Couple a 100MHz bandpass onto the output, then amplify and bandpass and buffer it a bit more. Probably fine to use a pair of 1st or 2nd order bandpasses, and choose a modest impedance (~200Ω?) so the crystal oscillator isn't heavily loaded at the 3rd harmonic, and a topology (i.e., series-first) so that it isn't loaded at the other harmonics.
That's basically it. Filters can be designed with calculator tools, and a few transformations if necessary (particularly beneficial for high Q resonators, e.g.
https://www.jrmagnetics.com/rf/doubtune/doubccl_c.php). The characteristic impedance is sqrt(L/C), the (system) Q is this impedance over or under the system impedance (i.e., around 200Ω, don't forget the terminator!) and the resonant frequency is 1 / (2*pi*sqrt(L*C)).
The amplifiers can be anything, but I'd prefer a:
- Grounded base/gate: low reverse feedback, good gain, downside: input impedance a bit too low here.
- Cascode: low reverse feedback, high gain, modest input impedance. Good choice.
- Diff pair, especially with input to one side, ground to the other input, output from the far collector/drain -- noninverting, good gain, no Miller effect (low reverse feedback).
- Emitter follower: low reverse feedback, low voltage gain, low output impedance. Good output buffer. Add a series terminating resistor (and also a coupling cap), to drive 50 ohm transmission line, or, whatever you're connecting to.
You can get away with 2N3904s, or whatever JFETs (say, MMBF4393, J309, etc.), but preferably a somewhat faster type like MMBTH10 or better. Biasing is textbook. To find input and output impedances, you're probably better off setting it up in SPICE, then tweaking the real circuit and going back and forth between SPICE and reality until you've got a reasonably correct model of the real deal, including actual part values and parasitics. It's a process, but it's not at all insurmountable, even if you aren't very experienced in tuning RF circuits.
Tim