Here's a way to think about the work-flow.
Let's invent a concept I'll call 'placement intelligence' (PI). PI is how intelligent a process, programme or person is when it comes to arranging things on pages in a document.
A word processor has very low PI. It will refuse to put something in too small a space, for example. Or it might be able to wrap text around something. But it will not try to figure out whether your image should go before the previous paragraph or after the next one or five pages later rather than here. It just slaps stuff where you say and, if it all spills over the place, it all spills over the place.
This may seem initially OK because you have very high PI. You are really good at figuring out where you want things to end up (even if you can't always persuade the software to put them there - we're talking design, not implementation).
The trouble is that if the stuff earlier changes, you have to revisit the placement to make sure it hasn't got screwed up because the word processor has too low a PI to make any serious effort to prevent this.
What about LaTeX? LaTeX has moderate PI. It is much more intelligent than the word processor. It can often tell it would be better if an image or table went a bit earlier or a bit later or, even, a lot later. At the same time, it is built on TeX which has a high but domain-specific PI. It is really good at working out when to break lines and, generally, pages of text.
However, LaTeX has much lower PI than you. Even TeX in its specialist domain has lower PI. Add images or tables and it gets trickier. This means that there are cases in which you have to intervene.
Manual intervention involves adding a one-off additional constraint which limits the degrees of freedom La(TeX has to exercise its PI. It says, e.g., 'whatever your other algorithms for placement say, this must (not) go here (there)!!' (There are interventions which are less forceful, but it is the most forceful which are generally used.)
Now the system must obey this additional constraint. So it has fewer and worse options (as far as the algorithms go). This is OK because it got it wrong in this case, so you have to impose this limitation.
But that's only true if the content of the document up to 'here (there)' is finalised. Otherwise, you don't know where 'this' would have ended up if LaTeX did its thing. So, in that case, you are forcing the system to consider fewer and worse options possibly for no benefit at all.
You can only tell if you need to intervene when the document content is finalised. At that point, you make changes one at a time going forwards. After any change, you recompile because what happens later may be changed by the intervention.
The goal is to intervene no more than necessary. This is a good idea because LaTeX has moderately high PI. If it had lower PI, it wouldn't be useful because it wouldn't make good choices anyway. If it had much higher PI, it wouldn't be needed, because it would make good choices all the time.
But it doesn't. It is in between. And that's why you get worse results if you intervene earlier than if you intervene only in the final stage.
For the record, this is the method I use to produce documents of 200-400 pages with lots of diagrams, pictures, tables etc. so don't say 'that's OK for a few pages but not a book'. It is of the greatest importance in a longer document. The longer the document, the greater the complexity, the more images, tables etc., the more crucial it is to wait until the content is finalised.
[I almost never intervene manually if I'm writing an article.]