There's a popular tip going around: tell the AI to greet you by name, and the moment it stops, that's your signal it's hallucinating and time for a new chat.
It's backwards. Saying your name is the easiest thing you ask of a model, so it's the last thing it drops. It will hallucinate long before it forgets your name, which means by the time the name disappears you've already been getting shaky answers for a while.
The real fix isn't a canary word. It's to keep chats short, and when you genuinely need a long one, hand off cleanly to a fresh chat with a summary. Here's the exact prompt I use to do it.
Give me a detailed handover summary of this entire conversation that I can paste into a new chat to continue exactly where we left off. Include: - The goal of this conversation - Key decisions made and why - Current status of the work - What's left to do - Any constraints, preferences, or rules I've set - Any files, names, or specific details that would be lost without this summary Write it as a direct briefing to the next chat, not as a recap for me.
How to use it
- When a chat is getting long, or before it starts drifting, paste the prompt above as your next message.
- Copy the summary it writes back.
- Open a new chat, paste the summary as your first message, and continue from there.
Using Claude Code? Do the same thing at around 60% context instead of reaching for /compact. Starting a fresh session primed with a handover summary is almost always cleaner and more efficient than a compacted one.
Ready to see how much capacity your firm is sitting on?
Book a 30-minute diagnostic call. No pitch deck, just a clear estimate of how many hours and dollars you can unlock with AI.