As an official policy, I'd like Felipe's answer with the first sentence missing. That is to say:
[W]e should avoid discussing preprints in general terms. If there is a specific question about a specific step in a proof and suitable context is given so that reading the paper is not a prerequisite, then it is on topic.
This avoids getting into discussions about whether or not a result is crackpot, and I suspect will in fact lead to all questions about crackpot preprints being closed.
I think we can all agree that discussing crackpot preprints should be off topic. My opinion is that we should avoid discussing preprints in general terms. If there is a specific question about a specific step in a proof and suitable context is given so that reading the paper is not a prerequisite, then it is on topic.
BTW: The Sha paper which I believe prompted this post falls clearly within the crackpot category.
I think the policy should be as follows:
On questions regarding significant portions of a preprint (any part larger than what fits comfortably in a page of a MathOverflow question), the response should be that the question is too broad for this forum, as it "asks about too much at once".
On questions regarding quality, style, and correctness of a preprint, the response should be that the question is again too broad for this forum. In particular, quality should be assessed by experts (or experts in training) in the field, style by editors interested in publishing the work, and correctness by anyone who is interested in vetting the paper, any condition of which excludes almost all the members of MathOverflow. (This is as opposed to specific and brief questions on quality, style, and correctness which may refer to a specific preprint or set of preprints. Such specific questions may be suitable, but should be phrased in a manner independent of any preprint.) Further, MathOverflow is not for referring to preprints of undetermined quality, style, and correctness, but only to work that has received acceptance by some portion of the mathematical community.
On questions regarding a specific step or proof technique, the question should be written as if the question was found in an accepted text, with sufficient background and motivation provided, and with enough specificity that a good answer can be formed. A link to the preprint can be provided, but only as an adjunct to the question, not as an advertisement for the work. Preferably the link should not be needed to answer the question.
On opinions about a preprint, courtesy suggests brief and cool responses from MathOverflow members, such as "I would recommend it", "I would not recommend it", or "Due to the claims on page x line y, I think the argument cannot be repaired (as it contradicts the established result in [A]) and I have no time nor inclination to suggest a repair."
On questions like "Is my work present in the literature already", the response should be that the question is too broad. On questions like "I did X, Y and Z to tackle question Q, which is clearly different from [A], [B], and [C], is this in the literature?", the proper response should be given by one who is familiar with the area and the literature; anything else should be given as a helpful comment suggesting additional literature to pursue or search terms to use that may have not been previously considered by the poster.
Gerhard "Have I Missed Any Variation?" Paseman, 2013.10.01