I’ve used AI daily in my work for the last 8 months and there’s a lot it does well – proofreading decks, drafting emails, brainstorming case study ideas. But for complex and nuanced workflows, it struggles to give anything more than generic outputs.

The problem is AI is a jack of all trades and a master of none. Because AI knows everything, you need to narrow your prompting focus to the one piece of context it needs to do your task.

After a ton of experimentation, I finally got ChatGPT to help me write decent scripts for courses. If you’re trying to do the same and struggling to get there, let me tell you how I did it.

Treat AI like a new hire

The annoying truth: Figuring out how to do a task with AI will take longer than doing the task yourself.

But trust me – the upfront investment  will pay off. Writing the first draft of a lesson script takes me about 5 hours. With ChatGPT, I get a higher-quality product in 2. As I refine the process, I expect to get that down to 1.

The trick is to stop treating AI like an all-knowing entity, and treat it like a new hire. How did you learn how to do your job, after all? You spent years in the weeds learning how everything works. You don’t need instructions now, because you have a ton of experience and context – but AI doesn’t have that context yet.

The secret to getting to a usable output: Instead of just telling AI to do the task, focus on training it to do the task – the way you would a new hire.

4 steps to training AI on your workflows

Here’s the process I used to go from cheesy scripts to decent V1:

#1: Build AI’s context incrementally

As you’re building up AI’s knowledge base, focus on one layer of information at a time, and make sure it completely understands what you explained before moving on.

For my task, I started at the very beginning – what is Section? Then, what kind of courses do we have? How do we create them? And so on.

After each new concept, I asked the follow up question: Do you understand this? When writing a script for a sprint for example, I asked it to tell me what it thought a sprint was, then corrected or supplemented its understanding with more information.

💡Two tips for building context:

1. Use GPT’s file upload feature. For everything you want to explain to ChatGPT, share a resource it can study – slide decks, documents, examples. If you don’t have something to share, write out a definition or tell AI to look it up on the internet.

2. Start with the least amount of context possible for each new concept. More information does NOT equal better output. AI is not good at discerning which information to pull for your prompt. It will pull from everything you give it, which is how you end up with generic and unhelpful responses. So just give it the most pertinent information.

#2: Ask AI how it can help with your workflow

The best way to determine where AI will be most helpful is to ask it. Start by telling it the task you have to accomplish, and have a back-and-forth conversation about how it can help.

For example, I told ChatGPT that I write scripts for courses, that I want its help to generate the copy, and the resources I had available, and asked how it could help. Once it gave me some recommended steps, I gave it more explicit context, and it gave me detailed steps to provide the right prompts.

#3: Try out your workflow with AI

Now it’s time to try out your workflow, using the prompts and instructions that AI gave you in the previous step.

💡Two tips for testing your workflow:

1. Be very specific about how you tell AI to use your materials.

For scripting, I always include a slide deck that has speaker notes. I told ChatGPT to follow the content on each slide for the structure of the script, but to pull the details of the script from the speaker notes. That helped ChatGPT to capture more of the nuance of how our instructors teach.

2. Expect these instructions to be iterative. I edited what ChatGPT gave me and sent it back for it to study and incorporate, and that vastly improved the results.

#4: Iterate, iterate, and iterate some more

You will probably have to iterate on this process more times than you initially expect. Even though I’ve gotten ChatGPT’s output to a good place, I’m still finetuning the process to make it faster and more easily repeatable.

💡Two tips for iterating:

1. Have ChatGPT ask you the questions it needs answers to. I’ve asked ChatGPT, “What questions do you want to ask, to help you do a better job at this task?” That creates a helpful feedback loop where it can respond to what I give it and vice versa.

2. Be willing to take direction. The more direction you’re able to take on what context it needs, the sharper you’ll become at prompting it. This has the potential to greatly reduce the number of steps you have to take to get to a good output.

It’s possible, it’s just not magic

If you’ve ever tried something with AI and decided it wasn’t faster, you were probably right – it’s not at first. Training AI is a time investment, but it’s worth making if you want AI to take over a task you do all the time. Here’s my parting advice:

• Your first goal should be to get a usable output. You won’t create the most efficient process right off the bat, but you’ll get past the hurdle of getting pure garbage back.
  
  
• Once you’re in the finetuning stage, focus on finding the minimum amount of steps to a usable output. That will help you determine if it makes sense to build a custom GPT.
  
  
• Be realistic about what step of the process AI is going to have. If it’s a complicated task, AI may only be helpful for a piece of it. Getting a shitty first draft is the longest part of scripting for me, and thankfully that’s something AI can do.

Want more tips like this? Join our AI academy and get access to unlimited courses to help you integrate AI into your work.