An Ask HN thread names Claude Code's addictive prompt loop

An engineer asked how to stop prompting Claude Code past the point where he still understood his own code, per the Ask HN thread. The habit stretched his workday. Multiple commenters converged independently on the same diagnosis: it's a variable-reward loop, where the next prompt always feels like the one that finally clicks.

The thread split into two kinds of fixes. One group fights the loop directly. Clear the session once it reaches 300k tokens and force the workflow to adapt to a blank context. Or set a hard stop alarm and physically move the laptop away. The other kind removes the reason to pull the lever so often. It spends the effort up front, on architecture and on chunking work into pieces small enough that the model works inside guardrails. That beats producing output reviewed in bulk after the fact.

One commenter who'd tried both reported that architecture-first works, because it changes what a prompt is for, and that the alarm-and-move method failed, repeatedly. The thread doesn't say why the guardrail approach holds where the willpower approach doesn't. It only shows the difference: one changes the environment, the other adds friction to staying in it.