[JauntyHatAngle]

Isn't this where the research plan implement loop comes in though? Assuming comprehension used effectively?

You should be learning alongside the llm through the research phase of anything. Updating your understanding of what is possible and best practices with rigorous checks and limiting scope to a high fidelity to leave little room for doubt. In-line commenting and questioning and asking for more passes on the living document of the area you are working on and then judiciously breaking it down further when you think there is too broad a scope for an llm to understand and synthesise properly.

If you do end up with too much vagueness, you need to limit scope more or break up the feature, implementation etc to be specific and applies enough to again, properly research and decide the plan.

I guess this is not so easy because lot of it depends on your own ability of reading comprehension, but I've had great success learning niche topics because I research (as a sub agent usually) essentially any topic that is mysterious until every level of the puzzle is properly mapped out to the specificity required.

Do I think most people are doing this? No. So I guess the statistics make sense. It's not intuitive to many people I think - because as you said, it's an embodiment of literature that is a tangled web of thought patterns and perspectives, so you need to pare it's answers down to the specific level, direction and area of ideas you want to get out of it. Way easier to do than it sounds, but it requires finesse in comprehension rather than getting lazy with it - normalcy of deviance comes to mind.