Text first, instruction second, then _. The text is rewritten per the instruction, and the instruction and underscore are removed with the request.
The order is the syntax: text first, instruction last.
you type
you get
No selection or dialog is involved. The text is what you just typed, and the instruction follows it on the same line.
you type
you get
What remains is the rewritten text only; the instruction is consumed with it. When the model cannot act on an instruction it stands aside and your line stays exactly as typed.
One press restores the line exactly as typed.
The rewrite is a cycle stop like any word alternative. Because a transform replaces a whole line, the original is kept in the rotation, and one press returns it.
There is no set of supported verbs.
"Make it caps", "make it polite", "fix the typos", "shorter": none is a registered command. The model reads the instruction, so the set of possible requests is what you can phrase, not what was implemented.
The instruction acts on all the text before it. A whole-text instruction rewrites the whole text; naming a target ("make the heading bold") changes only the target, wherever in the text it sits.
A rewrite typically takes around a second, a few seconds for long texts, and you may type during it. Parts you changed in the meantime are kept; only parts you left alone are replaced. Your typing always takes precedence over the arriving rewrite.
you type
you get
Model output writes emphasis as punctuation (**bold**, *italic*). Those markers are converted to the host's real styling before insertion, so literal asterisks do not reach your text; the one exception is a surface that renders markdown itself at send time, which keeps them.