What a canary actually is
Old coal miners carried a canary in a cage down the mine. The bird was more sensitive to invisible toxic gas than the humans were — so the moment it stopped singing, the miners knew to get out, before anyone got hurt. It was an early warning you couldn't argue with.
Peter Steinberger — the engineer behind a lot of the Claude power-user playbook — borrowed the idea for working with AI. You plant a tiny, visible signal in your setup. While the signal's there, everything's fine. The second it disappears, you know something's gone wrong before it causes real damage.
Here's the problem it solves
Claude doesn't go from perfect to hallucinating in one dramatic moment. It drifts. First it quietly drops one instruction from your rules. Then it starts making assumptions. Then it's confidently wrong — and you don't notice until you've burned an hour acting on bad output. The canary makes that first invisible slip impossible to miss.
The whole thing takes one line and two minutes. Let's set it up. 👇
Set up your canary (2 minutes)
The classic canary is dead simple: make Claude start every reply with your name. While it keeps saying your name, you know it's still reading your rules. Here's the exact instruction — paste it into your rules.
Now — where that line goes depends on which Claude you're using. Pick your spot:
Claude Code → your CLAUDE.md file
This is the rules file Claude Code reads at the start of every session. Add the line near the top so it's one of the first things loaded. (Global file lives at ~/.claude/CLAUDE.md; per-project one sits in the project root.)
Claude.ai Projects → custom instructions
Open your Project, click into its instructions/knowledge, and paste the line there. Every chat inside that Project now carries the canary.
Regular Claude chat → personal preferences
In Settings, find your personal preferences / "what Claude should know about you" and drop the line in. It'll apply across your chats.
That's it
Swap [YOUR NAME] for your actual name, save, and start a new session. Claude should now open every reply with your name. Your canary is live.
How to read the signal
This is the whole game, and it's gloriously simple. You only have to watch one thing: is Claude still saying your name?
🐤 Canary alive
Claude opens its reply with your name. Your rules are loading, it's reading them, everything's behaving. Carry on — you're safe.
💀 Canary dead
Claude forgets your name. That's your early warning: it's stopped fully reading your rules. Don't trust the next answers. Time to act.
The moment the canary dies, do this
Clear the session and start fresh. Don't try to nurse a drifting chat back to health — by the time the canary's gone, the context is already crowded and it'll keep slipping. A fresh session reloads your rules cleanly. Two seconds now saves you an afternoon of confidently-wrong output later.
That's the entire system. One line in, one thing to watch, one action when it dies. The genius is that it turns an invisible failure into an obvious one.
Better canaries (bonus)
Your name is the classic, but it can blend in — you might not notice it's gone. Here are stronger variants that are harder to miss. Pick whichever you'll actually spot.
🐤 The emoji canary
Tell Claude to start every reply with a specific emoji — a literal canary 🐤 works. An emoji vanishing is far more obvious at a glance than a missing word. Great if you skim replies fast.
🔢 The version-number canary
Ask Claude to end each reply with a tiny tag like [rules v1 ✓]. When the tag disappears, the rules dropped. Bonus: you can bump the number when you update your rules, to confirm it loaded the new version.
🗣️ The phrase canary
Pick an unusual short phrase only your rules would produce — e.g. "Right then —" to open every reply. The weirder and more specific, the more obvious its absence.
Don't stack too many
One clear canary beats three competing ones. Pick the single signal you're most likely to actually notice, and trust it. The point is a glance, not a checklist.
Why sessions drift in the first place
Quick plain-English on the "why", because understanding it makes you trust the canary more. Claude has a context window — basically a working memory for the conversation. Your rules, every message, every file, every long answer all sit inside it.
As a session gets longer, that working memory fills up. When it gets crowded, your rules — which you loaded right at the start — can get crowded out. Claude isn't being lazy; there's just more competing for its attention, so the earliest instructions are the first to lose grip.
How drift actually unfolds
One instruction slips. A small rule quietly stops being followed. Easy to miss — this is the slip the canary catches.
Assumptions creep in. With your guardrails fading, Claude starts filling gaps with guesses instead of your actual preferences.
Confidently wrong. It states the guesses as fact, in a perfectly confident tone — the most expensive failure, because it sounds right.
The canary catches it at Stage 1, where the fix is free: clear and restart. That's the whole point — you intervene before Stage 3 ever costs you anything.
One honest note
The canary is a brilliant little safety net — but it's a symptom-catcher, not a cure. The real upgrade is a rules file that's tight enough to drift slowly in the first place, plus habits like starting fresh sessions and keeping context lean. That's the difference between fighting Claude and having it run your business reliably. If you want a hand building setups that actually hold, that's exactly what Wright Mode is for. 👇
Ready for Claude that doesn't fall apart on you?
The canary catches the drift. These help you build setups that barely drift at all.
Wright Mode Membership
Join a community of women entrepreneurs building reliable AI setups — rules files, systems and workflows that hold up. Live calls, templates, and ongoing support.
Claude Masterclass
Learn how to use Claude like a pro. From prompting fundamentals to building real workflows that save hours every week.
Claude Code Masterclass
Go beyond the chat interface. Build automations, process data, and create tools with Claude Code — including rules files that actually stick.