How to Tell When a Long Chat Has Stopped Listening

This chat has been open since before lunch, and the scroll bar has shrunk to a sliver. Somewhere around turn 60 (your 60th message and its reply), you ask for the final draft. The draft comes back ignoring the one rule you set an hour ago.

Yet the AI sounds just as sure of itself as it did when it was getting things right.

Nothing warned you, and nothing will. A long chat doesn't break loudly but drifts, and you're the only one in a position to notice.

An AI chat can hold only so much at once. Picture the chat as a desk. While everything fits, the AI can see all of it, but keep piling things on and they slide off the back edge. That could be your early instructions, a detail from turn 12, or the tone you asked for.

The AI doesn't know what fell off, so it fills the gap with its best guess, and a guess reads exactly like a fact. (The context window page explains the full picture of how much a chat can hold.)

Nothing in the tool will fix that. But you can learn to spot the drift early, and then stop leaning on long chats at all.

Plant a tripwire at the top

You can't see the desk, but you can put something small on it at the start and watch whether it stays. Small habits slide off first, which makes a small habit a good alarm.

Before a long piece of work, give the chat one easy rule to follow all the way through. Pick one you'd miss the moment it went.

For the rest of this chat, end every reply with one line in square brackets that repeats the main goal I gave you at the start.

While that line is there and correct, the chat still has your goal in view. The day the line vanishes, or quietly turns into a goal you never gave, the drift has started.

Treat the line as an early signal and no more. A single reply can drop the line for no reason, so what counts is the line going missing again and again.

A steady line doesn't prove everything else is safe either, because a standing rule is one of the stickier things on the desk.

Your tool may have personal preferences, custom instructions or a memory feature. If so, put the rule there and it runs in every chat, so pick one you'd be happy to see every time.

Ask for a handover instead of arguing

When the tripwire goes, you'll want to type "you forgot what I said" and carry on. But that piles more onto a desk that's already overflowing, so ask the old chat to pack up for you instead.

We're moving to a fresh chat. Write a short handover. Include the goal, the decisions we've made, what's done, what's left, and any rules I gave you. Keep it under 200 words.

Read the handover before you trust it, because the chat that wrote it is the one that was drifting. If the handover has the goal wrong or has dropped a decision, fix it yourself. Then open a new chat and start with your corrected version.

A tripped alarm is a reason to check the last answer, though that answer may still turn out fine.

Keep what matters off the desk

The tripwire is a quick patch. The lasting fix is to stop treating one long chat as the place where everything lives.

Some things must never slide off, like your goal, your rules and your reference material. Put them somewhere that sits outside any one chat. In Claude that's a Project, and most tools have something similar.

Every new chat then starts with all of it back on the desk, and nothing gets buried under turn 40.

Between tasks, the handover carries forward whatever is worth keeping, and you're the one who decides what that is. Within a task, keep the chat short on purpose. One task, one chat.

When the task is done, close the chat.