Nice analysis! I would perhaps add that LLMs are pretty helpful at (2) in my experience, as long as you don't turn your brain off. (As a Google/SO replacement, mainly.)
I have about three modes, working with LLMs:
A) Just use them for (2), do (3) and (4) myself.
B) Just use them for (2) and (3), but with very fine grained instructions - to the point where it barely saves me time.
C) Use them for (2), (3) and (4), but with this workflow, it quickly degrades into slop for me.
I still get the best results with (A), but for a quick PoC, (C) is hard to beat.
What's hilarious (or sad) is that I see a lot of people throw them at (1) as well, which seems to have a high chance of inducing AI psychosis and creating software nobody understands, let alone needs.
I have about three modes, working with LLMs:
A) Just use them for (2), do (3) and (4) myself.
B) Just use them for (2) and (3), but with very fine grained instructions - to the point where it barely saves me time.
C) Use them for (2), (3) and (4), but with this workflow, it quickly degrades into slop for me.
I still get the best results with (A), but for a quick PoC, (C) is hard to beat.
What's hilarious (or sad) is that I see a lot of people throw them at (1) as well, which seems to have a high chance of inducing AI psychosis and creating software nobody understands, let alone needs.