10x10 — breadth then judgment

The fastest way to get a great result out of an AI that is not 100% reliable. Instead of fighting one mediocre output with round after round of corrections ("add this word, move that line") — which is slow because each round costs a full generation — you generate breadth, let a human pick, then converge. Human judgment is fast and high-value; spend the model on options, not on arguing.

Apply it by default whenever output quality matters and one shot is unlikely to land it.

Sandbox first — never touch the real thing until the pick is made

The variations are always generated in an isolated sandbox, never by mutating the real target. Copy the relevant slice into a throwaway space (a sandbox/ dir, /tmp, scratch files), generate all N options there, render/present them, and let the human choose. Only after a winner is picked do you apply that one change to the real data — the live deck, file, DB, document, etc. The real artifact is never in a half-edited state, and nothing is lost if a direction is rejected. This is non-negotiable: breadth is exploratory, so it stays quarantined until judgment lands.