◇ Could this help me?
Yes, and this is a useful corrective to the common advice to put everything into the project file. The research being cited is real and the inference drawn from it is sound: instruction-following degrades as instruction count rises, the agent's own system prompt consumes part of that budget before anything of yours is read, and a file padded with content that is irrelevant to the current task both wastes the budget and makes the file less likely to be consulted. The structural recommendation is the right one, since layout, architecture, and conventions apply to nearly every task while project inventories and personal background do not. General practice, with the honest caveat that instruction count is a proxy rather than a measurement, so the real test is whether the specific instructions that matter are being followed.
Audit my project instructions file by instruction count rather than length. Count the distinct instructions in it, then sort each one into three groups: applies to nearly every task, applies only to a specific kind of task, and does not change your behavior at all. Delete nothing yourself. Show me the groups and recommend what should move into a separate file that is only loaded when relevant, and what should be removed entirely because it describes rather than instructs. Then tell me which of the remaining universal instructions conflict with or duplicate each other, since those cost budget twice and give you a reason to follow neither.