Currently, the best approach is having a mix of Illumina and PacBio (or Nanopore) sequencing. First step would be to assemble with long reads alone, or a hybrid assembly with long reads and short reads. There are very good assemblers for the long read data alone (e.g. Flye, which already performs polishing with the long read data), I don't have experience with hybrid assemblers. A second step would be one round of long read data polyshing, depending on the assembler, the improvements can be dramatic. After that, at least one round of short read polishing, to correct for the homopolymer systematic errors.
Long read data still suffers from high error rate (or, for PacBio CSS, not high, but systematic errors at homopolymers), thus assemblies with long reads alone may have a high rate of missing genes, due to frame-shifting assembly errors. As Dave Carlson already noted, the gains in contiguity and percentage of the genome recovered can be a lot higher for long read data compared to short reads, though I never observed something as dramatic as his report.