ONE THING TO TRY
Try two prompts for the same real task: a broad request and a version with the relevant facts, a specific job, and the questions still to answer. Compare the results against what you actually need.
Give the goal a next step
Growing a business is a goal. Drafting a reply to a particular customer inquiry is a job. Define the information the task starts with, what it should produce, and what makes the result usable.
Use the facts that change the answer
In the episode’s fictional example, a customer asks about a website for a second location. Aaron talks through the business’s usual pricing, the customer’s situation, the proposed scope, and the questions that still need answers before a useful response can be drafted.
Talk it through
A voice memo or dictation can help you explain the task in your own words. Connecting AI to business information can supply context, but you still need to tell it what to do with that information. A large pile of files does not automatically make the assignment clear.
Compare two prompts
Use the same task twice: first with a broad request, then with relevant context and a specific job. Compare the responses, including their assumptions about pricing, timing, and missing information. Keep your judgment involved as you improve the instructions.
Links from this episode
Transcript
Why polished answers still miss
Hey, welcome back to another episode of Think Then Make. And today we're talking about giving AI a clear job. So you ask AI to help with something in your business, it gives you a super polished answer, it's like super formatted, it's got all the em dashes in the world. But maybe there's even some ideas that seem promising and the formatting is really clean, but you still can't use it. And it really isn't because it's not good content, it's because it doesn't sound like you. Maybe it assumes a service that you don't offer, maybe it solves a problem adjacent to one you're actually trying to solve. So you try again, make it better, make it more human, make it more strategic, make no mistakes, all the things we tell our AIs. And instead, we just get a different polished answer that still leaves you doing all the work. And so today, I kind of want to walk through this idea of a more useful starting point in giving AI a job you can describe and give yourself a way to judge whether that job was done. So this isn't really about finding a magic phrase, it's about making the work clear enough that you can actually review it. So we'll use a made-up customer inquiry as our example. You don't need to follow on screen or copy along. At the end of this, I'll give you a short brief you can say out loud.
Turn a broad goal into a specific job
But kind of what's this focus is, a wish sounds like this. Help me grow my business, improve our customer service, write better marketing, and make this more professional. Those are understandable goals. They don't tell you what a useful piece of work looks like though. So two people could hear improve our customer service and go in a completely different direction. One might rewrite emails, another might change response times, another might simplify the process customers have to go through to get help. But if you're that person receiving the assignment, you probably wanna ask a few questions before getting started. Now, AI can produce an answer even when those questions haven't been answered. But that can mean the missing information is less obvious because there's already something on the page. So the existence of an answer doesn't tell us whether the assignment was clear. We have to distinguish the broad goal from the next job. And so growing that business might be the goal, drafting a response to a specific kind of inquiry could be the job, improving customer service might be the goal, organizing the question customers repeatedly ask also could be the job. So a useful job always has an input and output and then a standard I can check. So we don't have to make it that complicated, but like what does the information start with? What should it produce? What would make that result usable or unusable?
A fictional second-location website inquiry
So I have to follow one inquiry. So imagine a small business that receives a request for custom work. A potential customer has written, I'm opening a second location, could you help me with the website and how much would that cost? I'm sure for those in the agency world, we've been asked a very similar question like this or maybe we're pitching that thing. So the person wants to be helpful, but there isn't enough information to price the work. So this is where you bring in AI. You bring in the old ChatGPT and you say, hey, if I ask AI to write a professional response, it has a lot of room to fill in the gaps. So it's just gonna start inventing prices. It's gonna start making up things and you're gonna be like, hey, pause Aaron, but I have it connected to my second brain. It may be, maybe AI still makes mistakes, but you have to give context.
Talk through the context before asking for a draft
So this is why I love when I'm doing anything with AI, I'm prompting, I try to give a specific goal on a specific job. So I'm not talking about the forward slash goal thing in ChatGPT or Claude, but I'm talking about, hey, so I'm looking to respond to this customer. I'll even take a screenshot saying, hey, here's a screenshot that they're asking about second location website, what's the cost gonna be? Then what I'm going to do, this is my practical, easy kind of process for getting the information is I will talk to text or do a voice memo or record myself talking out loud. I feel like it's a better way to process the thoughts here versus just trying to like type out a list of something. So I'll say, hey, here's what we normally charge for a website, or here's what we've done in the past. That'd be the first thing I'm gonna tell AI. The second thing I'm gonna be is give some context about the customer if the context already isn't there. And then I'm gonna talk about what's the website gonna include. And maybe I'm even gonna tell the AI, hey, we need to ask a few more questions so we can get better pricing information or better timelines. And so maybe it's, hey, help me form a response as a range of price. So that way we can ask a few more questions and make a better judgment of what that cost is going to be. That is a way better prompt and a clearer job description than going AI, find me a response to this. Now again, it might be good, but more than likely, it probably isn't gonna be super effective.
Give coding and writing tasks a clear direction
And so what we have to do is we have to start pivoting ourselves when we're writing prompts, writing posts, any of that fun stuff, in a way that's very clear with the job to AI. Same thing goes when you're coding with AI. You don't wanna just say, hey, make me an app and make me a billion dollars and make no mistakes. You could do that, but it's actually gonna make a lot of mistakes and it has zero context to what you're trying to do.
Connected information still needs a clear assignment
And I'll rewind back to the statement of, hey, Aaron, while he, AI, I almost said he, like there's a real person in the AI, but maybe the AI has all my context because it's connected to my CRM, my system, and my process, and blah, blah, blah, blah, blah. Awesome, and maybe it does, I don't know. That's your business. What I do know is we have our AI connected to many things in our business and it still messes up. It's not for a matter of skills or AGENTS.md files or prompts, it's a matter of we sometimes don't give the AI enough context and it gives us a crappy output. And now you're hearing that and you're like, well, my outputs aren't crappy. No, I'm sure a couple of them are and that's okay. So what we do is we give it a job. I'm not saying we give it a job description of saying, hey, you're the senior developer copywriter for Apple Inc., write me a post. No, I'm talking about we give it enough context so it understands the job it has to. The nice part, and this is where the beauty of your second brain or your business information is it can start pulling in some of that context as well and that's where the context becomes helpful because now it's not just assuming and making up stuff. You're actually giving a direction to go forward and move forward on.
Compare two prompts on the same task
So a couple quick things for you to try out in an ordinary use case for you to test. So testing an AI workflow doesn't mean setting up a complicated eval system. It just means starting with a few examples that represent the job and then a normal case study that would be a normal process for you. So we go back to the website issue. Try it, try two different prompts. Try one prompt where it says, hey, here's the question which is I have a client or a customer that's wanting to spin up another website for a second location. We need to give them a pricing and give them a deadline. Have that response. That's the whole prompt that you're gonna send in. The second one is going to be the one with more context. And I know you're going, well, of course, Aaron. I gave him more context but I also had to work. Yes, we have to work still with AI. We're not letting AI doing all the driving. We're like letting it be the co-pilot. We're letting it be the assistant. We're letting it be the helper to help speed up our process and system.
Try it this week
So this week, I want you to try that out. Try to give a little more context to your AI. And sometimes there can be such thing as too much context. So just dumping files upon files isn't an answer but just giving it a little more direction, giving a job, give it a specific task, give it a way to move forward. And I'm sure you're gonna see a bunch, a bunch more progress and better results in your systems and in your AI prompts and your outputs too. So check it out, give it a shot. Let us know how you're using AI. Would love to be a part of that conversation. Check us on Facebook. Other than that, folks, we'll see you on the next one.