[giladd]

> Perhaps there’s some kind of conservation law here: Any increases in programming speed will be offset by a corresponding increase in unnecessary features, rabbit holes, and diversions.

This resonates hard. LLMs enable true perfectionism, the ability to completely fulfil your vision for a project. This lets you add many features without burning out due to fatigue or boredom. However (as the author points out), most projects' original goal does not require these complementary features.

[sovenyr]

The flip side is that "fulfil your vision" requires actually having a vision before each session. Without explicit scope set upfront, the LLM defaults to filling time with adjacent improvements that feel like progress.

  I've started writing a one-line success criterion before opening the editor, the way @sambaumann's parent comment hints at — "the failure-mode-resistant version of this task is X." Not because the LLM can't figure out what I want, but because if I can't compress the goal to one        
  sentence, I'm not actually ready to build, I'm ready to wander.                                                                                                                                                                                                                               
                                                                                                                                                                                                                                                                                                
  The scope creep cure pre-LLM was fatigue. Now fatigue takes much longer to kick in, so the cure has to be cognitive instead.