AI - Personal Rules, Uses, and Thoughts

Common Ground Understanding

Throughout the changes over the past few years that I’ve been working I’ve seen folks from every walk using AI at a different capacity, some in… interesting ways, others in pretty standard, straightforward ways. I’d like to start by opening a few of my own rules I personally have developed inherently over time.

Common Ground Rules & Understanding:

1. AI is a tool.
2. Always take the time to understand the output and reasoning for the output.
3. AI models generate answers based on a large mathematical equation.
4. Tools are changing every other day, slow down and understand each tool’s use case before diving head first into building with it.
5. Use it to challenge your thinking or to expedite your execution.
6. Be conscious of your usage. (Opus is not required every time you are asking for information on a topic!)

Now, to move on to the uses.

Use 1: Research

When I refer to research here, it usually means a compilation of a bunch of different sources of information to make it something digestible before entering further. Take the example of if I wanted to understand the different API endpoints I could hit for Microsoft Azure, sure I could refer the documentation and start to manually map it out myself… but the tool expedites that learning into something visual and gives me a starting point where I could move in further from there and then start to pivot in more. (Also, please use ASCII outputs to enhance your understanding, don’t generate whole new images if you don’t need it. This is compute + time expensive.)

Another example is if I wanted to find something specific in some documentation but I can’t remember where I found it from or where it is, I use it to find the context of it quickly.

At times I use it as a starting point for research, but I find this can be 50/50. If we look deeper into the sources it uses then sometimes those sources may not be the best, which defeats the purpose of doing the research since I will need to redo it again myself. Best to limit its sources if this method is applied, which can also be unclear.

Use 2: Learning

This is a very touchy topic of mines. To be clear, I will always prefer a human being to teach me topics I’m unsure about, usually via Youtube or blogs or in person, since typically there’s plenty of information out there that I can start with as a form of interaction and consuming. Not only that, but there’s a person that I can go and reach out to if I have more questions or if something needs to be updated in the video or the article, as there’s accountability behind it.

Nonetheless, the AI method I primarily use is to challenge my understanding or my learning, or to dive very deep into a very specific aspect. For me, learning is always retained best when there’s a lot of friction in the way. In other words, a lot of mess ups and setbacks so that I get a vivid understanding by the time I finally get through all the messups.

There’s a tutor skill I had Claude built for me that provides me this friction and doesn’t like to let me get past it. I have a love/hate relationship with it.

The idea is that it will make me go through my own thinking and challenge my understanding by further prompting me into questions and avenues to go down and make me come to the realization. This is explicitly done with Haiku as I can receive the most value for minimal usage. If you skip this step and go straight to the answer, think about this perspective: have you retained that information since it was so direct? Do you know how many times the AI has been wrong, and then you have to course correct your thinking down the line? Did you understand just a very specific narrow piece of the overall concept, or did you understand it holisitically (e.g. think of ConfigMap v Kubernetes overall)? How many times did you re-read the output to actually understand it or did you just regurgitate the words down onto another page mindlessly?

The idea behind these questions is to illustrate my usage, which is meant to augment my learning and accelerate my own retention and understanding. Directly looking up questions and getting the answers only causes me to further prompt more since I’m naturally curious to understand things holistically, and sometimes things may have changed since it last received updated training data.

Use 3: Challenging my thinking, or speeding up my execution

Challenging my Thinking

This use tends to be a little tough to explain, so let me illustrate further.

Say you’re handed a specific task that is ambiguous in how you could execute upon it, so you need to start with an approach and then go from there. The easy way out would be to start your approach by asking for ideas from AI and then going from there. But is that really the best thought process?

I encourage you to start by thinking on your own - think about how you initially think about tackling the ambiguous problem, and reasoning behind it. Some questions I ask myself during this initial thought process:

• Why is this the solution I chose?
• Why is this the best route over all the others?
• What are the other types of solutions I thought of and why did I not find those to be the best?
• Based on the deadline, what is my recommended path forward the one I chose?

Why do I say to challenge yourself first? Why do I say to think before brainstorming with AI? Take a read on this New Yorker article by Kyle Chayka.

Essentially what I’m getting at is that AI is a mathematical, probability equation, where its output varies based on your input. Its output has been known to converge on similar ideas and similar points, so if you use it to map out all your solutions and brainstorm, you’ll also converge towards similar ideas and points as others, so then how are you differentiating yourself?

I veer towards using it to challenge my own thinking. Something framed out as the following (I don’t have a strict prompt, it’s a sporadic action I do that is usually performed against the strongest model for maximum results, and usually is when there are a wide range of approaches where I actually do want some external input):

I have been tasked to do X. My initial thinking on the approach to the solution is Y, and my reasoning behind it is Z. Based on the context I have given you, provide 5 other optimal options that I should consider? I have thought about potentially doing Option ZZ, but decided against it for ZZZ reasoning.

The primary reason behind this ‘Challenging my Thinking’ method is that it actually gives substance behind why I performed a certain action. If I simply ask the AI to think of the approach, this may not be something I can fully back up. While I also do that, I then reinforce my learning by either:

• A: Doubling down on my decision after reading the LLM output, since I still think for my use case, my approach is the best/most optimal
• B: Rerouting to using the AI’s approach and then plan out the next steps
• C: Adjust my approach to account for the AI’s approach.

Speeding Up my Execution

After I think out my approach, there’s a way of going about it in little steps. Maybe I need to develop some script, or maybe it needs to be some pattern analysis.

A few quick use ideas that come to mind when specifically speeding up execution:

• Document analysis: This has been repeated many times by others, but explicitly giving the AI documentation is one of the best ways for optimal results.
• Classification analysis: Let’s say I need to analyze a list of 100 use cases and they need to be classified by domain. I’ll use AI to support classification as well as provide documented rationale as to why, so that I can quickly review and for those that need a closer look, I can revise based on the ‘thought process’.
• Research (see Use #1).
• Scripting (see Final Thoughts section).

Final Thoughts

Ownership

AI is not a person and the model nor the companies hold accountability for its output, and I mean this primarily from a work perspective: You hold the reasoning and the understanding as to why you approached a task provided to you a certain way. It is your responsibility to thoroughly review and not only understand a solution, but understand why a solution works. Some example questions that come to mind, for example, when putting together a script may be:

• Why did the AI choose to use these implementations of methods/data structure?
• Why did the AI use these packages vs. others? Does that exactly fit my use case?
• Was there any other way I was thinking about it that it should’ve been? Why was it not written that way?

Others when putting together documentation:

• Was there any other sections I didn’t consider in this document? Why not?
• Does this documentation really cover everything I need? Review it word for word.

The point is, you need to take strong and clear ownership over anything you put together alongside AI. It is a tool and should be treated as such. These questions would have naturally come up if you were building on your own, but now you need to consciously think of these questions and have a good answer for them.

Automating

AI is non-deterministic, which is a fancy way of saying ‘its output can vary for the same input’. Scripting is more deterministic (you could implement non-deterministic algorithms, but that’s not really what I’m referring to here). I want to reinforce that you need to design and choose the correct implementation based on what you need to automate.

As an example, if you need to automate pulling in information via RSS feed at 10am each day, it’s probably best to use a script for this. You can use AI to develop it, sure, but it is best to use each tool for each type of job where possible, as you may find you run into headaches more when trying to get some things automated with AI alone.

Conscious Building

Be conscious about what you build and put out into the world. I shouldn’t need to say more than this on this side of the coin. At the same time, what I also mean is that you should be conscious as to what you use vs. what you don’t use for AI.

As an example, I don’t use AI for any of my writing on my blog. This is meant to be mental practice for how I explain concepts to others, from junior engineers to non-technical people, so that I train muscle memory and provide an open avenue for others on their own journey within the field.