Designing MEG studies
To start with, if you come from fMRI background, there are few important differences between the haemodynamic imaging (fMRI, PET) and neurophysiological imaging (MEG, EEG) - read about them on MEG_vs_fMRI page.
Other things you may want to consider:
Spatial resolution issues
MEG's sensitivity to deep sources is low: magnetic field decays rapidly with distance. If you expect your activations to come mostly from subcortical structures, think twice before attempting to study them in MEG. If you expect a complex pattern in which both subcortical and cortical activation overlap in time, telling them apart will not be trivial and may not be possible at all. There are ways to confront these problems, usually involving some prior assumptions about to the activation loci, which can be theory driven or based on data from other modality.
MEG's resolution is not uniform along the brain surface either. Sources tangential to the surface are picked up much better than those radial (perfectly radial sources cannot be recorded at all, but perfect radial sources probably do not exist or at least rare due to the brain's shape).
Close-by sources tend to blur into a single activation spot in MEG data, especially if they have similar orientation. Don't expect to be able to tell apart sources closer to each other than 1cm.
As previously discussed, MEG activation are fast and relatively short. Keep this in mind when putting your stimuli together. Typically, a simple stimulus (say short sound or picture) will cause changes in the MEG waveform starting as soon as 50ms or earlier. The response will have complex dynamics and will be over in 2-3 hundred milliseconds. Stick a few stimuli together and all these will overlap making your analysis much more difficult.
For the same reasons, it is good to have a silent (or constant) baseline, i.e. to allow the neurons to return to the basic state before the nest stimulus is presented. In practice, it means allowing a few hundred millisecond of stimulus-free time before presenting the next stimulus.
Typically, ERF design requires tens or hundreds of stimulus repetitions for a good SNR. ERFs from single trials will then be averaged together to yield a measurable responses.