Every: Context engineering is knowing when to start over
Mike Taylor, head of evals at Every, published this piece in October 2026 as a practical framework for understanding why AI writing assistance degrades over the course of a long chat thread — and what to do about it.
The core observation is that the problem is not the model. It is the accumulation of context. Taylor names this pattern “context rot,” and he describes seven ways it manifests.
Context distraction: Earlier examples and instructions constrain new answers. The model keeps following old patterns even when you have moved on to something different.
Context confusion: Irrelevant details that appeared earlier in the conversation surface in unrelated responses. If the model saw a framework someone used in a previous document, it may start applying it where it does not belong.
Context neglect: Models attend most to the beginning and end of their input. Important details buried in the middle get underweighted. The practical fix is to put the key instruction or constraint at the top of a message, not in a parenthetical halfway through.
Context drift: Over a long conversation, the model forgets standing preferences — spelling conventions, word count limits, tone notes. These need to be restated in each new session rather than assumed to persist.
Context clash: Models commit to early assumptions. If the brief changes midway through a thread, the earlier framing stays active. Starting with the complete brief in the first message is more reliable than building up to it gradually.
Context poisoning: Early errors, including ones you have corrected, keep resurfacing. Editing the original message is more effective than arguing with the model later in the thread.
Context underspecification: When details are missing, the model falls back on generic answers. Before leaving a long thread, Taylor recommends asking what information would have improved the responses — then using that insight to write a better brief for the next session.
The recommended practice is to start a fresh thread for each distinct writing task, rather than continuing a single conversation indefinitely. This is not about hitting a hard context limit. It is about recognizing that old context actively pulls responses in the wrong direction, even when the window is technically large enough to hold everything.
The practical infrastructure Taylor describes — Markdown files that store reusable context like style notes or structural templates, read at the start of each new session — is straightforward to set up and directly applicable to anyone who writes regularly with AI tools.
The article is most useful for writers who have already moved past initial AI experimentation and are trying to understand why outputs sometimes regress or feel less precise than expected.