I read How to keep thinking on a Saturday afternoon, and parts of it felt very similar to my own recent work. When the author describes skimming piles of AI slop and jumping across agent sessions, which leaves programmers feeling dumber, I felt strangely embarrassed: was he spying on me? His prescription that reading dense books and writing in your own words at night is unlikely to reconnect with thinking about your code.
Before we ask how to keep thinking in programming, it helps to be clear about thinking itself. Even with today's agents, thinking is still a programmer's core career skill. One example that both the author and I keep running into is a large refactor in a complicated codebase. I'm sure that it is one of the most thinking-heavy tasks in modern programming. A few paragraphs plus a simple design sketch will not hold: you get wordy nonsense and a chaotic structure. Then you spend hours chewing through weird terms and shapes the agent invented, and you end up as its slave. In a team of you and your agents, the leader is not whoever prompts, but whoever does more of the thinking.
There is a 10x speed-up button on your laptop. By nature, we crave the shortest path to done, so when a button blinks in the dark, we press it. We should not trap ourselves in a false dilemma: think more carefully or ship faster. Is there a way out of the dilemma? I think explaining is a great starter for thinking. It has helped me, and the practice is only two steps.
- First, classify the complexity of the task. On a simple task you do not need to understand every line. Save the slower thinking for a few kinds of tasks. From my experience, the list is short: a large new feature, a gnarly bug, a performance problem, or a large-component refactor.
- Then explain it to someone. That someone can be a PR reviewer, another agent session, or your future self. Ask whether they actually understand the code; if not, explain again.
These steps could be written into a skill and installed in your agent, so it could nudge you at the right time. I would rather not. Agents are strong in the virtual world of software; they cannot eat for you or exercise for you. They also cannot do the thinking for you.