Two things decide more submissions than anything else: every tool carrying a title, and the right read-only / destructive hints.

Anthropic's portal syncs your tool list and flags anything missing them. OpenAI additionally reads an open-world hint. Set them from what the tool DOES, not from its HTTP method — a POST that only reads is still read-only, and a tool that deletes something irreversibly is destructive even if it returns 200.

Getting this wrong costs a review cycle you cannot get back: the fix takes an hour, and the queue does not.