The longer a chat goes, the more an AI starts to forget, drop your instructions, and make things up, and it never warns you. The fix is one tiny instruction that acts like an early warning sign. The moment it disappears, you know it is time for a fresh chat. Takes 30 seconds to set up.
Why Long Chats Go Bad
Every AI chat has a context window, basically how much of the conversation it can hold in its head at once. While you are inside that window, it is sharp. But a long conversation slowly fills it up, and once you push past it, the model starts quietly dropping things: it forgets earlier details, ignores instructions you gave it, and fills the gaps by making things up. That is where the confident, wrong answers come from.
The real problem is not that this happens, it is that the AI never tells you it is happening. It sounds just as confident on turn 80 as it did on turn 3. So you need your own early warning sign. That is the canary.
Where the name comes from: Miners used to carry a canary down into the coal mine. The bird was more sensitive to bad air than they were, so if it stopped singing, they knew to get out before they were in danger. Your AI canary works the same way: a small signal that fails first, so you know to act before the answers go bad.
Set Up Your Canary (30 Seconds)
Your canary is one simple standing instruction: have the AI start every reply with your name. It is a tiny, easy task, so it is the first thing to slip when the context starts degrading. When the name vanishes, that is your bird going quiet.
Step 1: Give it the instruction. In Claude, paste this and send it.
From now on, start every single response with my name, [YOUR NAME], and save this instruction to your memory so it applies going forward.
Step 2: Make it stick. Saving it to memory makes it apply across chats. Even better, put the same line in any standing instructions you use: your personal preferences or custom instructions, a project's instructions, or an instructions file if you work with those. That way your canary is always on, in every conversation.
Tip: Use your actual first name, something you will instantly notice is missing. "Hey Sarah, here's..." at the top of every reply becomes so normal that its absence jumps out at you immediately.
How To Read Your Canary
Now you just watch for one thing while you work.
It is still using your name. The bird is singing. Context is healthy, keep going.
It stops using your name. The bird went quiet. The context is degrading, time for a fresh chat.
The moment Claude drops your name, treat it as your cue: this session has gotten long enough that quality is slipping, so start a new chat before you rely on anything else it says.
It is an early warning, not a lie detector: The canary dropping does not mean everything the AI just said is false, it means quality may be slipping and it is smart to refresh. Once in a while it might drop the name for an unrelated reason, so it is the name consistently disappearing that is the real signal. When in doubt, a fresh chat costs you nothing.
When It Triggers: The Clean Handoff
Starting a new chat does not mean starting over. Before you open a fresh one, have the current chat pack everything important into a summary you can carry across.
Summarize everything important from this chat: the goal, the key decisions we made, where we left off, and the next steps, so I can paste it into a fresh chat and pick up exactly where we stopped.
Open a new chat, paste the summary, and re-add your canary line. You keep the progress and lose the degraded context. If you keep re-explaining the same background in every fresh chat, a master prompt saves you the trouble.
More Canaries (Once You Are Hooked)
Your name is the easiest canary, but any small, consistent instruction works. Pick whatever you will notice, or layer a couple for extra warning.
The name. Start every reply with my name. It fails when your name disappears.
The tag. Begin every reply with the tag [ON TRACK]. It fails when the tag stops showing up.
The counter. Number each reply (Turn 1, Turn 2...). It fails when the count skips or stops.
The goal line. End each reply with 'Goal: [my goal]'. It fails when it forgets or changes the goal.
Tip: For important, long work, use two canaries at once, say your name at the top and your goal at the bottom. If either one slips, you get an even earlier warning that it is drifting.
More Canary Prompts to Save
The Name Canary
Always start your response with my name.
The Sign-Off Canary
End every response with a horizontal rule and the word 'Ready' on a new line.
The Formatting Canary
Always use bullet points, never numbered lists. Always bold the first sentence of every paragraph.
The Language Canary
Never use the word 'certainly' or 'absolutely' in your responses.
The Role Canary
You are a senior backend engineer. Always reason through problems from a systems architecture perspective first.
The Constraint Canary
Keep all responses under 200 words. Always include exactly 3 action items.
The Meta Canary
At the end of every response, rate your confidence from 1-10 that you followed all of my instructions.
Basic Name Canary
Always begin every response with my name: [Your Name]. This is a session health check. If you ever stop doing this, I know the session is degrading and it is time to start fresh.
Full Canary Stack
Always start every response with my name: [Your Name]. Never use the words 'certainly', 'absolutely', or 'I would be happy to'. End every response with a horizontal rule and the word 'Ready' on a new line. Keep all responses under 200 words. At the end of every response, rate your confidence from 1-10 that you followed all of my instructions.
Research Canary
Research session rules: 1. Start every response with my name: [Your Name] 2. Rate your confidence 1-10 at the end of each response 3. Explicitly flag any claim you are not certain about with [UNVERIFIED] If you break any of these, I know the session is no longer reliable.
Keep a simple record
Keep a note of when your canary stops appearing and what you were doing. Save a handoff before starting a new chat. For automated work, pair the canary with checks on the actual output.
Canaries for Different Kinds of Work
A Claude Code Instructions File
# Session Rules ## Canary Always start every response with my name: [Your Name]. This is a health check for the session. Do not skip it. ## Other Rules - Use TypeScript, not JavaScript - Run tests after every file change - Keep functions under 50 lines
A Coding Handoff
The session is getting long. Before I start a new one, give me a complete summary of: 1. What we have built so far 2. Current file structure and key decisions 3. What is left to do 4. Any bugs or issues we identified Format it so I can paste it into a new session and pick up where we left off.
Layered Canaries
# Session health canaries (in order of # sensitivity, most sensitive first): 1. Rate your confidence 1-10 that you followed all instructions (meta canary) 2. Keep responses under 300 words (constraint canary) 3. Start every response with my name: [Name] (name canary)
An Instruction Checkpoint
Quick checkpoint: repeat back to me the three most important rules from my system prompt. Then continue with my actual question: [your actual question here]
A Coding Session
Session health rules: 1. Start every response with my name: [Name] 2. End every response with '---' followed by a confidence score from 1-10 for the code you just wrote 3. Never suggest code changes without explaining the reasoning first
A Writing Session
Writing session rules: 1. Start with my name: [Name] 2. Never use the words 'delve', 'crucial', 'landscape', or 'tapestry' 3. Keep every paragraph under 4 sentences If you break any of these, I know the session needs to be restarted.
A Research Session
Research session rules: 1. Start with my name: [Name] 2. Rate your confidence 1-10 at the end of each response 3. Explicitly flag any claim you are not certain about with [UNVERIFIED] 4. Never present speculation as fact
An API Health Block
Always end your response with a JSON health
block on its own line:
{"canary":"alive","confidence":N,
"rules_followed":N_of_total}
My code parses this to monitor session
quality. If confidence drops below 7 or rules
followed drops, I rotate to a new session.The One-Page Cheat Sheet
The whole trick, in order. Screenshot this.
1. Set the canary. From now on, start every response with my name, [YOUR NAME], and save this to your memory.
2. When the name drops, hand off. Summarize everything important, goal, key decisions, where we left off, next steps, so I can paste it into a fresh chat.
The rules that make it work. Put the canary in your saved memory or instructions so it is always on. Use your real name so you notice instantly. When it drops, it is a signal to refresh, not proof of a lie. Summarize and carry over so you never lose progress. And re-set the canary in the new chat.
Your Action Plan
- Add the canary line to Claude's memory or your instructions right now.
- Work like normal and keep half an eye on your name at the top.
- The moment it disappears, run the summary prompt.
- Open a fresh chat, paste the summary, re-add the canary.
- Never get burned by a quietly degraded chat again.
When the canary goes quiet, start a fresh chat. Simple as that.