Using AI in software: Your cognitive capacity as the new bottleneck.

Almost a month ago I presented at Vibe Coders Tokyo, organised by two senior Google employees. While most talks on AI workflows are about which models to use, how many agents to use, and how to maintain context, instead I decided to focus on a topic no-one seems to talk about. Something that makes a major difference in productivity for anyone however, and I'm sure it will for you too.

The last few months here in Tokyo I worked 24/7 on my project, Kodo, to succesfully bring it to a closed beta state. I learned a lot. I needed to be as productive as possible, because my personal funds were limited. There was one major thorn in my side during the development with my workflow: managing and writing prompts, and using multiple agents got exhausting. You would think with AI the work load should be lower, but AI doesn't remove work. It changes the bottleneck to your brain power. Mainly, your cognitive capacity.

You might not realize it, but you're probably switching a lot between roles during your workflow. One minute you're coding your APIs, the other minute you're rethinking your front-end design, or maybe you get a Slack message interrupting you. Here, your brain has to switch context. It disengages, and it takes time to get back into deep focus; studies say about 20-25+ minutes. This is costly for your brain. People who frequently context switch can expect a 40% drop in productivity**. AI makes context switching happen way more often. Furthermore, monotonous work becomes automatic, and ideation and complex work (e.g. finding bugs) become the main human input. Even though workflows get more efficient, your brain gets more tired.

Here's my winning strategy: in general, try to keep context switches as minimal as possible by focusing on one context for longer durations, without interruptions. Flesh out one layer before moving to the next. Whenever your thoughts drift, write down your worries for later, and move back. When waiting for a prompt don't get distracted. Do something that's low stimulation to sustain your attention but doesn't distract you. What works for me is watching a livestreamer. Aside from that, treat your brain like a muscle. Give it enough breaks and focus your energy on the topic that actually matters for your product. For me that is UX. When it comes to writing back-end I choose to let AI do the most (with some personal input to keep software design/arch strong).

Extra: This as a strong signal for recruiters that want to recruit for the future. Doing leetcode style tests will not be a good first filter anymore. You will have to measure cognitive capacity.

Next steps: It would be useful to research how to minimize context switches in more detail. Large organisations should be fine, but smaller teams could use a more complete strategy.

Have a great evening!!! (I wrote this w/o AI, crazy right?)

**: sources are available on request, or through online lookup :)