Chris Oakman

There is a specific kind of intelligence that is almost never celebrated but is consistently effective: the intelligence that recognizes when the game being played is not the game worth playing.

The patterns you don’t confront don’t vanish, they just change hands: "You either face your demons, or they raise your children."

I'm trying to cultivate a practice of "calm vibe coding" 😌 It's exciting, even addictive, to push the limits with parallel coding agents... but I can't do my best UI prototyping work if I feel like an air traffic controller. Multitasked, frenetic attention is not the path to quality. My target is a smooth steady flow, efficient but methodical. Used correctly, I think AI can absolutely help rather than detract, but it does require care. I had a very nice rhythm for this back in the ancient days of single-file editing with Sonnet 3.5. The newer models are a bit trickier since they often go further out over their skis in less predictable ways; but I think they can still be an upgrade if used well. Things I'm trying out: - One main thread at a time. My mind needs hours to sink into a problem. Fine to have parallel agents doing scouting work on closely related tasks. Also OK to have background agents doing unrelated work if I'm not checking up on them, but it's tempting to check in which is a challenge. Relatedly... - Trying to be more mindful of what I do the moment after submitting a big task to an agent. Tempting moment to get distracted. Usually for UI prototyping I think the correct answer is to think ahead to the next idea: play with the current state, sketch, maybe start up an agent to do some next research. (Distraction was also a problem with manual coding, but not as bad because focus was required to make any progress.) - Calculator Construction Kits. Have the LLM produce parameterized designs that I can tweak via direct manipulation in the UI itself. The LLM is producing a scaffold/jig that I can use to dial in a design, staying in a deeper flow than if I was coding. - More batched planning upfront. I have always found planning / clarifying questions helpful with LLMs, but trying to invest even more in co-developing a very deep plan up front. There should be no major surprises after the plan is developed. And the agent is on the hook for helping me quickly develop a deep understanding of the existing context in the code. - Keep doing some edits by hand. Sometimes I fall into the trap of making small tweaks by agent, which is much slower than doing them by hand. Navigating the files also keeps me closer in touch with the structure of what's going on. At the farthest extreme, I may do a plan with the agent and then fully implement by hand--but for some things this feels inefficient now. Very much a work in progress. I'm curious for other ideas -- what are you trying?