Context Tracker

Level 1 · Course 10

Most agentic models ship with a 200K token context window. GPT-5 runs 400K. DeepSeek V4 and Gemini push 1M. But here's what nobody tells you: drift doesn't start at the limit — it starts much earlier. Research shows effective context is roughly 60-70% of what's advertised. Rules ghost. The agent fabricates from training data. Domains bleed into each other. And it happens at predictable percentages, not random thresholds. Your agent gets dumber the longer you talk — but you can see it coming.

This course installs a percentage-based context tracker tuned to your specific model's window. Your startup protocol should land at 10-15% of your context window — enough to load the architecture, not enough to fade. At 25%, you're in the sweet spot: full awareness, best execution. From 25% through 75%, you'll get the best work of the session — whatever project you're on, this is the band where it ships. At 75%, your agent alerts you: time to wind down. Note files in memory, archive to vault, set up the handoff for the next session. Call janitor mode. Update project files, write session notes, close clean. Then /new and pick up where you left off — fresh context, same work. No degradation. No lost momentum. Just a clean handoff from one instance to the next.

Most operators burn through their context window without knowing it, wondering why their agent was sharp at 10 AM and useless by 2 PM. This course doesn't fix the degradation — it's structural. But it makes it visible. Your agent gets a fuel gauge, and you get the steering wheel. You decide when to push and when to hand off.