Two-Way Prompts: SIMPLIFY your AI Agents, Agentic Workflows, Personal AI Assistant

here's a simple yet powerful idea that can help you take a large step toward useful and valuable agentic workflows I have to warn you this is one of those things that sounds really obvious after you hear it because it's hiding in plain sight the idea is simple it's called the two-way prompt so what is this why is it useful and how can it help you build better AI agent workflows two-way prompting happens all the time in real collaborative workspaces you are itively two or more agents prompting each other to drive outcomes two-way prompting happens all the time when you're at work with friends with family online in comment sections on PR reviews you ask a question your coworker responds they ask a question you respond now let's double click into what this looks like for your agentic tools right in agentic workflows you are the critical node in a communication process between you and your AI agents that are aiming to drive outcomes in most agent workflows we fire off one prompt or configure some system prompt and that's it but we're missing a ton of opportunity here that we can unlock using two-way prompts let me show you a concrete example with Ada so Ada of course is our proof of concept personal AI assistant and let me just go ahead and kick up this workflow so I can show you exactly how useful the 2A prompt can be AA let's create some example code Dan I don't see a URL on your clipboard please paste one into your editor Okay so so right away my assistant found that there wasn't a URL to create example code from and that's what this workflow is going to do for us this agentic workflow is going to create example code based on a URL that we give it so it's going to scrape it and create a nice single file where we can run code based on the documentation so my personal as assistant is actually prompting me for information so let me go ahead and open up this blog this is Simon Wilson's blog he's a prolific engineer a true chronic developer just creating products left and right he has this really really great blog post where he's concatenating an entire code base to prompt using claw three Opus right so basically he's taking a bunch of files combining them into a single prompt a single piece of text that you can then run a prompt on so what I'm going to do here is use this as our prompt value so I'm going to go and just paste this in this is going to be the example code that we want to look at so I'm going to save that file close and let our assistant look at that Dan I found the URL I'll scrape it and whip up some example code any preference on what the code should focus on okay awesome so you know another back and forth prompt this is the key value of the 2A prompt so now my assistant is asking me for additional information what do we want to focus on so you know basically I just want an endtoend example of runnable code where we can utilize Simon files to prompt for Claud so concise end to end example using Simon code to combine files to run a prompt against mirror his code AS closely as possible so I'm going to save and just close that sounds good Dan focusing on a concise endtoend example using Simon's code to combine files and run a prompt mirroring his code closely generating the code now awesome so you know my assistant is saying exactly what's happening um I really like this touch of the personal AI assistant it's a conversational um event where you're talking back and forth and really feels like we're talking to a computer in more collaborative way where we're driving outcomes together which is really really cool so you can see there there's the first version coming out this is going to run a prompt chain I'm not going to dig into the code in this video definitely check out the previous video where we walked through some of this code in detail we have the llm agent router you can see I've added a couple new agent workflows here but long story short you know we're running this example code workflow which has a agentic workflow that has this two-way prompt interaction and then we have a prompt chain where we scraping the content and then we're running several code generation prompt chains to make sure that we're just going tocode and then we're going to save it to a file I'll make sure that the previous video is in a description where we went in more depth of how exactly all this is working let's see how our assistant's doing Dan I've added getting your file cheer. to the working directory anything else awesome so we've restarted that Loop you can see that took quite a bit of time this agentic workflow is using the latest gbt and the text to speech speech to text functionality still a little slow we're working on that but let's go and and just see the results here so let me stop my assistant here and let's open up this file and let's take a look at this so nice this looks pretty spoton this looks exactly like Simon's code that looks nearly identical it's got some really clean comments for us explaining exactly what's happening on each line super useful I'm not going to run this it's not the point of this video but you can imagine this is going to give us a huge Head Start when we're looking at new code bases new documentation this workflow provides a lot of value and it provided it in a pretty clean simple way where our personal AI system is prompting us for information right so let's go ahead and run one more workflow I want to show this off in one more example but let's go ahead and create a view component so let's say python main 9 Ada let's create a new view component Dan let's generate a view component please choose an image okay so again we're starting to hop into a workflow here so my assistant is asking me to give it an image so we're using the new gp4 vision with Json model from open AI my assistant is prompting me the user the Builder the engineer to give it a file right so it needs a file to look at and we'll go ahead and use this uh pre- 10K Subs so this is a you know Simple Text view of before the channel hit 10K Subs so we'll just go ahead and drop that in and we'll leave this up so we can compare right so go ahead and we'll hit pre 10K Subs got it Dan generating that view component now so there's the basics four print message there's a component there's a prompt for gener sounds good Dan I finished the view component and called it Channel analytics card view any edits needed okay so here's another really cool two-way prompt Ada is opening up those files and it's showing me exactly what we have here so I have a simple view web application running here and all we're going to do is place this new component which should be placed right in the components directory nice all we're going to do is say uh what was that called Channel analytics and let's go ahead and get this imported and Bam so you can see there that's the component generated this looks pretty good so there's the current version there's the original version so not too bad it's got this blue background it's got green on the entire lower text and needs a little bit of space and so all I'm going to do here is create a simple prompt that's going to you know request some of these changes so in the prompt I'm just going to say let's add more padding and only the plus 645 should be green trailing text is gray so I'll just save that let that run close this as well hop back to our assistant and let's see how that plays out sounds good Dan I've updated the view component what should we tackle next awesome okay so that's finished let's go ahead and take a look at that nice so you can see here we have some additional padding not the exact padding that I wanted I wanted some vertical padding I should have been more specific there but we did get padding and we did get this trailing text in Gray and of course we can look at the component to see exactly how that manifested and yeah so as you would expect it added that span text Gray using tell wind awesome stuff here so really really simple examples but I hope that they show you you know the power and the potential of using these two-way prompts in your agentic workflows you can really get a lot more done and reduce complexity of your system by just having your AI agents prompt you the driver of the workflow and there are many manifestations of how this can look so that's a couple simple examples of what the 2-way prompt looks like you can imagine much more complex back and forth systems where your agent needs information from you where it's not sure on how to you know phrase and email or it doesn't know exactly how you want your code to look or it's asking questions about what to develop next and this can all be happening in the background right where your agent has done a bunch of work for you and it needs some review it needs some of your input to make next steps in a really concrete way so so if that makes sense and you can see the value of you know adding some two-way prompts to your agentic workflows hit the like hit the sub and let's address the elephant in the room right you might be thinking wait isn't this just human in the loop this is a like common PA of having the human provide feedback and 100% two-way prompts are multi-step human in the loop I would say the big differentiator is that you're not limited to just simple text prompt responses to just giving feedback you can provide all types of feedback in our example we showed off us selecting specific files we showed off us pasting URLs we want to be examined we showed off providing feedback for improvements on our view component and there are many many many more ways to to add multi-step human in the loop AKA two-way prompts into your agentic systems the key Insight here is that you the engineer the product Builder and the user are an essential guiding force in your agentic workflow right your workflow should be centered around you and your users right whoever's using your application whoever's using your product or tool this is the key Insight don't get caught up in trying to build a gentic workflows just because it's agentic right it's the new Kool-Aid to drink it's the cool thing everyone's doing it this is what engering is all about now like of course that's important but always focus on the end result for you and your users add two-way prompts and multi-step human in the loop to drive results in your agentic workflows this can be especially useful as stepping stones as you're developing your fully agentic workflows don't be afraid to create some you know partial agentic workflows with the human in the loop right using these two-way prompts you know especially as these co-pilots and assistants are built out and as you build your own agentic workflows keep in mind that you and your users are the critical asset in this communication process I want to close up with one more idea here and that is where this is all going right how I see this breaking down so I see three tiers of agentic workflows increasing in complexity as we go down the list there's humans prompt this is us we are effectively pulling information from our large language models and then we get to this next level where we're two-way prompting in the two-way prompts we're doing both pulling and pushing right our agents are pushing information back to us asking us questions I think right now when we're talking about agentic workflows and applications that are built on top of llms we're doing a lot of number one and some of number two right where the user and the agentic workflow talk back and forth to each other via a domain specific uiux and this is very powerful there's a lot to be uncovered here I predict that most applications will end up in the second Zone in the future this will all change and it'll be primarily agents pushing to us right our agents will be the one prompting us and this things accelerate to the next level where you wake up in the morning and your agents have been doing a ton of work for you overnight right they're responding to emails they're reviewing code they're writing code for you they're generating new ideas for your business they're setting up plans they're monitoring your key infrastructure they're you know they're looking at your gcp Cloud your zero Cloud making sure everything's stable they're reporting anomalies to you and then and you know when you wake up and you you know go to your device whatever app it is whatever tool you're using they're going to ask you a bunch of questions on things they need your input on right it's going to be this really interesting flippening of who's prompting who and that's really the key Insight here right you the engineer and the product Builder and your users are the essential guiding force and it's important to utilize that definitely dig into this idea put yourself in your agentic workflows more often so that you can simplify things and also really just drive the result that you're looking for right that is the reason we're building on these agentic systems it's a huge powerful unlock but it's really to drive a result for us and our users so that's the end game our agents will be prompting us right now we're building and working with tier one agenic workflows and two-way prompt workflows right where it's not just you firing off some initial prompt and your workflows doing everything you can create a lot more intricate interesting results with these two-way prompts so I think that's where things are going I hope this idea was helpful for you in future videos we'll continue discussing and building our personal AI assistant we'll be pushing our AI coding assistance and our AI coding skills even further and we'll keep building great AI agents that can operate these agentic workflows on our behalf thanks so much for watching if you enjoyed this video If you like this content if you like these ideas you know what to do stay focused keep building and I'll see you in the next one

Channel: IndyDevDan

Views: 12,212

Keywords: agentic, agentic workflow, ai agents, claude-3, opus, claude 3 opus, gpt4, gpt4v, gpt4 vision, gpt-4-turbo-2024-04-09, software, product, engineering, prompt engineering, prompt, prompt chain, prompt graph, prompt orchestration, ada, ada ai, two-way prompt, two way prompt, human in the loop

Id: sTruFeIO0iA

Length: 12min 56sec (776 seconds)

Published: Mon Apr 22 2024