◇ Could this help me?
Yes, with the numbers treated as heuristics rather than measurements. The underlying effect is real and well documented: quality degrades gradually as a context window fills rather than failing cleanly at the limit, and the practical consequences drawn here are all correct. The argument for manual over automatic compaction is the sharpest point in the video and is genuinely sound, because the automatic version runs at exactly the moment the model is least able to judge what matters. General practice, with one honest qualification: the specific percentages are rules of thumb that vary by model and by how repetitive the content is, so treat them as a prompt to watch for degradation rather than as a measured threshold.
I want to manage context deliberately instead of running sessions until they get bad. First, show me how to see how much of my context window a session is actually using. Then audit everything that loads automatically on every conversation, including my project instructions file and any global instructions, and tell me which parts of it are not relevant to most of what I do and should be moved out or linked instead of inlined. Finally, give me a short rule for when to compact and when to start fresh, and describe what degradation actually looks like in your responses so I can recognize it myself rather than relying on a percentage.