I'm going to preface this answer with the disclaimer that I have never used CRISPR/Cas, and there is a fair amount of speculation here.
But I think that efficient CRISPR mediated knock-in probably has three parts, targeting, integration, and selection.
The targeting is accomplished by the guide RNA, which is often introduced as DNA on the same plasmid that codes for the Cas9, or can be in vitro transcribed or synthesized as delivered as RNA. However, a poster suggests that a properly designed DNA fragment can be delivered and produce sufficient amounts of guide RNA in situ to produce efficient cutting.
The integration step is where the template DNA is inserted into the double strand break by homologous repair. Homologous repair efficiencies vary greatly from cell to cell, and is often less efficient than non-homologous end joining. However, using a single-stranded DNA might increase the efficiency.
However, even if only a small percentage of cells is correctly edited, it may be possible to select those cells and grow them. NEB describes a Cas9-GFP fusion (page 4) that can be used to bind GFP to a certain DNA sequence. It may be possible to produce CFP and YFP variants of this construct and use FRET to detect cells with the knock in. (This is where the speculation comes in) With the two fusion proteins and two guide RNAs for adjacent sequences in the knock-in gene, you might be able to detect cells where a CFP and YFP cas9 are close enough that you get FRET. This might allow you to sort cells by FACS, and then grow those cells.
Of course if your knock-in is making a membrane bound protein exposed to the surface, you could just use labeled antibodies to detect that protein.