AI Productivity Tip: Break Down Tasks, Not Brain Dump!

Full video transcript

Just going to share a little bit, some tips and tricks that we have noticed working with so many different companies and their teams and so on. A lot of these teams, the majority of them are certainly software development teams, but what I noticed is that although the medium is software, the overall idea is the way that people work and collaborate. Something that we found super interesting in our own surveys, and in the surveys that we’ve also commissioned by professional surveyors and market researchers, is the difference in self-reported productivity for people who are using various AI tools. We noticed by looking at the data that people self-reported a difference from up to 40% feeling that they are up to 40% more productive, down to smaller numbers, down to only 10% or so.

We took a closer look to see how people were using these tools, how they were organizing their workflows, and so on, and we started to develop an opinion for what accounts for that difference. There are a few things that accounted for that difference that are specific to software, but I’m not going to go into those. I’m going to go into how people worked and created their workflows, so I’ll give you just a few random tips and then we can open it up into a further discussion or questions.

The first thing that we noticed: the people who were more productive using these tools weren’t trying to tackle some complex project by taking a whole bite of the elephant at one time. What’s really interesting is that if you’ve ever worked with software developers, you’ll notice that they break things up into stories. So here’s a new website, and maybe one of the features is to log in, maybe another feature is single sign-on, maybe another feature is automatic email notifications, and so on down the line. The most productive engineers who were having AI write the code for them were actually the developers who maintained the discipline for breaking down the tasks that they were going to tackle, even though the AI was actually going to write some of the code for them.

So you got to think, let that sink in a little bit. You might think that maybe the most productive people are the people who are just going to try and toss the whole problem at an AI system and let the AI system figure it out and spit out the right answer, but that’s actually wrong, because in those cases they actually spent more time trying to refine and refine the answer, working and then reworking, working and then reworking. Whereas what was counterintuitive and impressive to me was that the developers who maintained the discipline (yes, it’s a new sexy world where the AI can work for you) what they actually did instead was they stuck to the very disciplined breakdown of tasks, prioritization of tasks, detail-oriented for each task, and then took one of those tasks and used an AI agent, like one of the tools that Jacine recommended, or just plain old ChatGPT in one of its models, and focused just on that task and iterated on that task until it got to a place where it was polished enough that they were happy with it. Then what they did is called that task done, and then pulled off the next task or reprioritized based on what they learned, reprioritized the remaining backlog of tasks, and then took the next one and then focused on that next one and kept going.

What was interesting, and the reason that was actually more productive to be more disciplined about how you work, not less disciplined, was because the context got smarter and smarter. So if the people took the time to be very detailed with these tickets, user requirements, test cases, things like that, and loaded them into the context of the ChatGPT or whatever tool that they were using, and once the code got to a good stable place, passed all the tests, was good, was correct, they manually checked it, then they loaded in the next task, keeping everything within the same context, building the context like layers of a cake over and over again, until it was this beautiful delicious cake that you would eat, only it was a software product or whatever.

So that’s tip number one, quick little tip for you all. Sometimes it feels, because these systems are really magical or they feel magical, it feels like I’m just going to give it a brain dump and start to work on that, and that works for the first draft 100%. But at a certain point, what you want to do is take a step back and really start to break it up into tasks, and then feed those tasks, tackle those tasks one at a time, just like you would any other complex project. I think the big misnomer is that yes, it makes complexity easier, but the tools for managing complexity are just tools for proper thinking, not necessarily magic wands, and especially not magic wands in the age of AI. So that’s the first little tip for you there, before you move on to tip two.

I want to add something: it’s also the reason why when you’re trying to offload what you’re doing, it should take days, weeks, not hours. I think, like Frank was saying, we’ve got these sexy tools that can do all of these things, whatever it is that you’re doing, you’ve developed so much knowledge around that subject, and you know so much, and so if you really spent the time to break it down, it’s going to take a lot longer, but the more work you do upfront, especially however we’ve identified it, whether it’s in software development or customer service or in sales, the better these tools are going to be in offloading the productivity to them. When you’re going in accomplishing it, it’s just proven time and time again, whether you’re 10x more productive, 4x, or 1x, it really depends on how you break it down.

Yeah, that’s 100% right, and this obviously applies for a big complex project, and you break it down and everything, and then you use AI in parallel with that structure. But it also works for very easy tasks. One example is I was making a meal plan for my daughter, who just moved into a dorm room last year, so she needed a budget and the foods that she was going to buy, how much we were going to spend, and so on. So it’s not a big multi-month campaign or product launch or whatever, but breaking it down into those kinds of requirements or tasks or steps actually yielded much better results with ChatGPT, in my case, which was the model I was using.

And actually this leads me to the second little point here: what we also noticed when people were the most effective with these tools was that they thought about, I’ll use the term, test cases for their prompts, or test cases for the results, and requirements for the results that were coming back from the model. So a good example here is taking the time to define the characteristics of the response upfront, and it’s a great thought exercise. So for example, if you’re working with documents or something like that, and you’re going to ask ChatGPT to write you a contract or to answer questions about a legal contract or whatever it is, you want to consider the specifications, if you will, of what the response should be. So it should be things like, okay, for this particular response, I need it to be jargon-free. For this particular response, I need a citation of the exact text that you’re going to be using to reason about your answer. For this particular response, I need it to be concise, no more than, you know, whatever X, Y, or Z.

And so they actually took the time to almost paint a template for the prompts for the models, and the AI models would then use that template and fill in that template. That cut down so much time, because if you take the time to figure out what you need, it’s a thought exercise where you figure out what you need, and the AI can use its infinite knowledge of all the world and all humans to start filling in the blanks, and it works much faster.