Lately, I have been experimenting with AI to improve how I work. On the one hand, I have been trying to automate some of my work with AI. On the other, I have started rethinking how I break tasks down: involving AI not only in carrying out what needs to be done, but also in improving how I do it.

I used to be a fairly committed practitioner of GTD (Getting Things Done), even if I was not especially good at following through. Put simply, GTD involves first capturing everything you need to do in an Inbox, then organizing, categorizing, and labeling those tasks. For larger, more complex tasks (Projects), you identify the next concrete action you can take: the Next Action. Once everything is organized, you can start working through your task lists.

It is a workable methodology. Recording tasks and breaking them down helps clarify vague problems and reduces the burden of remembering and planning everything in our heads.

In practice, though, I gradually noticed two problems.

First, maintaining the system takes considerable effort. Categorizing tasks, breaking down projects, listing next actions, and continually updating those lists as things progress all require time and energy.

Second, and more puzzling to me, even when I knew exactly what to do next, such as “open this website and start working on XXX,” I still felt strong resistance to doing it.

In other words, GTD can reduce uncertainty before we act, but breaking a task down into sufficiently concrete steps does not necessarily make it easy to carry out.

Of course, imagining future benefits can give our present actions a sense of long-term purpose. We should read a dry paper today so that we can master a skill in the future, or complete a few applications today to improve our career prospects. We know these things matter, and we are willing to work toward those goals. But while the future can give us reasons to act, we still live in the here and now.

A task being worthwhile, or something we ought to do in the long run, does not mean it is appealing right now. Without some thought going into the experience of doing it, the process can honestly be pretty unenjoyable, even when we rationally understand its value.

Beyond that, even after GTD has helped us identify the next action, carrying it out still requires many ongoing decisions, and those decisions consume cognitive resources of their own.

When reading a paper, for example, we need not only to understand its content, but also to keep deciding which sections deserve closer attention, which concepts require background study, when we can stop reading, and what exactly we should do next.

So I began asking AI to draw on research in education and psychology to redesign the Projects on my task lists. My hope was to make the process more engaging while reducing the burden of decisions I had to make on my own along the way.

Here is an example.

Suppose my goal is to develop a deep understanding of the Transformer architecture.

Using my old approach to breaking down tasks, I might make a list like this:

• T1: Read paper XXX and annotate it.
• T2: Watch the tutorials on website XXX.
• T3: Read the relevant source code.
• T4: Resolve a GitHub issue in the codebase.

From a task management perspective, this list seems perfectly reasonable. Every step has a clear object of action, and each relates to the final goal.

But when I actually started working through it, the experience was not very good.

Reading papers was tedious, and I often did not know how deeply I needed to understand them. Watching tutorials was easier, but I could easily slip into passively receiving information, spending a lot of time without knowing what I had actually learned. Jumping straight into a GitHub issue, meanwhile, could feel like too large a leap in difficulty, leaving me unsure where to begin.

The problem seemed to be that these tasks had not been designed around my current abilities and actual feedback, rather than any lack of clarity in the tasks themselves.

So I tried asking AI to take a different approach.

Instead of rushing to map out an entire learning path, it would begin with what I really wanted to achieve and identify the knowledge I needed at this stage. It would then use a few simple questions to assess my current understanding and design the next action based on my answers.

Rather than giving me a large amount of information at once, the process would get me actively involved as much as possible.

For instance, instead of explaining the entire Transformer architecture at the outset, AI would first ask me to explain what an MLP is in my own words.

If there were problems with my explanation, it might lower the difficulty, rephrase the question, and probe further. Once it had checked that I really understood the concept, it would ask me to write a simple implementation of an MLP.

Each step added just a little challenge beyond my current abilities, while giving me feedback fairly quickly. Learning this way, I kept going for three hours and still found it interesting.

Of course, this does not mean AI has invented a new teaching method. Active recall, adjusting task difficulty, and timely feedback have long been discussed in education. AI simply makes it possible to apply these methods to our everyday tasks at very low cost.

Beyond learning, I have also started using the same approach for vague, demanding tasks, such as projects involving a great deal of collecting, filtering, comparing, and judging information. The experience has generally been positive, though learning tasks still seem to benefit the most so far.

Still, this has led me to rethink task management itself.

In the past, I kept asking how to break tasks down clearly enough to help myself follow through. Looking back, it feels as though I was cutting my feet to fit the shoes: trying to make a person fit a system, while forgetting that the system is meant to serve the person. Perhaps a plan of action should not be something we settle on through reasoning before we begin and then try hard to follow. Instead, it should evolve through practice, continually adjusted and improved in response to actual feedback. I think good design needs to respect human nature and leave more room for the differences between people.