find-the-framing

State what exactly should happen, why the obvious approach fails, and what would make the answer obvious. Then apply moves like turning search into optimization, writing checkable behavior cases, or borrowing a framing from a distant domain where the problem is routine.

Use it when the path is unclear, the design has more moving parts than the problem, or before committing to something expensive to undo. First write down the dumbest thing that could work, since reframing is wasted if the simple approach already succeeds.

Look for emergent behavior the design was never explicitly built to handle, such as grid cells appearing in a model trained only on prediction. Confirming code does what you told it to is weak evidence compared to handling unanticipated cases.

Reframing is for problems that resist simple solutions. If the dumbest straightforward approach works, reframing is an expensive detour to the same answer. It also requires honest statement of where each proposal breaks and which assumptions make it silently wrong.

Shrink the problem to a toy scale, such as two dimensions or three states, and work it end to end by hand. Always announce the simplification explicitly, because an unlabeled simplification quietly becomes a false belief about the real system.