The conversation has developed its own weather system
Why ChatGPT Gets Worse in Long Conversations
Long chats accumulate old goals, abandoned drafts and corrections. Even when the conversation remains within a technical context limit, the current instruction can compete with a great deal of historical noise.
Why this happens
- Earlier instructions conflict with the current goal.
- Important constraints have become less prominent.
- Errors from previous answers are being treated as facts.
- The task has changed but the working brief has not been reset.
What to try next
- Ask for a concise current-state summary and correct it.
- Move the verified facts into a fresh conversation.
- Keep one external source-of-truth document.
- Split a large project into smaller, testable stages.
Frequently asked questions
When should I start a new conversation?
Do it when the goal has changed, the same mistake repeats or the chat contains several obsolete drafts.
Will a summary preserve everything?
Only if you review it. A summary can also preserve an incorrect assumption, so verify it before continuing.