Operator

Why your AI gets dumber the longer you use it

5 min read · By Tasha Shevchenko

It's not you, and it isn't broken. It's a known failure mode — with a two-part fix you can set up in five minutes.

If your AI feels sharp for the first ten minutes and then slowly turns to mush — ignoring your format, forgetting what you told it, answering like it never knew who you are — you're not imagining it. I felt exactly that a few months ago. It would start strong, then drift, then quietly stop following the rules I'd given it.

Turns out that's a real, documented failure mode. It has a name, a cause, and a fix. Here are all three.

Why it actually happens

Here's the part nobody tells you: every time you send a message, the AI re-reads the entire conversation from the top. Not just your last line — the whole pile, every single turn.

And the bigger that pile gets, the worse it performs. Researchers have a name for it: context rot. A 2025 study tested 18 of the top AI models — the ones behind every tool you've heard of — and found every single one got less reliable as the input got longer, even before the chat was anywhere near "full," even on simple tasks. More text in, worse output out.

On top of that, models are worst at using information stuck in the middle of a long stretch of text. They're sharp at the very start and the very end, and weakest in the middle — a documented effect researchers call "lost in the middle," where accuracy can drop by more than 30%.

Put those two together and you've got your problem. The rules you set thirty messages ago are now buried in the middle of a giant transcript — which is exactly where the model is least likely to follow them.

One honest clarification: your AI isn't "learning bad habits" that stick. Every brand-new chat starts clean. It's that one long, sprawling thread rots as it grows. That's good news — because it means the fix is entirely in your hands.

Fix #1: Give it a tripwire

First, a cheap way to see the drift coming before it costs you. Tell it to start every reply with your name.

Mine opens every answer with "Tasha." The moment it doesn't, I know the thread has drifted and it's stopped really following my instructions. It's not a fix — it's a warning light on the dashboard. And it has saved me from building real work on top of an AI that had quietly gone off the rails.

Fix #2: Stop burying your rules (the real fix)

Now the actual fix. Stop putting your rules in the chat. Anything you type mid-conversation sinks into that growing pile and rots right along with it.

Instead, put your rules where the tool reloads them every single time — your Project instructions (or your account's custom instructions). That text gets re-sent to the model on every message, sitting up top as standing orders instead of getting buried in the middle. It doesn't rot.

Three moves:

Here's the block I'd paste into your Project instructions — the name tripwire plus a few rules for how you want it to work. Make it yours:

Start every reply with my name, [your name]. If you ever stop, I'll know this chat has drifted and it's time to start fresh.

How I work — follow these every time:
- Write in plain English. No jargon, no hype, no filler.
- Give me your honest read, not flattery. If something's weak, say so.
- Ask one question at a time, not five.

About me and my business:
- [Who you are, what you do, who you serve, and what you're focused on right now.]

Do this now

Open the AI you use most, create a Project (or open your custom-instructions settings), and paste in the block above — your name tripwire plus three rules. Then run your next task in a brand-new chat. Five minutes, and you'll feel the difference today.