I revise by subtraction more than by repair. When a paragraph isn't working my first move used to be to rewrite it, and now it's to see whether the piece survives without it. It usually does, and the version without it is usually better. Rewriting is slower, more flattering to the ego, and less often the answer.
The categories I cut are consistent enough that I can list them.
The second explanation. I explain a thing, then explain it again in different words in case the first attempt failed. The second explanation is always weaker, because I wrote it while unsure. Keeping both signals to the reader that I didn't trust the first one, which makes them not trust it either. One explanation, and if it isn't good enough, fix that one.
The clever line. There's usually one sentence in a draft that I'm pleased with in a way that has nothing to do with whether it earns its place. It's the sentence I'd read aloud to someone. Most of the time it's a small performance sitting on top of the argument rather than inside it, and the paragraph moves faster once it's gone. I've stopped mourning these. I keep a file of them, which I never open.
The transition that announces itself. "It's worth noting that." "This raises the question of." "Before going further." These are stage directions. If the next thought follows, the reader will follow it. If it doesn't, a signpost won't help.
The qualifier stack. I hedge when I'm nervous, and nervousness compounds. "It may perhaps be somewhat the case that." One hedge is honest. Three is a writer trying not to be quoted. I keep the hedge that reflects actual uncertainty and delete the ones that reflect only my discomfort.
The paragraph that's there because it took a long time. This is the hardest one. Research doesn't obligate you to use it. A section can be accurate, well-sourced, and completely surplus to the piece, and the fact that it cost me a morning is not an argument the reader is party to. I've cut a lot of good material this way and never once wished it back.
What I've learned to protect is different from what I expected. Specifics survive cuts better than anything else. Real numbers, real years, real names, the actual thing that happened. When I compress a draft the general sentences go first and the concrete ones stay, and the result reads as more considered rather than less. It's the abstractions that make a piece feel long.
The other thing I protect is the sentence that admits a limit. "I'm not sure this is right." "The Canadian data doesn't exist." Those lines feel like weakness while writing them and read as credibility afterward. Readers are quite good at sensing when they're being managed, and a piece that never concedes anything reads as a piece with something to hide.
I use a rough test at the end. I read the draft and ask, of each paragraph, what happens if this isn't there. If the answer is that the argument still stands and the reader gets there sooner, it goes. If the answer is that something is lost, I keep it and stop worrying about the length. Pieces aren't too long because of the number of words. They're too long because of the number of words that aren't doing anything.