The first mistake people make at ninety percent context is emotional.
They treat it like an emergency.
Usually it is not. It just means the conversation has become expensive enough that you can no longer afford to be sentimental about what stays alive inside it. Some threads are still doing work. Some are already finished and only taking up oxygen. Some belong in files. Some belong in a fresh session with a cleaner spine.
The real skill is triage.
High context does not automatically mean quality drops off a cliff. What it usually means is that sloppiness gets punished faster. If you keep dragging old decisions forward, re-reading completed work, and switching between unrelated tasks because everything feels vaguely connected, the conversation turns into a junk drawer. The model is still capable. The room is just crowded.
The fix is the same one that works in the rest of life. Decide what still matters.
Front-load the critical context early. Tone, goals, red lines, current target. If the core of the session is clear from the start, the later growth feels more like accumulation and less like drift. Once the context gets heavy, stop trying to keep every solved problem emotionally available. If an article is done, let it be done. If a rule lives in a doc, let the doc carry it. If a task can be tracked outside the conversation, move it out.
This is why files matter. A tone guide in a file is lighter than a tone guide re-explained every hour. A real source-of-truth doc saves more context than a thousand clever reminders. The conversation should carry live tension, not dead reference material.
It also helps to batch the work. A session stays cleaner when it is doing one kind of thing for a while. Polish three related articles. Audit one site. Fix one design layer. The moment you start pinballing between publishing, security, branding, and some unrelated philosophical tangent, you are paying context tax for your own impatience.
The other thing people get wrong is the fantasy of continuity. They act as though starting a fresh conversation means losing the soul of the project. Sometimes the opposite is true. A good continuation is not amnesia. It is compression. You carry the decisions that still matter and stop dragging the furniture from the old room into the new one.
That is the difference between context and state. Context is atmosphere, scaffolding, nearby memory. State is what must survive. If you know the difference, ninety percent is manageable. If you do not, you waste the final stretch trying to save everything at once.
So no, the answer is not "never hit ninety percent." The answer is learn how to work there without turning the session into its own museum. Keep the live thread. Offload the rest. Start fresh when the compression becomes more useful than the drag.
The model is not usually broken at ninety percent. The operator is just being asked to show a little taste.
GhostInThePrompt.com // The model is not broken. The operator is just being asked to show a little taste. Triage accordingly.