Break up the script into steps, and place them in order in a dot-d style folder, like 1.foo.sh, 2.bar.sh, 3.baz.sh. Have a separate _context.sh which supplies the upfront variables or whatever are needed for each step. Ensure that each step can run on its own with just the context sourced. Then have a run.sh in the folder above that sources the context and the steps in order.
Now, that said, typically these kinds of things are dataflow or even buildsystem-adjacent, so I've also experimented with doing them in CMake (intermediate products managed with add_custom_target) or as Nix derivations (intermediate products are input-addressed in the Nix store), but I'd love to be aware of other tools that are better for expressing this kind of thing.