The problem: An ambiguous request gets a confident-sounding guess anyway, because stopping to ask feels slower than just picking an interpretation and moving. The guess turns into two hours of work that answers a question nobody asked, and the two hours weren’t lost to a hard problem, they were lost to nobody checking, before committing, whether being wrong here was cheap or expensive.
This is one test, run at the moment a gap would get filled with an assumption: is this load-bearing? If a wrong guess would compound (other decisions get built on it, it’s slow to undo, it touches something outside your own draft) ask, with two to four concrete options and a genuine escape hatch, never an open “what do you want here?” If the wrong guess stays cheap, state the assumption out loud and keep moving instead of stalling on it.
Part of Operator Skills, a collection of skills for running a real Claude Code setup.