The algorithm should reduce the number of comparisons, but also the access time: when the user discovers a task, it takes time for him to read it and recover the memory he has about that task. So when the user has selected the most important of two tasks, you may keep the winning task for the next round. A benefit of this method is that the most important tasks will be well sorted first, while the less important will stay longer in a random order: that’s great because the less important tasks will get implemented later, and things will have changed by then.

Another way to reduce the access time would be to show 5 tasks at a time instead of 2, and ask to sort these 5 tasks. When the system has buckets of 5 tasks that are already sorted, the next 5 cards that are presented should not be random: they should contain the top task of a bucket (top of bucket A) and the lowest ranked task of another bucket (bucket B). If the user says the lowest task of B is more important than the top of bucket A, you know the whole bucket B is more import than bucket A.

The algorithm should reduce the number of comparisons, but also the access time: when the user discovers a task, it takes time for him to read it and recover the memory he has about that task. So when the user has selected the most important of two tasks, you may keep the winning task for the next round. A benefit of this method is that the most important tasks will be well sorted first, while the less important will stay longer in a random order: that’s great because the less important tasks will get implemented later, and things will have changed by then.

Another way to reduce the access time would be to show 5 tasks at a time instead of 2, and ask to sort these 5 tasks.

And since you say you want to make it fun, maybe it should sometimes be 5, sometimes 2, sometimes sorting them, sometimes picking the most important, sometimes the least important. This can make the process less repetitive.