You don't classify each type of .e.g voice hallucinated with schizophrenia. You could for example apply your argument to say "well, is the voice threatening to kill you only if you don't study for your test? If so, isn't the net effect beneficial, and as such it's not really a mental illness? If you like being motivated by your voices, you don't suffer from schizophrenia, that's only for people who dislike their voices."

I certainly cannot prove that there are no situations in which hallucinating imaginary people giving you advice would not be net beneficial, in fact, there certainly are situations in which any given potential mental illness may be beneficial. There have been studies about certain potential mental illnesses being predominant (or at least overrepresented) in certain professions, sometimes to the professional's benefit (also: taking cocaine may be beneficial. Certain tulpas may be beneficial.).

Who knows, maybe an unknown grand-uncle will leave a fortune to you, predicated on you being a drug-addict. In which case being a drug-addict would have been beneficial.

People dabble in alcohol to get a social edge, they usually refrain from heroin. Which reference class is a tulpa most like?

You can put a "Your Mileage May Vary" disclaimer to any advice, but actually hallucinating persons who then interact with you seems like it should belong in the DSM (where it is) way more than it should belong in a self-help guide.

Maybe when plenty of people have used tulpas for decades, and a representative sample of them can be used to prove their safety, there will be enough evidence to switch the reference class, to introduce a special case in the form of "hallucinations are a common symptom of schizophrenia, except tulpas". Until then, the default case would be using the reference class of "effects of hallucinating people", which is presumed harmful unless shown to be otherwise.