By skyforbes i found this weird little pattern that lowkey fixes like half of my instruction drift issues. instead of letting the model jump straight into execution, u make it echo the task back in one short line first. something like here’s what i understand u want me to do. it kinda forces the model into a verification mindset instead of its usual overhelping mode so it stops adding random steps or assuming stuff u never said.
pairing that with a tiny ask before assuming line from one of the god of prompt sanity modules makes the output way tighter without turning the prompt into a whole essay. curious if anyone else does this or has other small checks that keep llms obedient without overengineering everything.