How to build a prompt in Aisle from scratch. Mitch walks through the Prompts feature: setting up a prompt, configuring structured outputs, and deploying it to your team.
This is the first in a series covering how Aisle actually works. More coming.
Video transcript
Hi, I'm Mitch and I'm going to walk you through the prompts feature of Aisle today. So this is the prompt library. This is where all of your teams AI tools will live. Right now, it's empty, so we're going to build a prompt from scratch. I'm going to build a prompt called Reddit Qualifier.
The idea here is that we will feed it Reddit threads, and it's going to evaluate them against some criteria. It's going to tell us whether Aisle should or shouldn't be engaging in that conversation. And eventually we're going to put this prompt inside a workflow, but for now we're just going to build it. So the prompter builder has a few key parts. You can add a system message, which sets the model's role and behavior.
The user message, which is the actual instruction that runs every time. And then there are variables. Which we'll get into in a minute. This is the inputs that might change every use. For this prompt, I want to set a structured output and enforce a JSON output.
For a model. I am happy to use sonnet, but there are a number of models available in here that I could use. All of those are configurable by your organization admin. I am also going to take the temperature down a little bit. This is a qualification task so I want it to stick to the criteria that we define and not get too creative.
I have already prepared this prompt off screen, so here is my system message. Essentially it is setting a persona of a lead qualification engine. Now the user message where we begin to structure the task, right? So I want it to evaluate these Reddit posts and return a JSON object. I'm going to set a variable using curly brace syntax, Reddit posts, and this is where we will inject the Reddit posts each time we wanna run this prompt.
I'm just going to paste in the rest of this prompt ' cause like all good prompts, it can get quite long. Great setting some criteria about. What is interesting and important for us you can see here that that variable is tracked. Now I'm going to structure the output that I want from this prompt, right? So I want it to return a posts array that has objects in it with these properties.
So we will have a post title. A score, which is a number, we will have a reason, a reply angle, and the URL to the post. Great. This looks good to me. Let's publish it.
So once a prompt is published, there's a couple of different ways that you could use it. First you can open it directly in chat and interact with it. That way your team can do the same. They can see it in the library they can run it with their own inputs and get the same expected. Results every time you manage the prompt centrally.
So they don't need to know how it works or how it's been improved. It's abstracted away from, from your team. Secondly, you can deploy it to an API endpoint. This means that you can call any prompt from your own application. You can pass in variables programmatically and get the response back, whether that be a structured JSON response or.
Just the, the LLM raw response itself. Your code base is calling Aisle and again, the prompt is abstracted away from the surface area. So when you update the prompt, the endpoint that is being called gets those updates automatically. You don't need to do any code changes. The third thing that you can do that we will do later is we can use it in a workflow.
Two other things worth knowing. About prompts in Aisle is that all prompts have logs. So you can see where it ran, how many tokens it used, how long it took, what came back who used it. You can turn these logs on and off. But this is one way that you can identify issues with a prompt.
The second thing is that every time you update a prompt, I automatically creates a new version. So as you iterate on things over time. You are collecting a version history and if you end up noticing some, some output degradation and you wanna roll back to a previous version of a prompt you can do that quite easily in one click. You're not guessing what's changed. The evidence is right there for you.
That's a quick overview of our prompts feature. stay tuned and we will dive into more features within the Aisle platform.