In the Whitewash of AI
Read about my journey how to cope with an overwhelming and numbing feeling of being behind.
5 min read
In the recent weeks I have made a somewhat unnerving observation within myself. I’ve been using AI in my everyday life intensively, professionally as well as privately, and discovered a drastic increase in restlessness. The todo list in my head was never ending and many of the things on the list I decided to offload to an agent, only to realise that within seconds, the temporary void was filled with the next item on the back of my head. Not only that, but the output of the prompts to learn about a topic or find the information relevant to me was seemingly lost within seconds of reading it. I get the impression that I tried too many things at once, and it felt like I increasingly demolished my attention span by mentally already being on the next task in line. And speaking with people around me I know that I’m not the only one in this situation. While there definitely is an increase in efficiency and speed with using AI, especially in software engineering, the question that many people and business now try to answer is how to put AI to use both economically and mentally sustainably.
In combination with the drastic increases of prices introduced last June, now is a good time to take a step back and figure out, completely without AI, how a healthy approach and usage could look in a professional environment. Surprisingly, none of the following points are fundamentally new or unheard of.
The first thing to prevent the overwhelming feeling is similar to what you should do being in any high pressure situation: stop for a couple of seconds and think. Is what you’re planning on doing worth doing? Is there a better way? Am I just too lazy to think myself? A small amount of discipline to honestly answer these questions goes a long way. Recently, I had AI create a small script to parse the current project state out of a board including pipeline status, status changes of tickets and currently assigned work. It felt great to have it print a neatly formatted state within the console just at my fingertips. When I showed this to my co-worker he righteously asked: “Why don’t you just use the Web UI?” While I love the idea of having many things in the terminal, I could indeed have saved some time and effort by asking myself if it really is worth it and must admit that perhaps my co-worker was right.
Once you’re confident that using AI for the task at hand is the right way, the prompt structure is key to good replies. As a rule of thumb, a good prompt can be summarized as such: “If you give the task at hand (or the prompt) to a competent stranger will the person have enough information to give a solid response?” Is it clear what needs to be done? Is the expected response format clear? If there is too much ambiguity then assumptions will be made based on probability which might not match to your case. To reduce this ambiguity, providing the context is necessary. It is from here where the new term of context engineering has become more widespread. Context engineering is central to improve the quality of the output, especially in non-trivial tasks. Putting effort into this in the beginning will save you tedious debugging work in software engineering or more generally speaking it will save you the headache of losing AI to a wrong path (that it always so persistently stays on even after a couple of iterations…).
Once you have crafted a solid prompt, processing the response follows. And here again, I have experienced that it’s incredibly easy to just take what has been said or done for the truth without giving it a second thought, especially due to AI’s convincing replies. Especially in work context, it’s absolutely essential that it is solidly verified what has been generated. At the end of the day, whether it’s a customer mail or pull request, it’s your name that’s on the line and the content has to be verified thoroughly.
The final piece that made the whole change exhausting for me was the additional context switching. Given you have multiple agents running you usually end up having multiple task each needing input after some time. Loading the asking agent and its problem in brain-memory every time again is mentally exhausting. It’s there where you could say why not delegate these overview tasks to an agent as well which can go on a couple iterations further. In the end I haven’t found a solid solution to make context switching less exhausting yet, apart from consciously setting the pace yourself and accepting that it could go faster but doesn’t have to. If you’re too drained to review anything that all your agents have created then that’s less than if you focus on having two in parallel and doing that properly. This also helped me getting away from chasing every small idea that popped up for barely more than two seconds.
All in all I have to say there are huge benefits in using AI, but as it’s such an individualistic tool it’s best if every user makes its own experiences and finds out what works best for them. It’s incredibly powerful, but it’s still a tool and you’re responsible for the output you’re giving. Despite the drawbacks, it’s key for companies to allow employees to work with it and let them find out their most efficient way of work. Because only then can the efficiency gains be leveraged sustainably - both mentally and economically.