Context rot
Context rot is the drop in an AI model's accuracy and recall as the amount of text stuffed into its context window grows, meaning a larger context window doesn't protect a task like personalization from getting less accurate as more source material gets crammed into one prompt.
Anthropic's own documentation states the phenomenon plainly: as token count grows, accuracy and recall degrade, a pattern the documentation names context rot. That makes curating what's actually in context just as important as how much space is technically available. It's not a competitor calling out a rival's weak spot. It's a vendor describing a limit in its own product, part of why it's worth taking seriously rather than dismissing as marketing spin about someone else.
Why a bigger context window doesn't fix personalization
The failure mode context rot describes shows up hardest in personalization work: dumping a full research dossier into one prompt and asking for a specific, accurate email about one person in the same pass. Stuff a large context window with ten sources and ask for three sentences of outreach copy, and accuracy on that one person's details can drop, regardless of how large the window doing the stuffing is. The model with the most room isn't protected from this. If anything, a bigger window is more tempting to overload, because the room is sitting right there.
In practice
Split research and personalization into two separate passes instead of one long prompt. Let the research pass gather broadly, then hand the personalization pass a narrow, curated slice, and require it to cite which specific source backs each claim it makes about the one person the email is about.
What people get wrong
Teams assume the model with the largest context window automatically produces the most accurate personalized output. Context window answers a different question, how much material a model can hold in one pass, not how accurately it uses that material once it's holding a lot of it at once. A model that wins on raw window size can still write outreach that reads like it was generated, because the failure mode isn't a lack of information. It's too much information, badly compressed.
