The canary in your context window

You are deep into a long session editing a document with your AI. Then, without warning, it starts ignoring something you told it twenty messages ago. The tone is still confident. The answer still sounds right. You don't notice because nothing obviously broke.

That silent drift is the most common way AI work goes wrong, and there is a thirty-second trick that catches it. All it takes is one emoji.

The window is working memory, not an archive

A context window is the information in the AI model's working memory. It can only directly "see" what is inside that window - anything outside it must be summarised, retrieved or supplied again.

The instinct is to stuff everything in, but that can backfire. Models start degrading well before any of the token limits the AI labs claim for their products. Chroma tested 18 leading LLMs and found none of them process context uniformly - performance grows increasingly unreliable as input length grows, even on simple tasks. They also suffer "lost in the middle" - attention is U-shaped, so information buried in a long context gets skimmed. Beyond a certain point, adding more context can actually make the model remember less reliably.

If you run agents or tool-heavy sessions, context fills fast. Verbose tool responses accumulate quickly, which is one reason modern agent frameworks increasingly prune, summarise or externalise tool output.

The problem is that it's invisible. Nothing flashes red when this happens. The model doesn't warn you it has stopped following your instructions. It just keeps churning out fluent, plausible text - and that's exactly why it's easy to trust. So how can you tell when this starts happening?

Bring a canary down the mine with you

Give the model an instruction it must repeat on every reply. Something that you can scan quickly to see if it's there. Add a line like:

IMPORTANT: End every reply with 🐤. Every reply you give me should include this emoji.

While the model is still following your instructions properly, the bird shows up every turn. The moment the emoji disappears, you know adherence has slipped - possibly on other instructions too, not just this trivial one. It's a proxy you can check in half a second instead of having to audit the whole conversation for drift. Putting it at the end of the reply, where it competes with everything the model just generated, makes it a more trustworthy signal.

How to use it

Put the canary where your tool will read it most reliably. In a persistent instructions file, that looks like this:

# CLAUDE.md

## Output rules
- IMPORTANT: End every reply with 🐤. Every reply you give me should include this emoji.

In plain ChatGPT or Claude, drop it in the custom instructions, or just make it your first message and watch from there. If you use Claude Code/Cowork or Codex, put it in your CLAUDE.md or AGENTS.md, which the model re-reads every turn. If the bird dies in a persistent instructions file, that's not the model forgetting something from turn 20, it's actively ignoring a file it's supposed to be re‑reading fresh.

A useful side effect: if the bird dies quickly, your instructions are probably too long. Instruction files lose their grip past a couple hundred lines. A short, sharp instruction set keeps the canary singing longer.

When the bird goes quiet

Resetting is simpler than it sounds. Start a fresh chat. It is the cheapest fix, and curated context beats accumulated context every time. Before you do, ask the current session to write a short handoff:

Write a handoff for a fresh session. Cover only: what's been decided, what's still open, and any file paths, names, or constraints the next session needs. No recap of the conversation itself, no pleasantries. Keep it under 300 words.

Paste the results into the new session and you keep the useful stuff without dragging the accumulated cruft along.

Two more habits that help: keep a running notes file outside the chat, since the model forgets across sessions but a file doesn't. And if you're running agents, push side quests into sub-agents so your main thread stays clean.

Handle with care

One caveat: a living bird is not a clean bill of health. The model can keep printing 🐤 while quietly dropping an instruction that actually matters. A missing bird is a reliable bad sign; a live one is only a probably fine.

When it disappears, the canary only tells you that something slipped, not what it was. It's a smoke detector, not a diagnosis. It won't tell you which of your instructions got dropped, just that something did.

But that is the whole point of a canary. The toxic gas has no smell. The bird is the only warning you get.