The Canary Trick
Catch a degrading AI chat before it wastes your afternoon
The Idea
In 1982 Van Halen buried one strange line deep in a 53-page tour contract: no brown M&Ms backstage. It read like a rock-star tantrum. It wasn't. The band played huge shows with heavy rigging, and the contract was full of safety-critical detail. The brown M&Ms were a tripwire — if a venue got that bit wrong, nobody had read the contract properly, and the dangerous stuff needed a second look.
You can give your AI the same tripwire — Claude, ChatGPT, Gemini, whichever you use. Plant one tiny, easy-to-spot instruction — and the moment it stops being followed, you know the rest of your instructions are slipping too. That instruction is your canary.
Set It Up In One Line
This is the line from my own settings. Swap in your name and paste it into the box your AI reads before every single chat — the pattern is the same everywhere: tap your initials, go into settings, and find the personalisation section.
Claude
Click your initials, then Settings, and paste the line into the instructions box (the field where you tell Claude how to behave in every chat).
ChatGPT
Click your profile, then Settings, then Personalization, and paste the line into Custom Instructions.
Gemini
Click your profile, then Settings, and look for Saved Info — the section where you tell Gemini things to remember about you. The label moves around, but it's the same idea.
That's it — done once, and it's set up for every chat from now on. Every healthy reply opens with your name, and it costs you nothing.
If You Build With Claude Code
Same trick, different home — the canary goes in your CLAUDE.md. Easiest way: let Claude do it. Open a session, paste this in, and pick a name — anything memorable and a bit silly works best, because you'll notice the instant it disappears.
Prefer to do it by hand? Drop this near the top of your CLAUDE.md and swap in your own name:
Why It Works
A chat doesn't go from sharp to useless in one step — it degrades gradually. Every time you send a message, the AI re-reads the whole conversation to reply, and as that pile grows the model's attention spreads thin and the oldest, smallest instructions get deprioritised first. (The full story of why is on the context window page.)
The name prefix is the least important thing in the whole conversation, which is exactly why it's useful. It's the first instruction to fall off the edge — so when it goes quiet, it's telling you the model is starting to forget, while everything still looks fine.
What Happens If You Miss It
Ignore the canary and the slide is predictable. Each stage is harder to spot than the last — and more expensive to undo.
A small instruction gets dropped
FirstThe name prefix is the cheapest thing in your context, so it's the first to go. Harmless on its own — but it's the smoke before the fire.
Your instructions start to drift
ThenThe rules you set at the start — the format you asked for, the things you told it to remember, the way you like things done — quietly stop being followed.
It starts guessing
ThenInstead of asking, it fills the gaps with assumptions. The answers get confident and wrong at the same time.
It hallucinates
FinallyMade-up facts, invented details, confident answers to questions you didn't ask. By now the chat is costing you more than it's giving back.
Reading The Signal
Name still there
The chat is healthy. Keep going — your instructions are still landing.
Name gone
Don't argue with it, don't start repeating yourself — start a fresh chat. Carry over what matters and you'll get a sharper model back, cheaper than fighting a fading one.
One Honest Caveat
This is a smoke alarm, not a lab instrument. A single dropped name doesn't prove the rest of the session is degrading — a model can miss one instruction and be fine on the next reply. Treat it as a cheap early warning that's worth a glance, not a hard verdict. In practice, the moment the name goes is a good moment to ask yourself whether it's time to start fresh.
Inspired by the agent-naming idea popularised by Peter Steinberger and the "canary" framing that grew out of the Hacker News discussion around it. This is my own write-up of the technique — go read the originals.