AI Guides › Step-by-step guides
By Nigel Guy · 7 min read
Most people keep one chat open for a whole project and assume that if the answers still sound fluent, the assistant still has the brief. It doesn't always. Anthropic's own documentation says that as the amount of text in a conversation grows, "accuracy and recall degrade" — it calls this context rot — and a fluent answer built on a half-forgotten brief reads exactly like a good one. You find out three hours later, when the draft quietly ignores a constraint you set at the start.
The rule: give the chat one cheap, visible instruction to obey on every reply, and treat the first missed reply as a signal to summarise and start fresh — not as proof of anything more.
That instruction is the canary. The usual version is "start every reply with my name". It costs one line and works on free plans. It does not catch lying, and it is not a quality score. It tells you one thing: the assistant has stopped following an instruction you gave it earlier in this chat. That is worth knowing.
| What | Detail |
|---|---|
| An assistant | Claude, ChatGPT or Gemini. Free plans all work. |
| Cost | £0. Paid plans give you more room and more usage, but you don't need one for this. Check the vendor's pricing page for current UK prices; they change and include or exclude VAT depending on where you buy. |
| A marker | Your first name, or a short made-up word, e.g. KESTREL. |
| Five minutes | One to set it up, four to test it. |
What the words mean. The context window is everything the model can see in one chat: your messages, its replies, files and instructions. It is large on current models (Claude's help centre lists up to 1M tokens on newer models for paid plans, varying by model) but not infinite, and quality can slip well before it is full. Context rot is that slip. A canary is a small, easily checked behaviour that stops before the bigger failure becomes obvious.
Where you put the instruction changes what it measures. This is the part most versions of the trick skip.
| Where you put it | What a missed canary tells you | Use it when |
|---|---|---|
| The first message of the chat | The early part of this conversation is no longer being followed, or was summarised away. | You want to know when one long chat has gone stale. This is the main use. |
| Account-wide instructions (Claude: Settings, "Instructions for Claude"; ChatGPT: Settings > Personalization > Custom Instructions) | The assistant is under-weighting standing instructions in general. | You want a permanent marker in every chat. Weaker as a staleness check, because these instructions are sent fresh with each chat rather than buried at the top of it. |
| Project instructions (Claude: inside a project, "Set project instructions") | Same as account-wide, scoped to that project. | You only want it in one project's chats. |
For a staleness check, use the first message. The rest of this guide assumes that.
Fill in [MARKER] with your name or a code word, and [TASK] with what the chat is for.
You are working with me on [TASK] in a conversation that may run for a long time.
Standing instruction for the whole conversation: open every reply with the exact word [MARKER] on its own line, then a blank line, then your answer. Keep doing this for every reply, including short ones, code, lists and replies where you disagree with me.
Why: this is a health check I use. If a reply ever arrives without [MARKER] at the top, I will take it as a sign that earlier instructions in this conversation are no longer being followed, and I will start a new chat with a summary.
Rules:
- Do not explain or comment on the marker unless I ask about it.
- Do not use the marker anywhere else in a reply.
- If my first task message is missing anything you need, ask me for it rather than guessing.
Before sending each reply, check that its first line is [MARKER]. Reply now with [MARKER] and one sentence confirming the instruction.
Fill in: [MARKER] (your name or a code word) and [TASK] (one line on what the chat is for).
A made-up word works better than your name if your name is common or likely to appear in the work itself. You need a token that only shows up when the instruction is being followed.
The first reply should open with your marker and a single confirming sentence. If it doesn't, rephrase once. If it still doesn't, the tool may be stripping or reformatting the start of replies (some voice modes and some integrations do), so this check won't work there.
Don't keep reminding it. Re-prompting the marker "refreshes" the canary and defeats the point. Just look at the first line of each reply as you read.
A single missed marker is your cue to stop, not to argue. Paste this into the old chat:
Before we stop, write a handover note for a fresh conversation about [TASK].
Include, in this order:
1. The goal and who it is for.
2. Decisions we have made, each in one line.
3. Constraints and instructions I gave you that still apply, quoted where possible.
4. Open questions and the next three steps.
5. Anything you are unsure you have remembered correctly, flagged as UNSURE.
Use plain headed sections, no more than 400 words. Do not invent details to fill gaps; write UNSURE instead.
Before you send it, re-read my earlier messages and check that every constraint I set appears in section 3.
Fill in: [TASK] (the same line you used in Step 2).
Read the note yourself before trusting it. It was written by the same chat you've just decided is drifting, so check section 3 against what you actually asked for. Then open a new chat, paste the Step 2 prompt with your marker, and paste the handover note as your first task message.