Why answers come out average.
A model learns from everything it can read and returns the most likely answer across all of it. For routine work, that's fine. A summary, a websearch, a checklist, a first draft. The trouble starts when you need the answer that fits your situation, because the model has no idea what your situation is.
Personalization is the old fix.
This is not a new problem. In the nineties we built a Bayesian network into Office that watched you work and inferred what you needed. It got cut for cost, but the diagnosis was right: software that doesn't know the user can only serve the average user, and there is no average user. The hard part hasn't changed either: with a small sample of someone's behavior, it's tough to know whether you're helping or guessing.
Structure is context.
Nobody will type their life story into a context box. The trick - the good kind of sneaky - is to give the data they'd enter anyway a little structure: projects, areas, tags, dates. Let the structure carry the meaning. The user thinks they're organizing. They're also telling the model who they are. The limit: this only works if the user organizes at least a little.
What to build.
Give personal context an explicit home in your product. A profile the model always sees. Examples attached to jobs that repeat. Defaults that don't need retyping. A simple test: if the user has to say it twice, the interface dropped it.