SYSTEMS
AI Canaries: How to Catch When Your AI Starts Lying to You 📚
AI does not go from perfect to broken overnight. It drifts. This guide shows you how to detect the drift before you start trusting outputs that are completely wrong.
AI does not go from perfect to broken overnight. It drifts. One message the model is following every instruction perfectly. Fifty messages later it is confidently generating information that is completely fabricated. The tone never changes. The confidence never drops. That is what makes it dangerous, and that is what makes canaries so useful.
What Is a Canary?
In the 1800s, coal miners brought canaries underground with them. The birds were sensitive to toxic gases. If the air went bad, the canary would stop singing and collapse before the gas reached dangerous levels for humans. When the canary dropped, miners knew to get out immediately.
That same idea has been borrowed all over technology. In software, a canary deployment rolls out a change to a small percentage of users first. If something breaks, you catch it before it hits everyone.
Peter Steinberger, creator of OpenClaw, popularized this technique for Claude. The idea is simple: add a tiny rule like "always start your response with my name" to your system prompt or CLAUDE.md file. As long as the model keeps doing it, you know the session is healthy. The moment it stops, something has shifted.
Why AI Sessions Degrade
AI degradation is not sudden. You will not get a warning or an error message. The quality just quietly slides. There are four specific reasons why this happens.
Context window limits. Every model has a maximum context window. Research consistently shows that information at the very beginning and very end gets the most attention. Information in the middle tends to get less weight. As your conversation grows longer, your original instructions end up in that low-attention zone.
Instruction dilution. When you start a conversation, the model has one or two messages to focus on. Fifty messages later, those original instructions are buried under thousands of tokens. The model unconsciously deprioritizes small rules in favor of the immediate task at hand.
Attention drift. The model processes your entire conversation every time it generates a response. As the conversation gets longer, it spreads its attention across more content. Your system prompt was 100% of the context in message one. By message fifty, it might be 2%.
Compounding errors. If the model makes a small mistake in message 10 and you do not catch it, that mistake becomes part of the conversation history. The model treats its own wrong answer as established fact and builds on top of it.
The Degradation Timeline
AI sessions follow a predictable pattern. Understanding this pattern is the whole reason canaries work.
7 Types of Canaries You Can Use
1. The Name Canary
What it tests: Whether your system prompt is still getting attention.
Best for: General use, Claude.md files, ChatGPT custom instructions.
Always start your response with my name.
2. The Sign-Off Canary
What it tests: Whether the model can follow formatting rules through to the end of its response.
Best for: Coding sessions, long-form writing, complex multi-step tasks.
End every response with a horizontal rule and the word 'Ready' on a new line.
3. The Formatting Canary
What it tests: Whether the model is still following your style rules.
Best for: Content creation, report writing, any work where output format matters.
Always use bullet points, never numbered lists. Always bold the first sentence of every paragraph.
4. The Language Canary
What it tests: Whether the model can suppress its default habits.
Best for: Writing tasks, brand voice work, editorial style enforcement.
Never use the word 'certainly' or 'absolutely' in your responses.
5. The Role Canary
What it tests: Whether the model is still in character.
Best for: Expert consultation, specialized research, roleplay-driven work.
You are a senior backend engineer. Always reason through problems from a systems architecture perspective first.
6. The Constraint Canary
What it tests: Whether the model can follow hard constraints.
Best for: Structured output, automated pipelines, situations where precision matters.
Keep all responses under 200 words. Always include exactly 3 action items.
7. The Meta Canary
What it tests: Whether the model is still self-aware about your instructions.
Best for: Research tasks, high-stakes work, situations where accuracy is critical.
At the end of every response, rate your confidence from 1-10 that you followed all of my instructions.
Setting Up Canaries in Different Tools
Claude (claude.ai, Desktop App). Go to project settings and add your canary to "Custom Instructions." This applies to every message in that project.
Claude Code (Terminal). Add your canary to the CLAUDE.md file in your project root. This file acts as a persistent system prompt that Claude reads at the start of every session.
ChatGPT. Open Settings, go to "Personalization," and click "Custom Instructions." Add your canary in the "How would you like ChatGPT to respond?" field.
Other tools (Gemini, Copilot, local models). Any tool that lets you set a system prompt supports canaries. For API calls, include it in the system role message.
What to Do When Your Canary Dies
Your canary just stopped singing. Now what?
Option 1: Start a new session. This is the safest option and the one you should default to. A fresh session gives the model a clean context window with your instructions at the beginning. In Claude Chat: open a new conversation. In Claude Code: run /clear or /compact. In ChatGPT: start a new chat.
Option 2: Remind the model. Send a message like: "You stopped using my name. Please re-read your instructions and continue following them." This can work for a few more messages, but it is a temporary fix. If the canary dies again after a reminder, start a new session.
Option 3: Save and migrate. For long coding or writing sessions, ask the model to summarize everything important from the current session. Start a fresh session, paste the summary, and restate your canary instruction.
Advanced Strategies for Production Use
Use multiple canaries. One canary tells you one thing is wrong. Two canaries tell you how wrong. If your name canary dies but your formatting canary is still alive, your system prompt is fading but not gone. If both die at once, the session is fully degraded.
Track canary death counts. Keep a simple log of how many messages each canary lasted. Over time you will learn your personal degradation curve and can plan session resets before quality drops.
Set canary thresholds for automated systems. If you are using AI in an automated pipeline, build in automatic restarts when a canary fails. Do not let a degraded model keep making decisions.
Combine canaries with output validation. A canary tells you the model is drifting. An output validator (like a regex check or a second model review) catches the actual errors. Use both for high-stakes work.
Copy-Paste Canary Prompts
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 (for high-stakes work):
Research session rules:
- Start every response with my name: [Your Name]
- Rate your confidence 1-10 at the end of each response
- 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.
The best workflows use canaries as early warning systems, not after-the-fact debugging tools. Set them up before you need them. Watch them. And when they stop singing, take it seriously.
Systems beat motivation every time. When you automate repetitive research, synthesis, or administrative workflows, you free up cognitive bandwidth for high-leverage strategy and revenue-generating execution. Implement this workflow once, document the edge cases, and make it part of your weekly operating system.