in a minute. But I would I would say this is the exact style of report I want or even give it a previous example. Like say, like, please make me a report in the style of this example and paste in a report that you like that it's done before that, you know, your typical report format.
14:58Actually give it that context. Or maybe you would do that with a project. Correct. Grant, you want to talk a little about projects and explain what is a project and, you know, structured workflow like that? Yeah, let's go ahead and pull up the ChatGPT window here and we can put some of this in context. So, all right. So as you can see here, this is your run of the mill ChatGPT window. There's something that we haven't talked about, but we'll get into as we discuss more about when to do, when to use, what type of AI prompts and tool for what, which is the types of models that there are. But for right now, we're going to just focus on projects.
15:41So let's say that you are using, you know, you just got the pro$20 a month version of Chat2BT and you are been using it for a couple of weeks now and you have these recurring prompts that you use a lot. Let's say it's for a daily task that you have to do. You know, one that I have to do is I have to, you know, format an email that I have to send out in the day. So I have a Word document and I need to put it in the right email structure that I need to send it out in email format. But instead of copying and pasting that every day and driving yourself nuts, you can create what's called a project.
16:17And both Claude and StratGPT have this feature. So I'm going to show examples of it in both. And it's so handy. Right here, it's this little folder icon. And you hit New Project. And then you give it a name. So, Corey, what should we name this? we should name this grant and cory's fantastical project okay great not i don't know that it's a word but i just was trying to be ridiculous it is today and you know what i would i would also say is like usually what i do when i uh create folders is i name them based on the task that i'm trying to accomplish so i usually do full like task-based folders and you'll see why in a minute instead of I have the Neuron Podcast showrunners, one of my favorites.
17:07There you go. Love that. Okay, so here it's like now you've opened a new chat window inside this folder, which means all of your chats will be organized inside this project. And you can have the option to add files or add instructions. Now, when you add files, that actually gives this project specifically context about, you know, things like documents that you upload to it. Or how you like a report. That's true. Yeah, you can you can put all sorts of things in here to have it referenced. And the way that we would do that is we would say something to the effect of, let's say we add a style guide right to our project, we would say like, based on the and then I just put it in XML, but you don't have to do that.
17:51the style guide in your project knowledge or maybe project folder helped me, you know, edit my article for today so that it matches our tone, right? So that's how you would reference the files in here. But then you can do this really handy thing called instructions. So instructions Instructions are basically where you would pre-save the prompt that you are copying and pasting every day. So that all you have to do, instead of typing out this, right, you just cut this and you put it in the instructions. So you cut the prompt that you would have typed into the context window and you just put it in your instructions and save it.
18:37now all you have to do is paste the context of what you want that task to reference into the window so for instance we said based on the style guide and your project knowledge help me edit my article for today so you would paste the article in and then it will automatically you know just hit enter and you're good to go yeah it'll even uh it'll it'll run through there it'll go over to reference the style guide and uh it's so easy so when when would you want to use a project as opposed to just a one-off prompt for me i use it when i've got something i know i'm going to be doing consistently like um once we've decided on a topic for this and we have kind of a base outline i have it generate a show notes doc for me each week um and i'll go in and i'll throw in my notes on everything we have, the things we want to go over, hit enter.
19:35And, uh, and it runs through 03, I believe. And then it spits me out. I don't have to give me a doc. I have to give it to me as plain text for no good reason whatsoever. Uh, but it, but it comes out with headers. So it copies and pastes into Google docs really easily, uh, pre-formatted. Um, there's really nothing to it and it's not perfect. You know, it always needs some tweaks, but I'm, I'm not looking for it to do the thing. I'm just looking for it to organize my thoughts into something I can go in and clean up. And that's essentially what it's doing for me. I have another that does edits for me.
20:12I have one that helps write ad copy. And that's all it does is, you know, I can just go in and be like, here's a link to a company and hit enter. And it spits out, you know, a 60 second ad read, for example, or something. It's not perfect. You still have to, again, just like with anything AI, you still have to go through it. You still have to make sure it works, that it's good. But it's all about getting to 80 % faster. Yeah, exactly. It's getting that first draft done, like solving the cold start problem. So what I would say is, in addition to what you just mentioned, after you run a prompt a couple of times and you see what does and doesn't work, when you include this request or that request or this feature, that feature, and let's say you structure the prompt, you get it in the way you like it, and a couple times you run it, it doesn't give you exactly what you want.
21:07So you look at what the prompt output was and you realize where it missed the mark on what you were looking for. And then you go back and you edit the prompt again and rerun it until you get to a point where it gets you that 80%, 95 % of the way there. then that's when you would bring this over to a project and say okay this is very reliable for me so i can just run it yeah um like yeah and to be clear those are things i've refined i should say that that they're not totally i just threw a loose prompt and they're very much things where i had done it a few times in a in a discussion in a regular chat and what i'll do is i'll go back in i'll go up to search and find the chat i used last time and uh once it's where i want it i just tell give me a prompt for a project that will do what we've done right here.
21:53And it will usually give me one that's just spot on and will work really well. Usually, even with spitting one out that way, it still needs a little tweak or two to get the prompt doing what I want because I don't know what the difference is between a chat and a project, but it always feels like there's just a little something lost between the two types of conversations. uh so i can see that but you know it's just a minor tweak a few minutes you're done let me give you an example i don't have to do it again ever well so you mentioned the podcast helper right so we've got my project folder here for working on the neuron i also created a project podcast helper with uh some podcasts and youtube request templates in the in the uh project knowledge here some show note examples and our style guide i've also got my prompt here you are a helpful podcast assistant helping us create content for the Neuron podcast.
22:50And I've got all my instructions in there. So I can take the show notes from this episode, which I just copied. And I like the way that Claude does this, where it just has a nice little, like, this is the document that you pasted here, which is nice. And then I can say... I wish I did that with all of them. Is there like a word count limit? I'm really... There is a word count limit. Yeah. Okay. So anytime you paste anything in here, I think it's above 600 words or around there. It will automatically kick it out to this. But if you hit. I love that. It's great. But if you don't want it to do that, you can also do, I think it's command shift V.
23:27And then you can paste it in and it'll just automatically. Can you alternatively force a shorter one to go into that file thing? You might be able to, but I don't know how. if you're reading if you're watching and you know how to do that shoot us a note i'd love to know the trick we want to hear it yeah so okay so based on these show notes uh help me write a nice little email for this episode i'm being a little tongue-in-cheek here i wouldn't phrase it this way help me write a nice little email for this episode uh and this helps because in my project knowledge here i have kind of like three or four different tasks that uh it could potentially do based on the context so that's why i'm giving it a little nudge um but it knows what to do based on this and so we're going to run it over here see what happens okay so this is thinking as we've talked about before thinking is when the reasoning model is running on the background and running through options before it spits out an output to you here yeah okay so here it got a little confused it said I'd be happy to help write an email for this episode.
24:34I want to make sure I'm giving you exactly what you need. Okay. Email draft instructions. This would create a full promotional email using a previous email as a template or a simpler promotional email. Just a nice email to promote the episode based on the show notes. I'll say give me both options, one and two. I say that so often. I want all of them. I get so greedy. Yeah. Yeah. This is just for fun. So, you know, why not? for sure for sure oh it needs okay so i i told it this it's smarter than you think it is because see i told it that i need to give it specific things so i need to give it the youtube episode transcript with time codes i need to give it the actual youtube link so it knows it knows knows what i want i like that all right so then now here's the simple version so grant you're probably making ai harder than it needs to be here's the thing most people just jump straight to building complex agents when a simple prompt would do the job.
25:33Others spend hours crafting elaborate prompts when just asking a direct question works better. So how do you actually decide between chatting with AI, writing a structured prompt, building a project, or deploying an agent? What do you think of this? And that's the million dollar question, isn't it? Yeah. And then it gives you a simple decision tree. No spoilers. So we'll switch out of this. But yeah, that's basically how projects work. Once you have a structured prompt that you no works 90 to 95 percent of the time you make that as the you set that as the custom instructions for the project and then you can basically one-off easily run through a bunch of different tasks in the day just by giving it the context it needs super quick and easy well i guess now the next question would be let's talk a little about agents okay so we have established what a basic prompt is we've established what a structured prompt is and we've established what a project folder is and when you would probably want to use all three of those things.
26:32So what about agents? Agents is what everyone is talking about these days. They are very exciting. They're the new buzzword. But when would you actually want to use an agent versus what we just described before? All the time, Grant. Well, first, let's define what an agent is, Corey. Okay. Cool. Cool it. All right. I know we're excited, but we need to just calm down here. Okay. Agents are AI that adapt to unexpected inputs and can make decisions with minimal human intervention. Put simply, they're autonomous systems that perceive, reason, and act towards a goal. Even simpler, they can reason, make decisions, and act.
27:11React. I love that. That's a really good, easy framework to follow. We'll shout out the creator of the React framework in the description. There you go. So anyway, so agents have a lot of pros and cons, right? You know, you can control. You actually give up control when you use an agent. You're basically saying, hey, here's the goal, and then you decide how it gets done. Yeah. Or do the thing while I sip coffee. Ta-ta. Or do other work. Ta-ta for now. Yeah. But they do have a lot of adaptability. So instead of that really rigid, structured prompt format that we just described where you're kind of manually going through exactly what it needs to do and even sometimes listing out the step-by-step instructions of what it needs to do, you're basically saying go ahead and do it however you need to as long as it accomplishes this goal.
28:04Here's the end result I want. Go figure that out. Do what it takes to do that. And come back to me when it's finished and you have something for me to take. Yeah. So, I mean, to a certain extent, that is kind of a personal decision that everyone has to make, which is when am I willing to give up control in favor of less, like, hand-holding? You know, that answer might be different depending on the task. That answer might be different depending on the person and their level of comfort with this stuff. A lot of times there are still things that agents can't do on their own. So it's also important to maybe test it out first before you say, okay, I'm going to now dedicate all my resources to an agent going out to do this.
28:46Also, keep in mind, agents, they can get pretty expensive because the agent is literally going out and spending a lot of thinking time. Think about all the tokens, how much it costs to send a prompt. Well, eventually, especially if you're using an agent over the API, that's going to add up quickly. Yeah, just like in the normal world, time is money. Time is money, right. So, you know, maybe it's not efficient to have an agent go out and do something. Maybe it's much more efficient to figure out how to do the 90 % version yourself and, you know, with the prompts that you have set up and save the agent for something where you're like, I don't know how the heck I would do this otherwise.
29:24I mean, do you agree? Do you have anything else to add to that? I absolutely agree. You know, for me, it's very much a, and is this something that I like to be in control of? Like, I'll give you a good example real quick. Right now, I'm using it to kind of augment my news gathering. I'm using ChatGPT agents specifically, which is like, there are different levels of agent, which we've talked about before. And if anyone needs, we can go into that. But right now, what I'm doing is in the morning, I sit down and I read through the AI headlines I find in all of the news sources that I read every day.
29:56I read a lot of news. And in going through those, I always know I am missing so much. So something I've been doing since JGBT Agent launched a couple weeks ago, aside from buying a Les Paul, I have been having it help me with that in the morning. So now I've got it scouring the Internet and finding me a dozen headlines from the last 12 hours. and it does it well and it consistently grabs the kind of stuff I never see where I'm looking. It's really good about finding different stuff. So I'm not necessarily using this instead of doing what I do. I'm using it to make the end result of what I do better, if that makes sense.
Read the full transcript
30:42Because I now have a more well-rounded set of articles I can share with the team. and uh you know i just think it gives me a better ability to ingest more without it without having to do the manual lift and i have other tests that i just have it go to and and it very much is though you're right it's a balancing act because like there are some things i'm not ready to let go of and some of them just for me personally it's less about being ready to let go of it and more about can it do it well? Like, you know, if it were up to me, I'd love to just have all the time in the world to, you know, sit around and draw and play guitar and, I don't know, take a nap.
31:23But the truth of the matter is... We ain't there yet. We ain't there yet. And it's not ready for every task. That's true. But what I will say is I am consistently amazed at the tasks that it's ready for that I didn't think it would be. Yeah, that's fair. Well, I mean, why is that, right? One of the reasons is because in order for an agent, let's use Chachapiti agent as the example, to accomplish any given task, especially a complicated one, it needs to have access to all of the tools that you would use to accomplish those tasks. So let's say, you know, you have access to your Google Drive, you have access to all of your logins for all the websites you're logged into, you know, the Chachapiti agent.
32:07In your Wall Street Journal account. Yeah, exactly. The chat to UT agent would need all of those logins to be able to go and read all the same stuff that you do and have access to the same information that you have. There are ways to give it a lot of that information, but there's also risks like we talked about in the chat to UT agent episode. Another reason why it might not be able to accomplish everything is because of hallucinations. And if you have, let's say, an agent that is constantly going and querying and potentially hallucinates somewhere in the process, the failure rate just gets more and more high and more and more likely the more steps in the process.
32:42Yeah, yeah. The earlier they fail, but also just the more steps in the process. The more complicated the process is, the more likely it is to fail somewhere along the way, which if it fails somewhere along the way, you know, Chattapity Agents is pretty smart. It's able to figure out when it's messed up and fix it a lot of the time. We don't know what all is going on in the back end to make that happen. But, I mean, it's able to pull up from a lot of mistakes, but sometimes it's not. And, you know, the finished result is the finished result. Like if you if you have an output that you need and you don't get that output, it doesn't matter if, you know, it got 90 percent of the way there.
33:17It didn't finish the job. So, yeah. What do I think I want to come down to the task? Exactly. And some tasks, 90 percent is good enough. No, you're right. You're right. You're right. And I think knowing yourself and what you're trying to accomplish is really going to help with, you know, okay, is this actually a 90 % task or is this a no, I need this 100 % of the way done or 99 % of the way done? Now, one other thing I want to say about using an agent. Now, you only, like, let's, again, we're going to use Chatubuti agent as an example here. when you give the agent the task, you are literally only giving it one message and then you're letting it go off and cook, right?
34:02You're letting it come back to you with the goal accomplished or not. So that means that you need to give it as much context possible about what a finished product means and enough tools in order to succeed at the job. So it kind of changes your prompt structure a little bit when you use it that way. For example, you want to say like, okay, this is the exact output that I'm looking for. Don't just say, hey, go off and do this, don't make me a spreadsheet like we did in the previous example. It needs to be, no, this is exactly what done looks like. It needs to have these exact columns. It needs to have these formulas or, you know, a version of that with as much detail as possible so that it can succeed as...
34:52And it needs to have that every single time you're triggered. Exactly. Like it needs to not just be right. It needs to be right almost all of the time to be of value when you're dealing with an agent, I would say. Yeah. And another thing, you know, besides giving as much context as possible is you can also use connectors, which, you know, the technical term for that is MCP, but that basically lets you connect it with your actual Google Drive so that the agent, when it goes out and it does something, it can actually reference Google Drive, get in there, you know, use tools that are related to Google Drive and, you know, be able to take action.
35:28So incredibly useful. It's awesome. Yeah. Why don't I go ahead and just very briefly share my screen again one more time. We're back in Grant and Corey's fantastical project. Let's say we wanted to use an agent inside this project. We can with tools. So that's how you would get to agent mode. And as we showed last time, if you type backslash, you can also call agent mode very quickly. Okay. So then we're back outside of our project window and to use a connector, for example, you would go down to the tools dropdown and select connectors, and then you would add a source. And there's a bunch of different connectors here that you can choose from more here you can even create your own by pasting in a name and the link like so uh with your authentication tokens there so that's how you would you know there's some pre-built ones in here google drive canva dropbox gmail so those are all great theoretically you could also tell it to drop its output there couldn't you What do you mean by that?
36:31I don't know. I don't know. I'm asking a question on the fly here that neither of us is prepared to answer. I know, and I know we're not. You know, if you know in the comments, let us know because Corey, you stumped me. I don't know. I was just thinking, wouldn't it be great if you could create an image and be like, drop this in my AI pictures folder? Oh, maybe. yeah and then so the up here at the top you see this like kind of technical looking jumble mess what is this oh four mini high well this is the model that you're using and it's worth explaining what the difference between the different models are just in brief basically when you go to this drop down and you know GPT-5 might make this obsolete in a week or so but just so people know you can choose by the time you watch this yeah that's true
37:25hopefully we don't have to record another podcast. But it has a list of these different models and they're all O something. It's either 4.0 or O3. What the heck is all this? Well, basically, even though ChatGPT OpenAI knowingly has terrible names for these and they're trying to fix that, this is the difference. GPT 4.0 is the one where if you just went to chat.com right now and asked a question, hey, what's up? it would answer you very quickly. And no thinking just gives you an answer, right? Now, any of the O-series models are reasoning models, and that means that they think about it beforehand.
38:05O4 Mini High just still happens to be the fastest at that. So if I say, and what's cool about this is you can actually change mid-conversation. So you'd be like, hey, use web search and tell me about the neuron. and then it will think about it before it goes off and does its task. Now, because we used web search here, you can actually watch it use the tool in the thinking window, which is kind of cool. Oh, never mind. Didn't do it. It's because I didn't turn on web search. That's why. It should do it on its own. You don't usually have to turn it on. It did it on the back end. Wow. Yeah. Yeah. You don't have to tell it to search anymore.
38:46Oh, that's cool. All right. It just knows when to use it. That's why I keep thinking we're watching them build GPT-5 in public because all the features we've been hearing about keep trickling out a little piece at a time. That makes sense. And then lastly, if you have to do something, let's say you pay for the$200 a month Pro version and you have this thing called O3 Pro, it'll think for a really, really long time before it'll answer you. So you can say, like, tell me the meaning of life. And then we'll just let that go. Maybe by the end of the episode, we'll... If it comes back and says 42, I'm going to fall out of my chair.
39:22Well, we'll see what happens. Just to show the same example on Anthropic real quick, if you go down here to the little tools tab, you can also see what connectors you have connected and topple them on or off with this manage connectors. You go in here, you can browse connectors. Getting a lot of options these days. more every day. Yeah, it's getting more. Or you can add a custom one. Which, we'll do a whole episode on MCPs and how that works at another time. But, yeah. Yeah, that's one of those tools that... MCPs are a thing that sounds really intimidating and complex. But in reality, it's a fancy word for connecting one tool to another.
40:10Every time you download a mobile game and pop in and connect it to your Google, that's a little different, but at its core, that's what you're looking to do here, is connect your Google to this. What it is is much more in-depth, and it can do a lot more than the other option, but that's really the best way to think of it and feel a little less intimidated by the idea, and it's not that hard. I would add one more amendment to that, and I would just say that it's like connecting an agent to a tool and then giving it the documentation to explain the tool so that it can actually use the tool. So it's like built-in functions, explanation of when and how to use the functions, and then permission to use the functions and access that data.
40:54To go do that. Yeah. Yeah, and that's a big thing. Yeah. That opens a lot of doors. All right. I'll keep you posted, Corey, on when we get the meaning of life from O3 Pro. Yeah, I want to know what it is. I want to know what it is. I've often wondered, and I'm curious to see what 03 Pearl with a good half hour under its belt comes back and tells us the meaning of life is. It's probably going to tell us that's a stupid question. There is no answer. Everyone makes it up for themselves. You decide your own meaning. Everyone makes it up for themselves. Yeah. Yeah. If there's not a 42 joke, I'll be really disappointed as a Douglas Adams.
41:33You know if it was Grok, that 42 joke would be front and center. It absolutely would be. Like it should be its only answer. Just, you know, maybe come back a few minutes later with something else. But, okay. So now we've talked about what these different methods are, kind of how you use them. But now we're going to get into the nitty-gritty and talk about how we decide which task is right for which. And we're going to start by basically presenting you with a little decision tree kind of thing here that's going to help you work through what tool in your belt is best for a specific task. And essentially it's the goal, the takeaway from this is to be able to start with the simplest tool that solves the problem and only move up to the next step if you have to.
42:26You know, that's the key. If a simple prompt gets you the results you need, don't go build an agent about it. You don't have to. Unless it's something that you need to do over and over. And that's what we're going to get into here with this little four-step framework we've put together to help you make decisions on the go. Okay. Yeah. So number one, use a chat for brainstorming and one-off questions. If you want to dive deep down some rabbit hole about the history of bunny rabbits and whether they really eat carrots, you can absolutely do that in a normal chat window. It is not a thing. If you're a rabbit truther and you're trying to find the evidence that supports your conspiracy theory, then yes, please use chat for brainstorming.
43:16Please use the chat window. That's all you need for that. and honestly, you can do an awful lot more there than what most people would think. There's a whole lot of stuff you can do in a chat window just asking it a question without having to do some fancy prompt, without having to get into all of that. You know, the whole mission of this all along has been, and we've heard Altman talk about this a few times as well as many of the others in the industry, that, you know, the idea is for you to not need to understand prompting to be able to use AI. The idea is for you to be able to just interface with it and get the results you want.
43:57And we've gotten a whole heck of a lot closer over the last two and a half years, I would say. Wouldn't you, Grant? Yeah, definitely. I mean, even just like all of the tools that you can use. I mean, you can call up web search just by asking for it. You can create an image just by asking for it. You can pull up ChatGPT Agent and be like, Go make me a spreadsheet or a presentation on this topic. There's a lot of things. Order me a pizza. Yeah. I still haven't tested that one. I'm not giving ChatGPT my credit card information. I haven't ordered anything, but I vow to have ordered something directly through ChatGPT agent before we record our next podcast and I'll report back.
44:37I want to hear how it goes, man. I will order something and spend actual dollars and take the risk. If it goes well, let's do it live on the pod again. The risk doesn't scare me at all because everyone has our data now. It's everywhere. I'm of no illusion that anything I do is protecting it. I'm literally using Google to log into it where they have access to everything. everything and from my youtube viewings to my you know purchases and web browsing habits they've got it all so uh the last thing i'm worried about is you know this sitting in a in a data center somewhere where it'll one day hopefully be deleted or if the court case goes right yeah read the neuron if you want to know what that's talking about prompts grant oh i'm sorry no you Sorry, go for it.
45:29Go for it. Okay. So, yeah. When would you say a structured prompt is the right call, Grant? So, we kind of hinted at this earlier, but a structured prompt would be for a multi-part or a formatted output. So, let me break that down. So, multi-part task would be you need the agent. Well, I'm using agent. Sorry. You need the AI to basically go out and do multiple things at once, reference multiple documents that you're sharing with it, or anything that's just like more complicated than a one-off task. I would definitely use a structured prompt for that. I would also use it if it's something repeatable that you're going to be doing every day, like we talked about earlier.
46:11But, you know, probably you want to push that more into project area. And then formatted output. So if you have something where you really want it to be like really structured in the way that it delivers you the result that you want, which honestly, probably a lot of things end up in that territory if you're having it create something for you. Like at a certain point, you just tell it like, I need you to do exactly this format layout. I need you to hyperlink all of the links. I need you to, you know, make sure the headings are sentence case and not heading case. Whatever it is, just really outline the hard format necessities, and it will probably do it for you.
46:51Yeah, it probably will. And the truth is, structured prompts also apply to the next two options we're going to throw at you about when to choose a project or an agent, because both of those involve structured prompting. You know, I mean, that's where you get into, you've probably heard the phrase chain of thought before, where essentially what that is is just a prompt where you're walking it. through the steps you wanted to take and the order you wanted to take them so that it gets to the end result in the way you think. Sometimes it can be limiting, though. The truth is, sometimes it may have a more efficient way than the way you would think through a problem or the way I would.
47:26There's a lot of things it can do simultaneously. The bitter lesson. The bitter lesson, yeah. Yeah, sometimes, is it possible that the way we do it is not perfect? Is it possible that us humans just aren't as good at just like brute force AI going out trying a million, 200 million different options and coming back with... It's like we're fallible or something. I don't understand. Yeah. So, okay. So from there, a project. Choose a project when what you need is a predictable outcome every time. Repeatable workflows where control, transparency matter. and the benefit is that if needed, the steps are right there, your instructions are right there.
48:12It can be updated quickly on the fly. If you're like, eh, this isn't working, you can go in and make tweaks as you go really easily without having to log into some other software and pull up a node where you have to go in and make changes. This is very much a, yeah, I need it to be mostly consistent, but occasionally I'm going to tweak that. And if that's the case, that's where this is good. If you're going to have to regularly swap out example documents, if, say, you know that we change how we do reports every 30 days here or something, if that's the case, a project is probably right for you because you can go in, pull one out, drop another in, you're out 20 seconds of your day.
48:52But when that's not the case, we'll go to agents and let Grant talk a little more about those. All right. So when you would want to use an agent is when the task requires autonomy, requires dynamic decision making and tool use. So basically breaking that down when you need it to just go out and figure it out yourself. I don't I can't be babysitting you the whole time. I need you to just do this for me. I need you to be able to make multiple decisions on the fly on your own. And I can't be holding your hand during that process. and I need you to be able to use tools like MCP connectors, which we showed earlier, so that you can go out and do it.
49:31So a great example of this, right, is, so we talked about N8N last week just very briefly and how people were using N8N to create agent workflows. Well, there's actually an MCP to work with N8N that you can basically, you know, chat with Claude and give it access to the N8N and MCP. and you can have it go and basically say, create the workflow for me. So that gives Claude access to the actual NNN workflow that you have set up and it can create the entire workflow. It's not going to put your API keys in for you, but it'll basically set up everything else. One week really soon, we're going to do that in a podcast because I think that is a really cool unlock that puts, you know, low-code agents in the hands of everyone.
50:25Like, it seems like such a smart direction to go. And, like, you know, to have it be able to even, you know, kind of walk you through what you do still need to go in and fill out manually, you know, that's such a powerful, powerful tool to just have at your fingertips. Something that you would have had to go to, you know, Hire an engineering firm or something four years ago. You can stumble through on your own good enough now. Yeah, yeah. Other examples we talked about last week in Chat2BT agent. Maybe you want to go out and make a spreadsheet for you, make a presentation for you, go out and research stuff for you.
51:03That's a pretty good use of agent where you're basically saying, here's my rules. This is what I need. This is what I don't want. Don't do this. make sure to include like avoids if you have any of those and then just like the output that you want like it needs to be in this format and you just let it go off and and see what it comes up with i'm gonna go grab a slice you keep working i'll be back you know uh you know that's also yeah i've got this thing that's got to be done by one o 'clock but i'm starving to death um yeah and uh there's just so many so many cool applications of that that i think it'd be a lot of fun i think there's a lot of criticism yeah go ahead no no just just one last point on this.
51:42There's a lot of criticism of people only using chat GPT like Google instead of using it like for what it's really for. And I think what we talked about in this hierarchy, just to put a pin in it, is yeah, there is the Google use cases. And actually, I think I read today that 70 % of people use it for search now who use it. So like if you need to search something, you know, it's probably, you're probably going to get closer to what you need using AI than than just a generic Google search. Google's still good. And Google's still good for, you know, when you know specifically what you need. When you don't know what you need, Chachupity Agent's usually pretty good for that.
52:19Or sorry. I use it less every week. Chachupity Web Search, not Agent. I don't know. Jury's still out on Agent if Agent's a good researcher. But it seems like it is. It seems like it is. It did me well, and I'm enjoying the heck out of the guitar I bought. But that's... Yeah. yeah but uh but that is just the first layer right that's the first layer that's simple things yes of course you can use it like google more complicated things yeah you're going to need to tweak it you're going to need to structure your prompts you're going to be thinking more like an engineer like how you would explain this to uh to a teammate or an intern then you know the next layer is I was like okay how do I organize all my recurring tasks and how do I set those up so that you know it's really easy for me to do those quickly I'm not thinking about it every time and then agent is all right, I'm going to trust you now and you go out and do this for me and I'm going to check your work.
53:09You know, and when you're working with agents, there are a few just little notes of caution to keep in mind that we've talked about before, but I want to run through them again here really quick. Number one, these don't need to be overengineered. You're adding cost. You're adding complexity. You know, every step you add is another place to fail. absolutely err on the side of keeping it simple as to the best of your ability I think is very vital and you know it's also money you know I mean you know these are running on whatever server they're on that costs money every call to an API costs you know fractions of a cent in some case but over time, it can build up quicker than you think.
53:57Though I am very impressed with how far$50 goes in AI credits. Like even two and a half years in, we're still at the, you know, $6 for a billion credits range, which is cheaper than it was when ChadGBT launched. So, so. Yeah. Yeah. It's pretty amazing. And, you know, reasoning tokens, those, those add up. You know, you can do a lot of reasoning in the app, but if you're using it with the API, those can get more expensive. And then I'm sure agent tokens, whenever they're their own thing, you know, that that's going to be even more expensive. So. Oh, yeah. Yeah, definitely will be. Definitely will be.
54:36I would also add to what you said, you know, in the same spirit of the framework that we created, start small and iterate. So, you know, let's say you have a basic task you need to do. Write the most basic prompt of it and see what it comes out with. See what you like, see what you don't like, and then change it from there. I think that's the single best way to get better at working with AI is to, you know, get the bad version out first, see the results, and then be like, okay, cool. I see that it uses these words that are really repetitive I don't like. I see that it didn't give me that output format that I expected because I didn't tell it what I wanted it to look like.
55:13It doesn't have the context about my company because I didn't give it that information. And then you can, you know, now work backwards from that and iterate and make a prompt that works 90, 95 % of the time. Yeah. Yeah. And, you know, you can always make it bigger. You can always add steps, you know, so start small. Number three, verify your outputs. Don't, like, you'll learn over time that there are areas you can trust and areas you can't. And the more you use, and I'm going to say a model, not a company. The more you work with a specific model, the more you learn its strengths and weaknesses.
55:53And you know, like, eh, that stuff's fine, but I've got to watch this pricing section or this, you know, these links. Sometimes these links 404 on me. Got to watch out for that, you know. And I think it's important to be just cognizant of that. Cognizant of that. Don't, I mean, you know, you don't have to kill yourself, but read what it puts out and make sure you trust it before you go stick your name on something. Yeah, yeah, definitely. And what does that look like in practice? It looks like, okay, let's say you say, use web search, go look up these facts, and it gives you the facts, but then it also gives you links.
56:29So that means go to the link, check to make sure the fact that it gave you is actually on the website that it gave you. Just like if you were to look at a Google OneBox result, and it gives you a nice little paragraph summary about some question you asked, you would probably want to go to the source and make sure that you understand the context of that paragraph. Like, wait a minute, does this make sense? Like, who wrote this? Like, you know, you want to verify it. Yeah, even those AI overviews on Google you've got to watch. Like, I've had pretty regular incidents where I ran across one that was just a little bit nonsense or a little off or maybe misunderstood the question.
57:06It's gotten a lot better since it was recommending glue on pizza.
57:12It's gotten way better than that. I forgot that. But yeah, no, now we're in the scary part of verifying where it's actually kind of hard to tell when it's wrong. So you really need to still be on your guard there, you know, and make sure that you're checking stuff. And it's not going to get easier. No. As they get smarter, the things where it makes the errors are going to be harder and harder for us to verify with our human knowledge. Not only that, as AI's agents are publishing things autonomously online, you're going to have content that could potentially be sourced that could be a poison pill, let's say.
57:49So you got to make sure that you use good sources too. Yes, so true. So true. Last thing related to that is always consider your data privacy and safety. So do I need to be giving ChatGPT access to my Google Drive and my Gmail for this task? No, then probably don't. do i need to spin up an agent instance where my logins are now on the cloud for buying a pizza no i could probably just use uber eats for that i don't i don't think i actually need to have agent buy me a pizza um cory's still going to but that's right i'm still going to but that's probably not you know the the best use of your time you know going out and researching a bunch of information for you and putting it in a spreadsheet and you know doing really technical stuff that is a good use of your time um yeah do you need to give it all of your logins to do that sometimes yes, sometimes no.
58:42So, yeah, you got to decide for yourself. Yeah. And, you know, be aware of whatever data the things you connect have, you know. Of course, now everything has your, you know, anytime you can use something like Apple Wallet instead of punching a credit card in, things like that, those are big wins. And it's a thing I'd love to see connected through AI soon is to be able to use my Apple wallet, Google wallet. I know Stripe is, I don't have any of the details of this off the top of my head, but I do know that Stripe is working on making payments really easy with AI. I know that ChatGPT is working with, I think it's, actually I don't want to misquote it, but they're working with somebody in order to make payments on the back end.
59:28And then in exchange for that, they're going to take a cut of those payments as one would. And that's fair. Yeah. So that's definitely coming. I pitched this idea to you, Corey, which is essentially maybe we need to think of creating a clean room equivalent, but for AI. So creating a clean account equivalent where you're basically making accounts specifically for the AI to use. You can even extend this to like, you know, buying a computer for the AI to use at some point once we get into the browsers and local AI, like just specifically for the AI. Like a Google Drive account that only contains things you're comfortable sharing.
1:00:06Exactly. So that you're not infecting, you know, you're not accidentally introducing all the other personal data you have on the computer with the, you know, stuff that you want the agents to do for work. And to be clear, odds are it would be fine anyways. But there's always a chance that it won't. And it's a matter of personal risk and your own choices and comfort levels. The more people who use these tools, the more these tools become a target for malicious actors. I'll say that. Yes. So keep that in line. Yeah, that's fair. I think it's time for the fun, exciting part of the show here where we ask our nerdy question.
1:00:45Let's do it. And today I have a good one that's relevant to what we're talking about. At least I think so. Grant, in your personal day-to-day use, what percentage is just talking to a model versus structured outputs versus projects or agents? If you were going to break your day down into a four-way pie chart, how would you roughly guesstimate you work? Okay, so for context, I write an AI newsletter. So I read a lot of AI news. I write a lot of AI news. And I'm working with AI to test it and figure out all this stuff. So based on that, I would say that I'm probably using structured prompts for 30 to 40 % of the tasks that I do.
1:01:35Where it's like, I know that I do this every day, use the prompt for that. I probably chat with the model for 20 to 25 % of the tasks that I do. Where it's like maybe a new thing that I'm doing for the first time. Or I have questions about an article that I just read and I'm trying to help understand it. things of that nature. Working in projects kind of overlaps with structured prompts for me. So I would put that in the, actually, I'm probably 60 to 70 % of my AI tasks. No, no, no. 80 to 90 % are inside projects. I'd say that's fair. Yeah. It's very rare that I'm just spinning up a random chat to do something.
1:02:14And that's because I've structured my workflow like that. Like I have it set. So, you know, I have a general purpose one that's for the neuron that I can chat with on all my one-off tasks. I have all my specific task-specific ones. So that's sort of how I have it set up. And then agents, I would say 10 % at the highest. These days I'm actually not using that much agents. It's more like experimental stuff or I have long-term agentic projects I'm working on. So that's my answer. What about you? I would say for me, just day-to-day in-and-out prompting makes up probably a third. As far as just prompting in chat, you know, dealing with standard tasks back and forth that vary a lot from day to day.
1:03:04I would say projects make up probably 20 % of my day. Well, no, closer to 30. that puts me at what 60 okay yeah this is a little off um i'm gonna put 20 at agents right now really i'm currently using about four of those connecting four of my credits a day and i am desperately going to need more of them quickly so that's that's a thing on my list but what happens is every day i keep finding something that's like that's so easy i should do that so i may need to I have some things I need to pause and look at that might be good for innate and workflows just because it's cheaper. Yeah. But yeah, I'm going to say.
1:03:5540%, no, 25 % simple one shot prompts back and forth, just communicating with it going. Let's say 20 % as structured prompts for a variety of reasons. and I'm meaning structured prompts like outside of a project. And I would then put projects at around 35%. It's a thing that increases by the week for me. I would say I'm using projects more every week. And the same applies to my work with ChaggyBT agents specifically right now, which I would say is probably 20%. So I don't know if I've added up to 100, but I think I might be 95. I don't know. I wasn't giving it to you. I wasn't either. I wasn't either really.
1:04:47I was just kind of thinking of each one individually as far as how much of what I do is this way. But I would say right now probably 20 % of my AI is using agent already. That's good. I mean, I know some people who make agents for everything. And I'm having to slow down because I know that I am going to run out of these credits soon. and have to figure out what is involved in getting more of them. Yeah, yeah, yeah, yeah. Because what you're talking about specifically is that everyone has a ChatGPT agent ticking clock countdown of how many times you can use that in a month just for the standard$20 or$200 tier.
1:05:27So the$200 tier gets about$400. The$20 tier gets about$40. I think Teams might be the same as the$20 level. Yeah, Teams is the same. I believe same 40. Okay. Yeah. And what I'm finding is that 40 is not bad. That's a respectable amount to have included for my$20 bill. However, I want so many more. Like there's a lot more I would do if I had it. I don't, I wouldn't say that I need 400. That's, that's so many. I mean, I'm apparently squandering them. So if you need me to do something for you, let me know. Yeah. We need to figure that out. Shout out in the comments. I'll do some tasks for y 'all.
1:06:12Dollar up out. Yeah. Yeah. I'm now like drop shipping my agent requests. Drop shipping agent requests. I love it. And, you know, it's been really useful. But I think that's about it for today. We'd love to thank everyone for taking the time out to watch us and listen. If you hung on until the end, please take a moment to like, subscribe, leave a comment. It's the best thing you can do to help us continue making these podcasts and bringing you the best of what we're learning in our own day-to-day AI journeys. If you haven't yet, please join more than a half million others who read the Neuron's daily AI newsletter every day.
1:06:56Visit theneuron.ai today to sign up for free at your convenience. It's a really good newsletter. I hope you'd enjoy it. And on that note, we'll see you back here next week. Farewell for now, humans.
From the publisher
A lot of people aren't sure whether they should just chat with an AI model, craft a structured prompt, spin up a project, or unleash a full-blown agent. In this episode, we break down the differences between these approaches and share a practical decision-making framework. We'll show how simple prompts excel for quick, isolated tasks, why structured prompts improve clarity and focus, when a project (workflow) is better for predictable, repeatable processes, and where autonomous agents shine for dynamic, open-ended problems. Along the way we'll demo real examples, share tips for avoiding unnecessary complexity, and help listeners decide which tool fits their use case.
Subscribe to The Neuron newsletter: https://theneuron.ai