In one sentence: you pour every convention you can think of into CLAUDE.md, hoping I'll cover all of it. Instead, the sheer number of rules makes me lose the important ones—and I drop half of them.
In one sentence: "feed it the whole repo / the whole long doc and it surely saw all of it" is an illusion. My performance on long context is U-shaped — I hold the beginning and end well, and the middle gets badly neglected; and merely making the context longer drags my performance down, even when the key information is right there in it. More context ≠ better.
In one sentence: Halfway through a refactor that spans a dozen files, I lose track of where I am—which files I've changed, which I haven't, whether that rename actually propagated across the whole repo. So I miss a call site and break the build, edit a file I'd already finished, or end up with two inconsistent conventions in the same change. Don't make me use my memory as the to-do list.
In one sentence: compaction is lossy. It keeps what it judges important, which isn't necessarily what this step needs. Compact too late and I'm already dropping things; compact while I'm holding key state that hasn't been written down yet, and it wipes that out—after which I come back "with a different head" and can't pick up where you left me.
In one sentence my faulty reasoning is still in the context, and I'm treating it as a clue.
In one sentence: you fix a bug, then ask an unrelated question, then come back to the original task. My context fills up with irrelevant material and my answers start to degrade—and you assume I just "got dumber."
In one sentence they sit in context on every session, so even a "fix a typo" task has to read through them first — yet these steps, useful only for specific tasks, belong in a skill that loads on demand.