I’ve written about context rot before, but it’s worth another look. Mike Taylor at Every just published an exceptional clear explanation.
What is context rot?
Context rot is what happens when a long AI chat starts giving you worse answers. The model hangs on to old instructions, repeats earlier mistakes, or skips details buried in the middle of the conversation. It won’t warn you. The output just slips.
Mike is head of evals at Every and coauthor of Prompt Engineering for Generative AI. His article, “Context Engineering Is Knowing When to Start Over,” is for paid subscribers, so I’ll just hit a few key points.
Seven ways a chat goes sideways
Mike lays out seven ways context goes wrong. Four stood out to me:
Old patterns box in new answers. Everything earlier in the thread becomes an example the model follows, even after you’ve changed direction.
Irrelevant details steer the response. If something is in the chat, the model assumes it matters and tries to use it. More instructions can make results worse, not better.
The middle gets lost. Models pay the most attention to the start and end of what you give them. A key detail in the middle of a long transcript or a stack of PDFs can get skipped.
Mistakes stick around. Once the AI gets something wrong, that error stays in the thread. Telling it “that’s wrong” doesn’t erase it. Your correction and the original error are both still there.
The other three cover goals that drift out of focus, early guesses that clash with new information, and prompts that leave out too much. Each one comes with examples and a fix.
What to do about it
His core advice is simple: start a fresh chat when the task changes, even if it’s the same project. A few other tips I liked:
Put your full brief in the first message instead of feeding in details a bit at a time.
When the AI makes a mistake, edit your earlier message and regenerate rather than arguing with it.
Before you leave a long thread, ask the model for a short handoff summary of key facts and decisions. Paste that into the new chat.
Keep the preferences you always want (tone, length, spelling) in a saved document and paste them in at the start.
He also shares a good question to ask the AI before you leave a thread. You’ll find it in the full article.
Why I pay for Every
I’ve subscribed to Every for a long time. Recently I upgraded to their premium level, because the information is that good for keeping up with AI.
This article is a good example of why. It’s practical, it’s based on real experiments, and it tells you what to do next. If you want to keep up with AI without chasing every headline, I’d recommend subscribing at every.to.
;-)
Ernie
P.S. Here’s Mike’s full article: Context Engineering Is Knowing When to Start Over. If you’re not a subscriber yet, it’s a good place to start.