What is a feature class in GIS? A working definition

If you've spent any time in ArcGIS Pro or QGIS, you already use feature classes without thinking about the term much. But when someone asks you to define one cleanly, like during onboarding, or when writing a data management plan for a client, it's worth nailing down.
A feature class is a collection of geographic features that share the same geometry type (point, line, or polygon) and the same set of attribute fields. It lives inside a geodatabase, either as a standalone table or grouped with related feature classes inside a feature dataset. Every record in a feature class has a shape field and whatever attributes the schema defines, so "wetland_boundaries" has one row per wetland polygon, each carrying the same columns: delineation date, cowardin class, surveyor initials, whatever your QA/QC workflow requires.
That's the textbook answer, and it's accurate. What trips people up is confusing the feature class with the layer that displays it.
Feature class vs. layer
A feature class is data. It's the actual geometry and attribute table sitting in your geodatabase, on disk, with a schema that doesn't change unless you edit it. A layer is a reference to that data plus the symbology, labeling, and visibility rules you've applied in a specific map document. You can point five different layers at the same feature class: one symbolized by wetland type, one showing only features flagged for resurvey, one with labels on and one without. Delete all five layers and the feature class is untouched. Delete the feature class and every layer pointing to it breaks.
This matters in practice because clients and project managers often use the words interchangeably, and that's usually fine until someone asks you to "update the layer" when what they actually need is a schema change to the underlying feature class, or vice versa. Knowing which one you're being asked to touch saves a round of confused emails.
It also matters when you're handing off a deliverable. A shapefile is roughly analogous to a single feature class, but it's capped at 2 GB, has a 10-character field name limit, and can't store null values properly. A geodatabase feature class doesn't have those limits, which is why most consultancies standardize on file geodatabases for anything that's going to get edited, versioned, or joined across projects.
Where the standard toolkit runs out
Feature classes work fine once you know what you're mapping. The harder problem shows up before that: you've got a feature type specific to this client, this site, this regulatory ask, and no classifier ships with a category for it. Stockpile boundaries by material, a particular invasive species rosette, storm-damaged canopy gaps, informal access tracks across a lease boundary. None of these are in a standard land-cover product, and standing up a bespoke classification model for a single deliverable rarely pencils out against the project budget.
That's the gap Few-Shot Mapping is built around: you label a handful of instances of the feature on your own imagery, and the workflow learns that one class well enough to find every other instance across your project area, without a months-long modeling effort for a category that only exists on this one job.
Once the detections come back, they land in your workflow the way any other feature class would: same geometry type, same attribute schema, ready to join into the geodatabase and symbolize however the deliverable needs. The feature class definition doesn't change. What changes is how much manual digitizing it took to get there.
If you've got a custom feature type you keep digitizing by hand project after project, it's worth finding out whether a handful of labeled examples can do that work for you instead.