maybe you can preselect good ideas, build up guidelines describing most common pitfalls, extrapolate from ideas you already vetted etc and run on autopilot on a safe-ish subset