Both aoforce and even more Numforce require well converged SCF-/DFT-calculations (e.g. $scfconv 8 and jobex [-ri] -gcart 4).
The maximum core memory the program aoforce is allowed to allocate should be defined in the data group $maxcor; the recommended value is about 50% of the available (physical) core memory (in case of RI-calculations subtract the memory specified in $ricore).
To start aoforce in the lowest eigenvalue search mode, use the keyword $les. For its use as well as other keywords dealing with the calculation of only some irreps, see the Referenceguide part of this manual.
Numforce additionally requires the file gradient and will not work, if the calculation is not done at a stationary point of the molecular total energy. For reliable results, always use Numforce with the option -central (i.e. central differences) and be aware of effects due to the step length (option -d real;, default value is 0.02 a.u.). It is strongly recommended to use Numforce in DFT calculations only with the option weight derivatives in $dft, since this provides more accurate gradients and thus frequencies, see Section 21.2.10.
The Numforce script can be run for different levels of theory, which means that the binaries it calls have to be specified additionally. To perform calculations using the RI approximation, call Numforce with the option -ri. MP2 and CC2 calculations are requested via the options -level mp2 and -level cc2, respectively. NumForce works also on the RI-RPA level with -level rirpa (note: the -ri option must be used in this case). To select the correct option(s), use the explanations you get by calling NumForce -h.

For a review of theory and implementation see refs.  [158,159].