@Stephan Ralescu no, you are right. It isn't necessarily true. I just treat it as a sign of unclearness, as the GSD method usually prevents Claude Code from compacting the context window if you clear it after every GSD slash command finishes. Automatic compacting is what I would call the root cause for a rotten context window. Like the one classic AI chatbots suffer from, as they don't have any mechanics for context engineering. It is just a chat history that gets compressed eventually. But there are other kinds of problems with context engineering that are called context rot, e.g., when you save much information in a context window, but the model is incapable of working reliably on it, often measured with failing needle-in-the-haystack tests (Source: Chroma). This is also called within-window drift. And I doubt that this is a big problem with Anthropic's models inside of Claude Code unless you write several times that dark mode is definitely out of scope in several documents but mention one single time that you want a dark mode. How did you come to the understanding that 60% is the context rot threshold?

I may have to correct myself 😅 Asked colleagues. Some talked about an anecdotal "dumb zone" above 60% in Opus and Sonnet 4.5. One mentioned this paper: https://arxiv.org/html/2502.05167v3 I will look closer at it.

Since yesterday (1.9.0) there is a new command `/gsd:set-profile` with three modes ["quality", "balanced", "budget"]. You can find (most likely edit) the modes in `get-shit-done/references/model-profiles.md` but I did not try it. I just use Opus all the time 😅